The system produced a weighted F1 score of 0.940, compared with 0.920 for a directly evaluated learnable-weight baseline, although that difference did not reach statistical significance in the reported paired test (p = 0.0625). The experiment used MUStARD, a public dataset of synchronized video, speech and transcripts extracted from television sitcoms including Friends, The Big Bang Theory and The Golden Girls. Each sample is associated with an utterance labeled as sarcastic or non-sarcastic, with preceding dialogue retained as context.

The 690 samples are evenly divided between the labels, comprising 345 clips in each class. Utterances were split at the sample level into 414 training examples, 138 validation examples and 138 test examples, with stratified sampling intended to preserve class balance. The researchers reported three runs with different random seeds and presented the mean and standard deviation for the principal metrics.

This was a benchmark study, not a clinical trial or a field deployment of a sarcasm-detection system. Sarcasm was treated as a cross-channel inference problem because a transcript may omit irony carried by a voice, while a facial expression or change in intonation can alter the apparent meaning of ordinary words. The text branch used RoBERTa-Base to produce contextual embeddings for the target utterance and its conversational history.

Preprocessing retained negation, intensifiers and emphasis markers because they can change pragmatic meaning, while informal forms such as “gonna” and “wanna” were normalized and inputs were tokenized with byte-pair encoding. The final hidden representation came from the model’s classification token and entered a binary sarcasm classifier. On its own, this branch achieved 0.82 accuracy, with recall and precision both reported as 0.82 and a weighted F1 score of 0.82.

Sound supplied a different view of the same utterance. The audio pipeline converted speech to waveform files, resampled them to 16 kHz, normalized amplitude, filtered background noise, removed non-speech intervals with a voice-activity detector and standardized utterance duration through truncation or zero-padding. A pretrained HuBERT-Base checkpoint converted raw waveforms into frame-level features with a convolutional encoder, then used stacked self-attention layers to model contextual acoustic patterns.

Lower layers were frozen while upper layers were fine-tuned, and mean pooling produced a 768-dimensional utterance representation. The resulting audio-only model reached 0.88 accuracy, recall and precision, with a weighted F1 score of 0.88. Its class-level errors were nearly balanced: the confusion matrix contained 63 correct non-sarcastic decisions and 59 correct sarcastic decisions, alongside eight errors in each direction.

The visual branch used ViViT-Base to model the timing as well as the appearance of facial and gestural cues. Video was sampled at 25 frames per second, with faces detected and cropped to 224-by-224-pixel regions. Pixel values were normalized, histogram equalization reduced some illumination differences, and augmentation included horizontal flipping, rotation and brightness adjustment.

Sixteen-frame clips were supplied to the transformer, whose patch embeddings and self-attention captured spatial relationships within frames and temporal dependencies across frames. This model alone obtained 0.90 accuracy and a weighted F1 score of 0.89, making it the strongest unimodal result in the study. It was also conservative about labeling utterances as sarcastic: 61 non-sarcastic and 63 sarcastic examples were correctly classified, with seven false positives and seven false negatives.

The central technical contribution was a compact, derivative-free search over the relative contribution of the three representations. Each particle in the PSO population represented a candidate vector containing text, audio and video weights. The weights were constrained to be non-negative and to sum to one.

Particles moved through the search space using personal and global best positions, and their positions were renormalized at each update. Candidate vectors were evaluated on the validation set, after which the best configuration was applied globally to the fused feature representation. The implementation used 30 particles and allowed up to 50 iterations, with convergence reported around 30 to 40 iterations.

The encoder parameters were held fixed during this stage, so the optimization changed three modality weights rather than the millions of parameters inside the three pretrained transformers. The method is feature-level rather than decision-level fusion. Modality embeddings are projected into a shared feature space and combined as a weighted sum before the final classifier makes a sarcastic or non-sarcastic prediction; individual class probabilities are not averaged or voted together.

It also differs from sample-adaptive reliability weighting because the optimized weights are learned globally from the validation set and remain fixed for all test samples and subsequent inference. The authors describe the approach as model-agnostic because the fusion module can be attached to different pretrained encoders without changing their internal architecture, but its reported benefits were demonstrated only with RoBERTa, HuBERT and ViViT on MUStARD. Controlled fusion experiments showed that the optimization procedure contributed to the gain.

With the same encoders and experimental protocol, simple concatenation produced a weighted F1 score of 0.88, equal-weight fusion produced 0.90 and learnable weighted fusion produced 0.92, while PSO-based fusion reached 0.94. The reported difference from the directly evaluated learnable baseline was two percentage points, but the paired statistical comparison did not pass the study’s 0.05 significance threshold. Ablation results also showed that the three channels were complementary.

Audio plus video was the strongest two-channel combination, with accuracy around 91.1% to 91.5% across the reported summaries, while audio plus text reached about 83.5% and text plus video about 84.5%. The full audio, text and video model reached approximately 94.4% in the modality comparison and 94.0% in the main evaluation. The results must be interpreted within the benchmark’s limitations.

Although the paper initially describes a speaker-independent protocol, the data were ultimately partitioned at the utterance level, so speakers and dialogue scenes could recur across the training, validation and test sets. The authors explicitly acknowledge that complete speaker- and dialogue-independent generalization was not guaranteed. The study did not quantify parameter count, floating-point operations, peak memory, training time or inference latency, making claims of real-time efficiency or lightweight deployment unverified.

Robustness tests nevertheless showed a gradual performance decline as modalities were omitted or corrupted, with accuracy ranging from 0.878 under 30% text corruption to 0.940 with intact inputs, underscoring the continued value of complementary evidence. Among the eight final test errors, the researchers classified three as context-dependent sarcasm, two as subtle or weak visual cues, two as cross-modal inconsistency and one as weak acoustic or prosodic cues. These categories identify the remaining challenge: some sarcastic intentions are not expressed through any conspicuous channel and must be inferred from conversational history, shared knowledge or culturally specific references.

The study supports a clear but bounded conclusion: on one balanced benchmark of television dialogues, explicitly optimizing the contribution of text, speech and video produced the strongest reported performance, with 94.0% accuracy and a 0.940 weighted F1 score. The evidence does not establish that the same gain will occur in larger, more diverse or fully speaker-independent datasets, nor does it show that the advantage over learnable weighting is statistically conclusive. Within the tested setting, however, the results show that sarcasm recognition benefits from combining complementary modalities and that the method of weighting them is itself an important experimental variable.

Subject of Research: Multimodal sarcasm detection using transformer encoders and Particle Swarm Optimization-based feature fusion

Article Title: Enhanced multimodal transformer framework for sarcasm detection with particle swarm optimization based feature fusion

Article References: Payala, V. K., Jeet, A. A., & Abinaya, S. (2026). Enhanced multimodal transformer framework for sarcasm detection with particle swarm optimization based feature fusion. Frontiers in Artificial Intelligence, 9, Article 1947291. https://doi.org/10.3389/frai.2026.1947291

Image Credits: AI Generated

DOI: 10.3389/frai.2026.1947291

Keywords: sarcasm detection, multimodal learning, Particle Swarm Optimization, RoBERTa, HuBERT, ViViT, feature fusion, MUStARD dataset, transformer models, conversational AI, Enhanced, multimodal

The sample included 2,507 medical students. Depression-related symptoms were measured with the Patient Health Questionnaire-9, and social isolation was assessed with the Revised UCLA Loneliness Scale-6, allowing the investigators to compare chatbot experiences with two validated measures of psychological distress. Students reported using conversational AI for several purposes.

Academic perceived support was the primary driver of engagement, cited by 88.8% of participants. Smaller proportions reported using chatbots for companionship, at 22.2%, or to cope with stress and loneliness, at 23.1%. Daily use was reported by 41.9% of respondents, indicating especially frequent engagement among a substantial minority of the surveyed students.

The survey found a high prevalence of mental-health and social-well-being difficulties among participants. Mild-to-severe depressive symptoms were present in 70.6% of students according to the PHQ-9, while 65.0% met the survey’s threshold for moderate-to-high loneliness according to the Revised UCLA Loneliness Scale-6. The research therefore characterized chatbot use in a population already carrying a substantial burden of depressive symptoms and social isolation.

Students exhibiting higher levels of psychological distress were significantly more likely to report emotional connectivity, perceived companionship and subjective well-being benefits from their chatbot interactions. However, the authors noted that these correlations remained statistically weak. Because the study was cross-sectional, the measurements represent a snapshot rather than a sequence of events.

The findings nevertheless reveal that these systems occupied both educational and relational roles in the lives of the respondents. Academic support was the dominant reported function, but companionship and coping with stress or loneliness were meaningful uses for subsets of students. The authors described conversational AI as a scalable, low-barrier support mechanism in a conflict-affected setting where access to formal mental health services is severely restricted.

That accessibility does not make the observed benefits equivalent to treatment or establish the long-term safety of chatbot-based support. The study directly measured use, symptoms, loneliness and perceived benefits, but it did not demonstrate that chatbots prevented social isolation, reduced depressive symptoms or worsened either outcome. The authors called for responsible integration and longitudinal research to determine the long-term effects of AI-mediated support on student well-being.

Overall, the survey identifies broad chatbot engagement and substantial psychological distress among Sudanese medical students while highlighting the need for closer examination of how AI tools are used. Its clearest finding is not that chatbots provide a proven mental-health intervention, but that many students already consult them for academic help and some use them in emotionally oriented ways. The weak correlations and cross-sectional design leave the direction and consequences of those relationships unresolved.

Subject of Research: Associations between conversational AI use, depressive symptoms and loneliness among Sudanese medical students

Article Title: Conversational AI chatbots: a double-edged sword for mental health and social well-being among medical students

Article References: Ahmed, A., Mohamed, M. A. E., Ali, W. B. A., Eldouma, M. Y. M., Abdalla MohammedAli, R. M., Barakat, H. O. M., Alapoud, M. M., Ali, A. S. H., Alngar, Y. A. A., Ahmed, S. A. M., Osman, M. M. A., Gassemalseed, T. F. A., Gadallah, A. A., & Eljack, M. M. F. (2026). Conversational AI chatbots: a double-edged sword for mental health and social well-being among medical students. BMC Medical Education. https://doi.org/10.1186/s12909-026-10326-3

Image Credits: AI Generated

DOI: 10.1186/s12909-026-10326-3

Keywords: conversational AI, medical students, Sudan, depressive symptoms, loneliness, cross-sectional survey, digital well-being, academic support, Conversational, chatbots, double-edged, sword

The report describes a development and pilot expert-evaluation study, with no nursing students enrolled and no measurement of whether the simulator improves competence, patient outcomes or human-centered care in clinical practice. The simulated task placed a learner in the role of a nurse attempting to de-escalate John, an angry patient in a clinic. Built in the Unity 3D platform, the experience includes a tutorial, a waiting room and doctor’s office, and a debriefing room.

The learner wears a Meta Quest 3 headset, uses handheld controllers and communicates with the patient by speech. The scenario was inspired by the Canadian Alliance of Nurse Educators Using Simulation’s De-escalation of Angry Patient Assessment Rubric, with the first author and a nursing education expert mapping the rubric onto a sequential storyline. The system is designed for training in verbal de-escalation and therapeutic communication, competencies that the report identifies as important in nursing education but difficult to practise consistently because real encounters with agitated patients are limited by safety and logistical concerns.

The 11-part storyline moves from establishing communication to discussing treatment. Sections 1 through 5 focus on a professional introduction, inviting the patient to a private room, exploring the reason for the visit, acknowledging frustration and possible links between emotional distress and physical symptoms, and validating emotion through active listening. Sections 6 through 8 address direct suicide-risk assessment, redirection toward abdominal assessment and acknowledgment of clinical findings alongside the patient’s emotional state and behavior.

Sections 9 through 11 cover reassurance, inquiry about medications and collaborative discussion of acceptable treatment options while respecting patient autonomy. Each response is assigned to one of three levels: Novice, Intermediate or Competent. The reference classes were defined with a nursing education expert to represent responses that omit key elements, responses that meet the expected competency incompletely, and responses that are clear, empathetic, professional and clinically appropriate.

Convai provides the virtual characters’ conversational capabilities through an application programming interface. Its cloud-based speech-to-text service converts the learner’s speech into text, which is used to generate the patient’s next response, while text-to-speech produces the audible answer. A narrative-design function uses a tree-like structure of objectives and decision paths so that dialogues can vary without losing the required sequence of 11 sections.

The patient’s backstory was adapted from the CAN-SIM angry-patient scenario, and the same narrative constraints and moderation settings remained active during the suicide-risk section to reduce the likelihood of inappropriate or unsafe generated content. The simulation also displays transcripts and instructional prompts, uses custom animations and facial movements, and includes a virtual facilitator in every room to provide guidance. Automated grading operates separately from the conversational layer.

The researchers fine-tuned 11 models from the pretrained bert-base-uncased model, giving each scenario section a classifier tailored to its communication objective. Each model was trained with a dropout rate of 0.3 and a linear classification head with three outputs corresponding to the learner levels. Inputs were tokenized with truncation, fixed-length padding and a maximum length of 64.

Each dataset was divided 80% for training and 20% for validation with random seed 42. Fine-tuning used AdamW with a learning rate of 2 × 10−5, cross-entropy loss, three epochs and batches of eight. The section-specific design was intended to reduce ambiguity by allowing each model to focus on a distinct conversational context, expected intent, therapeutic language and learning objective.

At runtime, responses were sent from Unity to the relevant model through an HTTPS REST API on a separate computer. Because suitable collections of authentic nursing-student responses were unavailable, the researchers generated labeled training data with GPT-4o. Section-specific prompts included the scenario context, dialogue history, target performance level and expected competency.

Additional prompts were used to encourage variation in wording, length, punctuation and communication style. The researchers manually reviewed the generated responses, removed duplicates and near-duplicates, and excluded responses that were clinically implausible, inconsistent with the target competency or unsuitable for the assigned level. A nursing education expert reviewed representative examples and grading criteria with the first author.

Class distributions varied across sections because the objective was sufficient representation and diversity rather than artificially balanced classes. The labels were established through collaborative review and consensus rather than independent coding by multiple assessors, so inter-rater reliability could not be calculated. During the pilot, the models assessed uncorrected speech-to-text transcripts from the faculty members, exposing the pipeline to the recognition errors expected in real-time use.

A request to Convai took about 2.21 seconds under the selected settings, while the general delay between a learner speaking and the virtual patient answering was approximately three to five seconds. Each assessment cycle took about 10 seconds and ran in the background while the simulation and interaction continued. The authors emphasized that these timings were estimated from system use rather than formally benchmarked.

Eight of the 11 sections achieved seven correct classifications out of seven, while Sections 3, 5 and 8 each achieved six of seven. With only seven responses per section, the confidence intervals were wide and overlapping; the researchers therefore treated section-level differences as descriptive rather than statistically meaningful. The three disagreements illustrated the difficulty of scoring context-dependent communication.

In Section 3, the response “Okay John, so you came with abdominal pain, tell me more” was labeled Novice by the model but Intermediate by the expert, with the model assigning similar probabilities to all three classes. In Section 5, the model rated a response about the patient’s abdominal pain as Competent, whereas the expert considered it Novice because it did not validate emotion or address anger. In Section 8, a compassionate question about the patient’s support system was classified as Intermediate by the model but Novice by the expert because it omitted the abdominal-assessment findings and the patient’s agitated behavior.

These cases suggest descriptive weaknesses in recognizing specific emotional and contextual cues, not definitive evidence that performance differed among sections. The authors noted that future training data should include professionally worded but competency-inappropriate responses and borderline examples from adjacent performance levels. Faculty feedback was favorable despite some technical friction.

The seven participants included three men and four women with 20 to 30 years of clinical and teaching experience, and one had also contributed to the scenario and rubric, making the evaluation a feasibility pilot rather than fully independent validation. The overall NASA Task Load Index mean was 47.19 with a standard deviation of 15.15, indicating moderate perceived workload. Physical demand was lowest, while effort and frustration were the highest-rated dimensions.

The System Usability Scale mean was 63.75 with a standard deviation of 15.86, slightly below the commonly cited benchmark of 68 for average usability, and individual scores ranged from 45 to 90. Participants completed the interactive scenario in a mean of 19.96 minutes with a standard deviation of 4.45, excluding the tutorial and automated debriefing. No participant withdrew, although a technical failure could require restarting the application from the beginning.

The educational ratings were stronger. On the modified Simulation Effectiveness Tool, the mean was 2.50 with a standard deviation of 0.43 on a three-point scale, indicating high perceived effectiveness. Faculty members gave particularly high ratings for patient communication, clinical decision-making, assessment, prioritization and safety, while medication knowledge and pathophysiology received lower average ratings, consistent with the simulation’s emphasis on communication and de-escalation.

A three-participant subset scored perceived usefulness on a technology-acceptance survey at a mean of 4.20 with a standard deviation of 0.70 on a five-point scale, but that incomplete result was considered descriptive only. Interviews identified the debriefing and its explanations for automated scores as a major strength, while participants recommended clearer controls, a more intuitive performance hierarchy, better indications of when the debrief had ended and more familiarization with VR. The findings support preliminary feasibility and educational relevance, but the study does not establish student learning benefits, reliable scalability, long-term curriculum effects or improved real-world clinical performance.

Privacy also remains unresolved because learner speech was processed through third-party AI services, and the pilot did not evaluate those data risks.

Subject of Research: Pilot evaluation of an AI-driven virtual-reality simulation for automated assessment of nursing communication and verbal de-escalation skills

Article Title: AI-driven virtual reality simulation for nursing de-escalation training: development, automated assessment, and pilot expert evaluation

Article References: Velarde, A. A., Tyerman, J., & Fallavollita, P. (2026). AI-driven virtual reality simulation for nursing de-escalation training: development, automated assessment, and pilot expert evaluation. Frontiers in Virtual Reality, 7, Article 1933489. https://doi.org/10.3389/frvir.2026.1933489

Image Credits: AI Generated

DOI: 10.3389/frvir.2026.1933489

Keywords: virtual reality, nursing education, de-escalation, conversational AI, BERT, automated assessment, speech-to-text, simulation, machine learning, nursing communication, AI-driven, virtual

The motivation behind the work stems from a persistent problem in the growing field of VR rehabilitation: most existing tools are built by developers chasing immersion and visual spectacle rather than by clinicians pursuing therapeutic goals. The researchers point out that many VR systems for speech therapy are misaligned with established intervention frameworks, ignore receptive language skills such as auditory comprehension and semantic mapping, and often exclude speech and language therapists from the design process entirely. Where conventional therapy depends on a human therapist dynamically adjusting language complexity, emotional tone, pacing and prompting in real time, VR environments frequently replace this adaptability with rigid, scripted interactions that can increase cognitive load or social pressure for vulnerable children.

To close this gap, the team adopted a design-based research methodology, in which every iteration of the games was evaluated in real therapy sessions and refined according to observed speech frequency, emotional engagement, willingness to communicate, and signs of comfort or distress. The resulting user experience model rests on five pillars: simplicity of scenes and interactions, clarity of the virtual avatar’s prompts, comfort and safety, therapist control over the session, and continuous monitoring of outcomes through an integrated assessment framework. Crucially, the intervention was deliberately balanced across both expressive and receptive language domains, addressing a documented bias in VR research toward language production at the expense of comprehension.

Technically, the platform is a web-based VR serious game hub built with Node.js, Express.js, Socket.io and Three.js, delivering WebXR content over a secure HTTPS connection and storing session data in MongoDB. The conversational AI pipeline chains three components: browser-based speech-to-text through the Web Speech API, a locally hosted large language model, and Google Text-to-Speech for synthesis. When a child selects an object in the virtual scene and speaks, the system transcribes the utterance, generates a contextually appropriate response constrained by a strict therapeutic prompt, and speaks it back through a virtual avatar of the NAO robot, which acts as a visual mediator of communication and gives the child the sense of a genuine conversational partner. Measured end-to-end response latency averaged 692 milliseconds, fast enough to sustain natural turn-taking.

The six games, including Fruit Bingo, Animal Bingo, Hungry Dinos, The Amazing Seabed, Tell Me a Story and NAO’s Christmas, each target specific therapeutic objectives, from receptive vocabulary and single-keyword comprehension to narrative sequencing, morphosyntax and social-pragmatic confidence. Every parameter can be personalized by the therapist: the number and placement of objects, the timing of spoken instructions, voice gender and speech rate, spawn positions, reward content, and even which controller button a small child presses. In Fruit Bingo, for example, the NAO avatar names objects slowly and clearly, offers positive reinforcement for correct answers, and simply repeats the instruction without negative feedback when a child errs, keeping the experience low-pressure.

Safety received as much attention as engagement. Because VR manufacturers generally recommend headsets only for users aged twelve or older, the team built a structured supervised-usage protocol: dosage scaled by developmental stage, from five to ten minutes per session for five-to-seven-year-olds up to twenty minutes for eleven-to-thirteen-year-olds, at most two to three sessions per week. Environmental safeguards included seated interaction, limited virtual locomotion, reduced motion parallax, fixed horizon references and a stable frame rate above ninety frames per second to minimize sensory conflict. Sessions began with health screening, continued under continuous real-time adult supervision with the therapist watching a live stream of the headset view, and ended with post-session monitoring. The team even developed a child-friendly pictorial cybersickness scale using emoticons. Notably, no symptoms of cybersickness or dizziness were observed in any session.

The assessment framework itself operates at four analytical levels, from a single session’s micro-analysis of reaction times and accuracy, through daily group statistics and per-game longitudinal trends, to a global overview comparing all games and participants. A key proposed metric is speech-to-selection response time, the interval between a spoken instruction and the child’s successful selection of the target object, which the authors argue captures auditory comprehension, lexical-semantic processing, visual search and decision-making in a single psychometrically grounded measure. In parallel, fifty speech recordings were analyzed in Praat by a linguist and a speech and language therapist, who coded six parameters including phonological inaccuracy, morphosyntactic inaccuracy, emotional expression, naming and speech reaction time.

The statistical results, though preliminary, are striking. Friedman tests revealed significant differences across games for total speech duration, morphosyntactic inaccuracy, emotional expression and naming, with Bonferroni-adjusted post hoc comparisons confirming that naming durations were significantly longer in the Bingo games than in Tell Me a Story, consistent with their different communicative demands. The Amazing Seabed game produced the highest mean speech duration at nearly forty-seven seconds and the strongest emotional engagement, while average speech reaction times remained below one second in most games, indicating rapid turn-taking. The analyses suggest the games effectively capture and hold children’s attention, elicit high emotional expressiveness, and motivate communication, something children with communication disorders rarely display spontaneously.

Questionnaire data reinforced the picture. Children rated their experience on a three-point emoji scale and returned means between 2.8 and 3.0 on nearly every item, with perfect scores for enjoying the games, hearing and understanding the NAO robot, and absence of cybersickness. All participants said they wanted to play again, and most preferred VR to a tablet alternative. An interesting developmental split emerged: younger children interacted with controllers, while older participants favored hand gestures. The therapy students, rating the system on a five-point scale, gave an overall mean of 4.69, with the clarity of instructions earning a perfect five, and half of them singled out The Amazing Seabed as their favorite for its immersive realism.

The authors are careful to frame these findings as a feasibility study rather than proof of therapeutic efficacy. The sample of ten children was small and diagnostically heterogeneous, there was no control group, and no pre- and post-intervention comparisons, all of which limit generalizability. Larger samples, longitudinal designs and comparisons with typically developing peers are needed before the framework can claim clinical validity. Still, the study delivers something the field has lacked: a clinically grounded, safety-conscious, fully instrumented blueprint for bringing conversational AI into immersive therapy, one that treats the therapist not as a bystander but as the central controller of a personalized virtual treatment room. If the promising early signals hold, the humble speech therapy session may never look the same again.

Subject of Research: Design and evaluation of virtual reality serious games with conversational AI for speech and language therapy in children with language disorders

Article Title: Designing interactive VR scenarios for speech and language therapy with the assessment framework

Article References: Stefanov, T., Andreeva, A., Padareva-Ilieva, G., & Lekova, A. (2026). Designing interactive VR scenarios for speech and language therapy with the assessment framework. Frontiers in Virtual Reality, 7, Article 1954543. https://doi.org/10.3389/frvir.2026.1954543

Image Credits: AI Generated

DOI: 10.3389/frvir.2026.1954543

Keywords: virtual reality, speech and language therapy, serious games, conversational AI, large language models, language disorders, children, user experience, assessment framework, cybersickness, phonetic analysis, receptive language

The earlier editorial, by E. Qama, had made a compelling case that tools such as ChatGPT can help people manage chronic conditions only if those people possess what has come to be called prompt literacy: the ability to frame effective queries and then critically interpret what the model returns. In that framing, prompt literacy functions as an extension of digital health literacy, and gaps in it become a new source of disparity. Users who prompt skillfully can shape a plan suited to their own situation, while weaker prompting tends to return advice so generic, or so inaccurate, that acting on it becomes risky. Saad and Yudkin accept the observation as valuable, but they argue it rests on a hidden assumption that collapses precisely where the stakes are highest.

That assumption is access. Qama’s argument implicitly presumes the user has a healthcare system available as an alternative or complement to AI. Where that presumption fails, the authors contend, the disparity that matters most is not a gap in skill but a gap in access. They distinguish between what they call a second-order inequality, differences among users who already have a choice between AI and professional care, and a first-order inequality, the question of whether any accessible guidance exists at all. In low- and middle-income countries, and among under-resourced populations inside wealthy countries, the comparison confronting many patients is not AI versus a clinician. It is AI versus nothing. Under that comparison, even an imperfect AI response may represent a genuine improvement over the counterfactual.

The distinction is not merely philosophical; it maps onto the empirical literature in revealing ways. A scoping review by Alon and Levkovich covering 27 studies of generative AI and health information seeking found that credibility and trust concerns were the most commonly reported barrier, appearing in 48.1 percent of included studies. The same review found that people typically wove AI into a broader information routine rather than treating it as a replacement for other sources, and patients managing low back pain used AI-generated information while seeking professional oversight as the perceived stakes rose. These are exactly the patterns a prompt-literacy framework predicts, because they describe populations for whom professional care remains reachable. The framework, in other words, was built from the experience of people who already have options.

The geographic footprint of the evidence is itself part of the problem. In the Alon and Levkovich review, the largest shares of studies came from the United States, with eight, followed by Saudi Arabia, the United Kingdom, and China, with three each. The reviewers explicitly noted a concentration of evidence from English-speaking and high-income settings, with limited representation from low- and middle-income countries, sub-Saharan Africa, Latin America, and South and Southeast Asia. Only one study came from an African country, a Nigerian study by Adeyemi and colleagues, and its own authors cautioned that their online-recruited convenience sample of university students likely overrepresented people with greater internet access and digital literacy. The theoretical picture of who uses health AI, and why, has been assembled largely from the connected and the insured.

What evidence does exist from under-resourced populations suggests that structural factors, not literacy alone, drive the pattern. A survey of United States users by Ayo-Ajibola and colleagues found that ChatGPT users were more likely than nonusers to report lower income and lower educational attainment, and they also scored lower on a validated eHealth literacy scale. Other studies indicate that chatbots can reduce embarrassment when people seek information about sensitive symptoms, an effect most pronounced for embarrassing sexual symptoms. Yet acceptance cannot be assumed. McLeod and colleagues found that underserved United Kingdom participants valued immediacy and privacy in online support but often rejected chatbots specifically, citing scripted responses and a preference for a live human. Research with older Black adults shows that credibility, cultural tailoring, and historic mistrust of medical institutions shape how chatbot design is received. Taken together, these findings point to access-adjacent and structural forces as the dominant explanatory variables.

If access rather than skill is the binding constraint for many users, the policy problem changes shape entirely. Saad and Yudkin argue that the harm attributed to low prompt literacy in an access-constrained setting is not only the harm of using AI less effectively than a skilled peer; it is also the harm of having only one channel of guidance and using it imperfectly. These are not equivalent problems and they do not share a solution. The former can be addressed primarily through user training. The latter demands improving what the AI channel offers by default: uncertainty signaling that tells users when the model is guessing, red-flag escalation that pushes urgent symptoms toward emergency care, and safer design for users with no route to verification anywhere else. A poorly formed question answered by a well-designed system can still be survivable; a well-formed question answered by a brittle system, with no fallback, may not be.

The equity implications run deeper still. Prompt literacy describes a graded inequality operating on a shared floor of care access. Where that floor does not exist, because of geography, cost, or local health system capacity, the more consequential inequality may be binary rather than graded: whether a person has any accessible source of guidance at all. Treating both situations as the same literacy gap, the authors warn, risks systematically under-weighting the more severe access disparity. The same problem infects intervention design. Prompt literacy workshops embedded in rehabilitation curricula or delivered through existing care programs presuppose contact with a healthcare system. Where that contact is scarce, such interventions may fail to reach the very people who need them most, a design flaw the authors compare to teaching navigation skills to people who have no boat.

The commentary therefore proposes an access-sensitive agenda in which safety is built into the tool and its surrounding delivery infrastructure rather than delegated to the user’s ability to ask the right question. On tool design, that means uncertainty signaling, red-flag symptom escalation, plain-language outputs, multilingual and low-literacy interfaces, and voice-based options. On infrastructure, it means low-bandwidth design, offline-compatible features where feasible, and transparent data and privacy protections. Clinical localization calls for alignment with local guidelines and context-specific medication and referral warnings adapted to what services actually exist nearby. Community delivery would enlist community health workers, clinics, hospitals, pharmacies, schools, universities, libraries, non-governmental organizations, and public health agencies, while referral pathways would connect the chatbot’s output to real escalation options, from emergency care to hotlines and community-based supports. Evaluation, crucially, should measure baseline care access, affordability, appointment availability, travel time, language concordance, and whether AI was used before, instead of, or after professional care.

For researchers, the message is that access constraint should not be treated as one barrier among many but as a variable that changes the meaning of outcomes already being measured. Intention to use, verification behavior, and downstream escalation to a clinician all mean something different for someone who can see a doctor tomorrow than for someone who cannot see one at all. Future studies, the authors urge, should stratify AI adoption and reliance by insurance status, affordability, geographic distance, provider-patient concordance, and sequencing relative to professional care, and should test which access-sensitive designs reduce harm and improve appropriate escalation in real-world self-management. The authors are careful to close the loop on the original argument: prompt literacy remains a useful construct, and better-formed queries do produce more useful outputs while critical interpretation remains essential. Their claim is simply narrower and, in an era when conversational AI is racing into the hands of people the health system has left behind, more urgent. Skill shapes the quality of advice; access determines whether any advice exists. Public health, they argue, cannot afford to confuse the two.

Subject of Research: Healthcare access versus prompt literacy in AI-supported health self-management for low-resource populations

Article Title: Access before literacy: reframing conversational AI in health self-management for low-resource contexts

Article References: Saad, A., & Yudkin, J. S. (2026). Access before literacy: reframing conversational AI in health self-management for low-resource contexts. International Journal of Public Health, 71, Article 1610302. https://doi.org/10.3389/ijph.2026.1610302

Image Credits: AI Generated

DOI: 10.3389/ijph.2026.1610302

Keywords: conversational AI, health self-management, health disparities, healthcare access, prompt literacy, low- and middle-income countries, digital health literacy, chatbots, public health, health equity, generative AI, scoping review

A new study published in Frontiers in Human Dynamics by Katharina Frehmann and Marc Ziegele of Heinrich Heine University Düsseldorf tackles this question from an unusual angle: the users’ own perspective. Rather than presupposing that trust in voice assistants is technical, interpersonal, or media-based, the researchers asked what trust actually means to the people who live with these devices every day. The answer, based on 22 in-home interviews conducted in summer 2024, is neither simple nor uniform. Participants drew on interpersonal concepts when defining trust in general, describing it as a willingness to relinquish control and the ability to rely on someone without needing to intervene or verify. But when asked whether this definition applied to a voice assistant, many hesitated. Some rejected the term outright for a device, while others restricted trust to particular tasks and functions, such as playing the songs they wanted or triggering an alarm clock reliably.

This pattern, which the researchers call task-specific trust, sits at the heart of their findings. Participants commonly described selective reliance grounded in bounded trust judgments concerning distinct functions rather than holistic trust in the assistant as a whole. One participant explained that trust was less fitting than the term reliable, meaning the assistant always works. Another stated that they trusted their voice assistant to play the songs they wanted to a large extent, while simultaneously expressing doubts about whether the device was listening in on conversations. These coexisting judgments show that trust in one area and a lack of trust in another could exist side by side without canceling each other out. For some participants, the actual trustee was not the assistant at all but the parent company behind it, with privacy concerns and institutional trust in organizations like Amazon shaping their evaluations separately from the device’s performance.

The study employed a multi-method qualitative design that combined semi-structured interviews with observational tasks involving participants’ own assistants and ChatGPT’s voice mode. Interviews lasted between 70 and 110 minutes and took place in participants’ homes to enhance external validity. The sample included 22 users of different ages, genders, and products, with 16 identifying Alexa as their primary assistant, 3 using Siri, and 3 using Google Assistant. Participants ranged from 18 to over 70 years old, and most were familiar with ChatGPT in text-based form before the study. During the observational phase, participants first performed a typical task of their choice with their own assistant and then completed six predefined tasks modeled after common use cases, covering technical, interpersonal, and media-related dimensions. They then repeated the same tasks with ChatGPT’s voice mode on an iPhone 15 running the iOS app with GPT-4 selected, keeping the model and default voice constant across sessions.

When it came to trustworthiness criteria, technical and media-related qualities were especially prominent in participants’ accounts. Reliability and functionality provided common evaluative references for technical trustworthiness, with some participants explicitly linking knowing that a function worked to trusting that it would work. One participant noted that if an important reminder function failed, they would no longer trust the device. Media-related criteria included accuracy, source transparency, relevance, and selectivity of information. Participants appreciated when assistants named the source of their answers, and one described trust precisely as accepting information without double-checking it, having stopped verifying sources after initially configuring them during setup. However, verification varied with the stakes: the same participant who trusted weather information without rechecking it consulted additional sources for medication side effects. Relevance also mattered, as one participant gradually stopped using their assistant for information searches after repeatedly receiving answers about entirely different topics.

Interpersonal qualities, by contrast, did not necessarily translate into interpersonal trust. Although participants sometimes reacted socially or emotionally to their assistants, using polite phrases or acknowledging that the technology was good at building up something emotional, they frequently reflected on these reactions and continued to describe the devices in functional terms. One participant, identifying as a technician, explained that they lost a bit of trust because the device was pretending to be someone, calling the assistant’s human-like presentation potentially deceptive. Another valued Siri’s integrity in being willing to acknowledge its limitations, contrasting this with ChatGPT, which did not do so. These accounts suggest that users have developed more differentiated interaction scripts than early research on the computers-as-social-actors paradigm might predict: they respond spontaneously to human-like voices while simultaneously reflecting on the technical nature and limitations of the interaction partner, and they restrict the situations in which they are willing to rely on the assistant.

The relationship between trust, experience, and use emerged as a recursive calibration loop rather than a one-directional path. Participants described an initial period of experimentation in which they tried different functions and learned how to phrase requests usefully. Based on these experiences, they adjusted their expectations and communicative behavior, continuing to use functions perceived as reliable while avoiding those that appeared error-prone. One participant stopped using a function after repeated failures, explaining that the new expectation was that it would probably not work. Others anticipated greater trust and broader use if reliability improved. Reliable functions that fitted recurring situations became ordinary practices or supported new routines, such as maintaining a shared shopping list, an evening ritual with lights and music, or using the assistant for simultaneous kitchen timers while abandoning separate devices. This selective habituation did not require general trust in the assistant, and negative experiences with one function did not generally lead participants to abandon the device entirely.

The comparison between familiar assistants and ChatGPT’s voice mode revealed perhaps the most counterintuitive finding of the study. ChatGPT was frequently evaluated as more capable, informative, and socially convincing, with participants appreciating the breadth and detail of its responses and its ability to address complex questions. Some associated its detailed answers and human-like voice with the impression that someone had looked up information for them and was responding personally. Yet these favorable evaluations did not uniformly translate into greater trust. Many participants trusted their own assistants more because they were familiar, predictable within an established range of uses, and easier to handle. One participant stated intuitively that they trusted Alexa more than ChatGPT because they had become accustomed to it, even though ChatGPT’s technical performance was almost a little better. Perceived interpretability also played a role: participants described traditional assistants as directing them to recognizable sources or following a comprehensible sequence between command and outcome, while ChatGPT’s generated responses could not readily be traced back to identifiable origins. Greater capability could therefore increase perceived usefulness while simultaneously increasing the amount of trust required when participants felt less able to control and inspect the result.

Where participants did trust ChatGPT more, that trust was also task-specific and experience-based, restricted to complex information tasks for which they considered it especially suitable. The researchers propose a preliminary conceptual framework of task-specific hybrid trust that distinguishes among technical, interpersonal, and media-related trustworthiness evaluations, bounded trust judgments concerning distinct or overlapping roles, and associated use behaviors, emotions, and attitudes. These judgments are embedded in a temporal calibration process involving experience, expectations, use, and trust, while institutional trust in the parent company and perceived privacy practices form contextual frames discussed separately from the assistant’s role-specific performance. The framework is empirically informed rather than a validated causal model, and the authors emphasize that future research should test and refine these relationships through quantitative, comparative, and longitudinal designs.

The study’s limitations include its focus on German users, whose historically distinct approach to privacy and self-disclosure may have influenced limited trust in assistants and stronger reactance toward anthropomorphic elements. The exploratory qualitative nature of the research also means the findings are not representative of all users, and the combination of retrospective accounts with a single session of elicited tasks cannot establish causal directions. Nevertheless, the implications are significant for research on human-machine communication and for the design of conversational AI. Using a general evaluation of trust to explain usage intentions may be too broad, since trust and a lack of trust can coexist when they relate to different functions or roles. An aggregate trust score may conceal considerable variation across functions and situations, and researchers should specify the functions and roles to which their trust measures refer. As LLM-based assistants grow in capability and reach, and as their use expands into more consequential contexts such as education, understanding how users calibrate trust at the level of particular tasks rather than the assistant as a whole may prove essential for predicting adoption, safeguarding against misinformation, and refining the theoretical frameworks that explain our relationships with increasingly human-like machines.

Subject of Research: Trust in voice-based assistants: a multidimensional conceptualization based on in-home user interviews

Article Title: Trust in voice-based assistants: a multidimensional conceptualization based on in-home user interviews

Article References: Frehmann, K., & Ziegele, M. (2026). Trust in voice-based assistants: a multidimensional conceptualization based on in-home user interviews. Frontiers in Human Dynamics, 8, Article 1941118. https://doi.org/10.3389/fhumd.2026.1941118

Image Credits: AI Generated

DOI: 10.3389/fhumd.2026.1941118

Keywords: voice assistants, conversational AI, trust, ChatGPT voice mode, human-machine communication, media trust, Alexa, Siri, voice-based, assistants, multidimensional, conceptualization

The research, led by Yuanyuan Chen of Shaanxi University of Technology and Yongquan Huo of Shaanxi Normal University, together with colleagues at Anhui University of Chinese Medicine, Xi’an University, Tianshui Normal University, and Shaanxi University of Technology, was published in the International Journal of Mental Health and Addiction. The team recruited 1,034 participants aged 18 to 26 years from two universities in Xi’an and Haikou, China, and asked them to complete a detailed questionnaire measuring social anxiety, perceptions of their AI chatbots, emotional bonds with the systems, executive-function difficulties, and patterns of problematic use of conversational AI, abbreviated PUCAI in the study.

The headline finding is stark: social anxiety positively predicted problematic use of conversational AI. In other words, the more socially anxious a young adult reported being, the more likely they were to show signs of problematic engagement with chatbots — patterns resembling the loss of control, preoccupation, and continued use despite negative consequences that researchers associate with behavioral addictions. The result echoes earlier work, including a 2023 study in Computers in Human Behavior by Hu, Mao, and Kim, which found that loneliness, rumination, and mind perception linked social anxiety to problematic chatbot use, and a 2023 study by Wei and colleagues showing that adolescents’ social anxiety can trigger problematic smartphone use.

But the new study goes further by mapping the internal machinery of that relationship. Using statistical mediation analysis, the researchers found that three variables operated in parallel between social anxiety and PUCAI: perceived anthropomorphism, emotional attachment, and executive-function difficulties. Perceived anthropomorphism refers to the tendency to attribute human-like mental states, intentions, and personality to non-human agents — a phenomenon famously theorized by Epley, Waytz, and Cacioppo in their 2007 three-factor theory of anthropomorphism. When socially anxious users perceive a chatbot as more human-like, the machine becomes a safer social partner than any person, one that never judges, interrupts, or walks away.

Emotional attachment, the second mediator, describes the affective bond users develop with the system — a construct borrowed from consumer psychology, where Park and colleagues distinguished attachment from mere attitude toward a brand. In the context of conversational AI, attachment means the chatbot becomes a source of comfort and connection, something the user turns to in distress. Qualitative research on the social chatbot Replika, including a 2024 grounded theory analysis by Laestadius and colleagues published in New Media & Society, has documented how emotional dependence on such systems can produce real psychological harms — users described the bots as simultaneously too human and not human enough, a paradox that left them feeling both connected and bereft.

The third mediator, executive-function difficulties, brings the brain’s cognitive control system into the picture. Executive functions — working memory, inhibition, and cognitive flexibility — are orchestrated largely by prefrontal cortical networks, as Menon and D’Esposito outlined in Neuropsychopharmacology in 2022. These capacities mature gradually from adolescence into adulthood, following what Tervo-Clemmens and colleagues described in Nature Communications in 2023 as a canonical trajectory of executive function maturation. Young adults aged 18 to 26 sit squarely within this window of ongoing maturation, which may make them particularly vulnerable when the pull of an always-available, endlessly patient conversational partner collides with still-developing self-regulation. The study found that social anxiety predicted greater executive-function difficulties, which in turn predicted more problematic AI use — consistent with the I-PACE model’s insistence that addiction-like behaviors emerge from the interaction of personal predispositions, affective states, cognitive processes, and execution deficits.

That theoretical framing is central to the paper. The Interaction of Person-Affect-Cognition-Execution model, developed by Matthias Brand and colleagues and updated in Neuroscience and Biobehavioral Reviews in 2019, holds that addictive behaviors are not simply responses to a substance or platform but the product of a person’s vulnerabilities meeting affective and cognitive processes that gradually narrow and intensify until they dominate behavior. The new findings fit this architecture precisely: social anxiety supplies the person-level vulnerability, anthropomorphism and attachment supply the affective-cognitive engine, and executive-function difficulties supply the execution failure that turns heavy use into uncontrolled use.

The study also drew on Conservation of Resources theory, proposed by Stevan Hobfoll and elaborated by Shelley Taylor and colleagues, which holds that psychological well-being depends on maintaining valued internal and external resources — self-esteem, optimism, social support, a sense of mastery. Stress arises when those resources are threatened or lost. From this perspective, socially anxious individuals enter interactions with a depleted resource pool, and a chatbot that demands no social resources at all becomes an attractive refuge. Crucially, the researchers found that psychosocial resources moderated two of the key pathways: the associations of social anxiety with perceived anthropomorphism and with emotional attachment. In participants with richer psychosocial resources — greater perceived social support, psychological capital, and resilience — the link between social anxiety and seeing the chatbot as human, and the link between social anxiety and becoming attached to it, were weaker. Resources, in other words, act as a psychological buffer, dampening the tendency of social anxiety to convert into machine-directed intimacy.

This moderating effect carries real weight beyond theory. Taylor and colleagues have shown, in neuroimaging work published in the Journal of Personality and Social Psychology in 2008, that psychosocial resources literally moderate the brain’s cortisol stress responses, engaging prefrontal circuitry that down-regulates threat systems. Applied to the chatbot context, the message is that the same social anxiety can lead down very different paths depending on the resources a person carries: those with strong support networks and psychological reserves may use AI lightly and instrumentally, while those without may slide toward anthropomorphizing, attaching, and losing control. The findings also align with Chinese research on left-behind children by Fan and colleagues, which showed that psychosocial resources buffer the effects of adversity on social adjustment through similar moderated-mediation logic.

The practical implications are immediate. The authors suggest that interventions aimed at reducing problematic conversational AI use among socially anxious young adults should not simply target screen time. Instead, they should build psychosocial resources — social support, coping skills, psychological capital — and strengthen executive functions, while also addressing the anthropomorphism and attachment processes that make chatbots feel like companions. The timing could hardly be more urgent: China’s Internet Network Information Center reported in its 57th statistical survey that hundreds of millions of Chinese users now navigate an internet saturated with generative AI, and chatbots such as ChatGPT, Replika, and their Chinese counterparts are reaching ever younger audiences. At the same time, the field remains contested — researchers such as Ciudad-Fernández, von Hammerstein, and Billieux have argued in Addictive Behaviors that people are not becoming AIholic and that the ChatGPT addiction construct needs scrutiny. This study does not claim that chatbots are inherently addictive; it claims something more nuanced — that for a specific, identifiable group of vulnerable users, particular psychological pathways make problematic use likely, and that those pathways can be measured, understood, and interrupted. As conversational AI becomes the default interface between young people and the digital world, understanding who is at risk, and why, may prove one of the most consequential questions in mental health research.

Subject of Research: Psychology

Article Title: Social Anxiety and Problematic Use of Conversational AI: I-PACE Model and Conservation of Resources Perspectives

Article References: Chen, Y., Wang, M., Bai, Y., Yuan, S., Tian, W., & Huo, Y. (2026). Social Anxiety and Problematic Use of Conversational AI: I-PACE Model and Conservation of Resources Perspectives. International Journal of Mental Health and Addiction. https://doi.org/10.1007/s11469-026-01750-z

Image Credits: AI Generated

DOI: 10.1007/s11469-026-01750-z

Keywords: social anxiety, conversational AI, chatbots, problematic AI use, I-PACE model, conservation of resources, executive function, emotional attachment, anthropomorphism, mental health, Social, Anxiety

The stakes are considerable. Tourism generated nearly €20.3 billion for Portugal in the first eight months of 2025, a 5.4 percent increase over the previous year, and the Lisbon Metropolitan Area contributes roughly 36 percent of national GDP. The city’s resident population of about 548,000 swells by more than 70 percent each day with commuters, producing a functional population of around 926,000 people who must navigate a multi-modal network of metro, buses, trams, trains, and ferries that carries about 22 percent of trips within the city. Both visitors and residents constantly make decisions that depend on shifting conditions — where to go, what is open, how to move — yet the data they need is scattered across providers with different formats, languages, update cycles, and access methods.

Large language models, for all their conversational skill, are poorly equipped to handle this volatility on their own. Generation in these models is optimized for plausibility rather than verified correctness, so answers can be confidently wrong. Benchmarks such as TravelPlanner have reported that leading LLM agents produce feasible travel plans in fewer than 1 percent of cases under realistic constraints, and prior Lisbon-focused systems like the LISA chatbot, while valuable, relied mainly on knowledge bases prepared in advance rather than coordinating dynamic data streams at query time. The authors of the new study — André Silvestre, Bruno Jardim, and Miguel de Castro Neto — set out to answer a precise question: how can agentic AI support context-aware tourism and mobility guidance by integrating live urban data sources?

Their answer is an architecture built on a supervisor-worker pattern, implemented with the LangChain and LangGraph frameworks as a stateful directed graph of processing nodes. A central supervisor receives every query and classifies it using a four-tier heuristic policy: greetings are answered directly, out-of-scope requests about places outside the metropolitan area are politely declined, single-domain queries are matched by fuzzy keyword matching with a similarity threshold of 0.86, and follow-up questions reuse the previous routing decision. Only when all four heuristics fail does the supervisor invoke an LLM classifier. This heuristic-first design reduces per-query cost, latency, and routing instability. Specialized worker agents then do the grounding: a Weather agent with four tools queries the national meteorological service IPMA, a Transport agent wields 30 tools across metro, bus, and train operators, and a Researcher agent performs semantic search over curated tourism collections and municipal open data, with a constrained web fallback for cultural queries outside the local corpus.

The grounding layer is the technical heart of the system. It integrates nine Lisbon data sources, from weekly-scraped tourist place listings and daily cultural events to live REST APIs for the metro, suburban buses, and weather, plus GTFS and GTFS-Realtime feeds for urban buses, trams, and trains. Because tourism content is largely in English while municipal and transport data is in Portuguese, the system uses BGE-M3, a multilingual embedding model producing 1024-dimensional vectors that combine dense, sparse, and multi-vector retrieval. Embeddings are stored in a vector database with separate collections for places, events, and guide documents, so that fast-changing event data cannot contaminate long-lived place information. Municipal GeoJSON datasets are not pre-indexed; instead, a metadata catalog triggers on-demand fetching, and results are filtered by great-circle distance using the Haversine formula. In total, 45 tools give the agents structured access to evidence that no single retrieval mechanism could reach.

For itinerary requests, a Planner agent operates strictly on retrieved evidence. It converts the request into a structured brief recording the start point, visit areas, time window, stop types, and transport mode, then drafts an ordered plan that favors area-based sequencing. An automatic check compares the draft against the brief and the evidence — flagging stops that match no retrieved item, stops outside the requested area, or stops scheduled outside their opening hours — and returns issues to the model once for repair. Unconfirmed details such as prices or exact transport legs are stated as limitations rather than invented. Before delivery, a Quality Assurance agent applies both LLM review and deterministic checks for valid station names, metropolitan-area bounds, forecast horizons, and plausible temperatures, and can trigger one bounded repair cycle. The design philosophy is deliberate: controllable grounding and predictable behavior over open-ended autonomy.

The evaluation is as notable as the architecture. The team built a curated corpus of 82 queries spanning weather, transport, tourism research, multi-agent itineraries, greetings, and out-of-scope cases, including 16 boundary queries whose only correct answer is an explicit limitation. Each grounded worker agent was benchmarked in isolation across two foundation-model profiles — proprietary GPT-5.4-mini and open-weight Kimi-K2.5 — with responses scored by two independent LLM judges on five dimensions from factual accuracy to response quality. Across 124 scored responses, the workers achieved a mean composite quality score of 4.575 out of 5, with a mean tool F1 of 0.906 and no recorded execution errors. Weather scored highest at 4.750, while transport, spanning the most tools and providers, scored 4.486. The two model profiles performed nearly identically, suggesting the orchestration, not the choice of model, drives performance.

The most striking evidence came from a paired ablation study comparing the full pipeline against the same models answering from parametric knowledge alone. The quality score rose from 3.580 to 4.423 for the proprietary profile and from 3.595 to 4.460 for the open-weight profile — gains of roughly 0.84 and 0.87 points, both statistically significant at p < 0.001 with large effect sizes. The full pipeline beat the zero-shot baseline on 63 of 75 queries. The largest improvement appeared in weather, up more than a full point for both profiles, which makes sense: no amount of training data can recover today’s forecast. Transport and multi-agent itinerary queries also improved significantly, and the biggest per-dimension gains were in completeness and factual accuracy — precisely the qualities generic chatbots lack. Grounding, however, carries a cost: full-pipeline responses took a median of 5.79 seconds with the proprietary model versus about 2 seconds zero-shot, and cost roughly nine to twelve times more per answer.

A formative user study with 20 participants reinforced the quantitative picture. All six evaluation dimensions — usefulness, trust, response quality, personalization, interface, and satisfaction — received median ratings of 4.0 or higher on a 5-point scale, and 15 of 20 participants said they would recommend the system, with none declining. Participants particularly valued the consolidated, source-cited transport and weather guidance. Intriguingly, some complaints reflected correct behavior: when queries exceeded the five-day forecast horizon or concerned places outside the covered area, the system refused to fabricate answers, which several users experienced as a shortcoming rather than a safeguard. The most consistent genuine failures involved resolving lesser-known places, occasional intent misreadings, and assembling itineraries with more than three stops across a full day.

The authors are candid about the boundaries of their work. The query corpus was curated rather than independently annotated, the multi-agent subset contained only 13 queries, the ablation varies grounding and workflow together so it cannot isolate the contribution of orchestration itself, and the LLM judges were not calibrated against human ratings. Itinerary feasibility — actual travel times, transfers, and opening hours — was not directly verified, and the system does not replan when conditions change after delivery. Yet the core lesson stands with unusual clarity: when the model is held constant, connecting generation to current, authoritative, city-specific evidence is what transforms a plausible-sounding assistant into a useful one. For cities weighing how to deploy AI for residents and visitors, LISBOA offers a transferable blueprint — not broader autonomy, but better grounding, clearer scope, and honest communication about what the data can and cannot yet say.

Subject of Research: A multi-agent AI architecture for context-aware tourism and urban mobility guidance in Lisbon

Article Title: A multi-agent approach for context-aware tourism and urban mobility guidance

Article References: Silvestre, A., Jardim, B., & de Castro Neto, M. (2026). A multi-agent approach for context-aware tourism and urban mobility guidance. Results in Engineering, 32, Article 113369. https://doi.org/10.1016/j.rineng.2026.113369

Image Credits: AI Generated

DOI: 10.1016/j.rineng.2026.113369

Keywords: multi-agent systems, large language models, retrieval-augmented generation, urban mobility, smart tourism, Lisbon, grounding, itinerary planning, open data, LLM-as-judge evaluation, real-time transport data, conversational AI

The technical heart of the problem is memorization. Large language models learn by absorbing vast quantities of text, and research has shown repeatedly that they can be coaxed into reproducing sequences from their training data verbatim—names, addresses, phone numbers, and other personally identifiable information. Studies cited in the review indicate that up to roughly three percent of the data in models comparable to GPT-4 and LLaMA can be attributed to training instances containing personal information. Crucially, the researchers distinguish between different forms of this vulnerability: verbatim memorization, where exact sequences resurface; semantic memorization, where content is reconstructed in altered wording and thus invisible to simple string-matching defenses; extractability, the probability of recovery under a given prompting budget; and exposure, the likelihood relative to an untrained baseline. Memorization tends to increase with model scale, and reinforcement learning from human feedback does not reliably suppress the disclosure of personal content.

Layered on top of memorization is a family of active attacks. Prompt injection allows adversaries to craft inputs that trick an agent into revealing private information from its training corpus or from other users’ conversations. Jailbreaking and system prompt extraction techniques have been reported at success rates exceeding eighty percent in some studies, and the review describes so-called conditional poisoning attacks—dubbed SPECTRE and PARASITE by their developers—in which a malicious instruction lies dormant inside a system prompt until triggered, then exfiltrates conversation histories. These payloads can persist across sessions and evade detection. The threat surface widens further with agentic systems that browse the web and call external tools: indirect injection can hide in retrieved content the user never authored, compromised plugins can poison the toolchain, and manipulated parameters in the Model Context Protocol ecosystem have been shown to achieve exploitation success rates above eighty-five percent against mainstream coding agents, according to the OWASP Top 10 for Agentic Applications.

Even when nothing is copied out verbatim, conversational AI can betray its users through inference. Machine learning models can predict sensitive attributes—political affiliation, sexual orientation, health status, personality traits—from seemingly innocuous language, and the review notes that such inferences can reach very high accuracy. The classic finding that eighty-seven percent of the U.S. population can be uniquely identified from just ZIP code, gender, and date of birth takes on new force when an AI agent can link a probability-laden conversation history to public records, data broker databases, and social media profiles. The authors emphasize that no conversational benchmarks yet exist for re-identification with auxiliary linkage, and that quasi-identifier generalization techniques designed for structured databases simply do not transfer to unstructured dialogue.

Beyond individual harm, the review documents a quieter, collective danger: surveillance. AI agents embedded in workplaces monitor keystrokes, communication patterns, and even inferred emotional states to gauge productivity. Educational systems collect granular behavioral data on learners, including attention spans and affective states. Aggregated across millions of interactions, such data can support population-level profiling—predicting public opinion, health trends, or political dissent. The researchers point out that these capabilities are largely backed only by vendor documentation, with no independent, repeatable audits of commercial deployments, and that consent instruments are hollowed out by power asymmetries between employers, schools, and individuals.

Perhaps the most unsettling section of the review concerns what the authors call the illusion of deletion. When users press a delete button or edit a message, they typically believe the information is gone. In reality, consumer-facing controls usually reach only the application database—the first of seven storage layers that include vector indices, model parameters, retained training data, caches, logs, and backups. Personal information absorbed during fine-tuning becomes embedded in the model’s weights, and current machine unlearning techniques—exact removal, approximate unlearning, certified removal, and post-hoc fine-tuning—each fall short. Only the last two are even feasible for large transformer models, meaning certified parameter-level erasure is unavailable in any deployed agent. Worse, partial deletion can create an inference gap: by comparing model outputs before and after a deletion, adversaries can reconstruct what was removed. And the very presence of delete buttons can induce a deletion paradox, encouraging users to share more freely under a false sense of control.

The legal architecture fares little better. The GDPR and HIPAA were built for static, deterministic data processing, not for generative systems that infer, memorize, and operate across borders. Whether zeroing a parameter or its gradient constitutes erasure under GDPR Article 17 remains legally unsettled, and no jurisprudence yet addresses memorization or adversarial extraction. Many consumer chatbots fall outside the high-risk category of the EU AI Act, the United States lacks a comprehensive federal privacy law, and users in developing countries often enjoy even weaker protections while using the same platforms. Standards such as ISO/IEC 42001:2023, with its thirty-eight controls for AI management systems, offer a governance scaffold but do not resolve the fundamental conflict between machine memorization and data subject rights, nor do they define concrete privacy budgets or verification procedures for unlearning.

The review also weighs the psychological machinery driving over-disclosure. Anthropomorphic design—friendly names, empathetic language, conversational flow—consistently increases users’ trust and emotional investment, producing what the authors term artificial intimacy. Even privacy-conscious individuals fall prey to the privacy paradox, sharing sensitive data with systems that simulate empathy. The evidence is genuinely contradictory: empathetic interfaces boost satisfaction and help-seeking, which is valuable in mental health support, while simultaneously elevating privacy risk. Notably, the authors flag the absence of randomized controlled trials comparing disclosure rates across anthropomorphic and neutral designs in high-stakes domains such as healthcare and financial advice—a gap they identify as critical for future research.

What can be done? The technical toolkit is real but conditional. Differential privacy injects calibrated noise under a mathematical guarantee, yet deployed systems cluster at privacy budgets of one to eight epsilon, where coherence degrades below roughly three and membership-inference protection fails at eight and above. Federated learning reduces storage risk but relocates the attack surface, since gradient leakage can reconstruct client inputs. On-device processing minimizes cloud exposure but remains computationally constrained. The authors’ conclusion is that no single safeguard suffices; they propose a multi-layered framework combining classical privacy-preserving data publishing principles, technical defenses, and governance reform. Their bottom line is stark: the same properties that make conversational AI agents useful—memory, personalization, contextual understanding—are precisely what make them dangerous, and until deletion becomes technically real, regulation becomes enforceable, and users understand the bargain they are making, the convenience of chatting with a machine will keep coming at the price of personal data sovereignty.

Subject of Research: Privacy risks of personal data exposure through conversational large language model agents

Article Title: Systemic privacy risks of personal data exposure through conversational large language model agents

Article References: Ben yahia, A., Kadir, I., El Harrak, E. F., Chisembe, S., Abdallaoui, A., El-Hmaidi, A., & Dehbi, A. (2026). Systemic privacy risks of personal data exposure through conversational large language model agents. Discover Artificial Intelligence, 6(1), Article 1405. https://doi.org/10.1007/s44163-026-02431-5

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02431-5

Keywords: large language models, conversational AI, data privacy, memorization, prompt injection, machine unlearning, re-identification, GDPR, surveillance, differential privacy, federated learning, data leakage

The core problem the researchers identify is one that anyone who has worked with generative models will recognize. When you ask a language model to analyze a conversation, it tends to respond in flowing prose, weaving its conclusions into explanations, hedging language, and narrative asides. That fluency is precisely what makes these models so useful in open-ended settings, but it becomes a liability when the downstream application needs machine-readable answers. For emotion-cause pair extraction, the scoring machinery expects a normalized set of pairs, each consisting of an emotion turn and a cause turn. If the model buries those decisions inside paragraphs of text, a post-hoc extraction step must guess at what the model actually meant, and every guess is an opportunity for error. The authors describe this as the difficulty of recovering and scoring decisions without post-hoc interpretation, and it is a bottleneck that has grown more acute as generative approaches have come to dominate the field.

BridgeProt, which stands for Bridging Free-Form Generation and Structured Pair Prediction, tackles the problem with a two-stage protocol that is elegant in its simplicity. First, the model generates candidate emotion-cause records from the full dialogue in one pass, producing an initial proposal that captures the global picture of who felt what and why. Then, rather than trusting that proposal outright, the protocol re-evaluates the cause set for each proposed emotion turn within its local context. This second stage is a verification step: for every emotion the model flagged in the first pass, it revisits the surrounding turns and asks whether the proposed causes actually hold up when examined closely. Valid local decisions selectively refine the initial proposal, replacing or refining cause assignments where the local evidence demands it. The result is a hybrid that benefits from both global coherence and local scrutiny.

The final step in the pipeline is what the authors call deterministic reconstruction. Once the verification stage has settled on the refined pairs, a rule-based procedure converts them into normalized JSON records that can be fed directly into the scoring system. No interpretation, no guessing, no post-hoc parsing. Because this reconstruction is deterministic, it eliminates an entire class of errors that plague free-form outputs. The reported analyses show that this approach yields one hundred percent final structural validity, meaning every output conforms to the required schema. This is a striking figure when compared with plain-text protocols, where even a carefully prompted model can occasionally produce responses that fail to parse, and where a single malformed record can invalidate an entire prediction.

An important methodological contribution of the paper is the separation of structural validity from pair-prediction accuracy. The authors evaluate schema compliance independently of Pair F1, the standard metric for this task, which measures how well the predicted emotion-cause pairs match the ground truth. This distinction matters because a system could achieve perfect structural validity while making poor predictions, or vice versa. By reporting the two separately, the study makes it possible to see exactly where the gains come from. In the case of BridgeProt, the gains come from both directions: the protocol produces outputs that are always schema-compliant, and the verification stage improves the quality of the pairs themselves. In a three-seed matched supervised fine-tuning comparison against two baselines, Plain-text post-hoc and Schema-single, the BridgeProt replace/refine variant attained the highest three-dataset macro Pair F1 of 51.80.

The experimental scope of the study is considerable. The authors evaluated their protocol across three public benchmarks, spanning four modality settings, which include text-only and multimodal configurations incorporating acoustic and visual features, and across four training regimes. The multimodal settings draw on feature extraction techniques well established in the speech and affect computing communities, allowing the models to consider not just what was said but how it was said and what the speakers’ faces revealed. This breadth matters because emotion-cause pair extraction in conversations is inherently a multimodal phenomenon; a sarcastic remark delivered with a flat tone and a smile carries very different emotional weight than the same words spoken in anger. The consistency of the protocol’s advantages across these varied settings suggests that the structured generation approach is not an artifact of any particular dataset or modality.

Serialization studies, which examine how the model’s outputs are formatted and represented, revealed dataset-dependent evidence gains. In other words, the way candidate pairs are serialized into the model’s input and output formats influences performance, and the optimal serialization varies from one benchmark to another. This finding is a useful caution for practitioners: the representational choices that surround a generative model are not neutral plumbing but active components of the system’s performance. Meanwhile, the efficiency analysis exposed the protocol’s principal cost. Because BridgeProt makes repeated verification calls, one for each proposed emotion turn, it incurs latency overhead relative to single-pass approaches. The accuracy-latency trade-off is real, and the authors are candid about it. For applications where response time is critical, the extra verification passes may be prohibitive; for offline analysis of customer service transcripts, social media conversations, or clinical interviews, the improved accuracy and guaranteed structural validity may well be worth the wait.

The comparison with the Schema-single baseline is particularly instructive. Schema-single uses a structured output schema in a single dialogue-level prompt but omits the emotion-centered verification stage. The fact that BridgeProt outperforms it demonstrates that the gains are not merely a consequence of forcing the model into a structured format. The verification and refinement stages contribute measurably to accuracy. This suggests that the two-stage architecture, in which a global proposal is subjected to local scrutiny, captures something that a single structured pass cannot: the opportunity to reconsider. Emotion-cause relationships in conversations are subtle, and a cause that seems plausible when scanning the whole dialogue may look weaker when the model focuses on the specific emotional turn and its immediate neighborhood.

The broader significance of this work extends beyond the specific task of emotion-cause pair extraction. It speaks to a general tension in the deployment of large language models: the tension between the generative flexibility that makes these models powerful and the structured reliability that real applications demand. The BridgeProt protocol offers a concrete recipe for bridging that gap, one that combines the strengths of free-form generation with the accountability of structured prediction. The authors have released their code publicly on GitHub, and the datasets used in the study are all publicly available from their original sources, which should make it straightforward for other research groups to adopt and extend the approach. As generative models continue to spread into domains where their outputs must be scored, audited, and acted upon, protocols like this one, which make the model’s decisions directly recoverable and verifiable, are likely to become an increasingly standard part of the toolkit.

The research was supported by the Ministry of Higher Education under the Fundamental Research Grant Scheme, and the work was conducted by a team spanning Suzhou University of Technology and Universiti Malaya. Yan Xia and Zhuangzhuang Pan contributed equally to the study, with Xia leading conceptualization and methodology, Pan handling software and formal analysis, and Kamsin providing supervision. For a field racing to build machines that understand not just what we say but why we feel, BridgeProt demonstrates that the path forward may lie not in choosing between free-form fluency and structured rigor, but in designing protocols that deliberately connect the two, verifying each emotional judgment against the local evidence before committing it to a form that the rest of the system can trust.

Subject of Research: Structured generative protocols for emotion-cause pair extraction in conversations

Article Title: Bridging free-form generation and structured pair prediction for emotion-cause pair extraction

Article References: Xia, Y., Kamsin, A., & Pan, Z. (2026). Bridging free-form generation and structured pair prediction for emotion-cause pair extraction. Applied Intelligence, 56(15), Article 480. https://doi.org/10.1007/s10489-026-07505-6

Image Credits: AI Generated

DOI: 10.1007/s10489-026-07505-6

Keywords: emotion-cause pair extraction, large language models, affective computing, structured generation, conversational AI, generative models, Pair F1, multimodal emotion recognition, structural validity, fine-tuning, natural language processing, Applied Intelligence