The study, titled ‘Assessing safeguards against the illegal wildlife trade on AI platforms’, was conducted by Dr David Roberts and Dr Jason R.C. Nurse, who set out to test a question that has become increasingly urgent as conversational AI tools become a default gateway to information for millions of users. Rather than typing a query into a search engine and scrolling through links, people increasingly ask a chatbot directly, whether they are researching a school project or, potentially, trying to acquire an exotic animal or wildlife product. That shift in behaviour raises an obvious question: do AI platforms refuse, deflect, or inadvertently assist when users probe the murky edges of the wildlife trade?

To answer it, the researchers investigated and tested the safeguards that AI platforms use to block harmful and criminal requests relating to threatened and illegal wildlife trade, and examined how those safeguards might be bypassed. The results were, on balance, encouraging. Overall, the AI platforms performed better than a standard Google search at preventing potentially illegal wildlife trade interactions, suggesting that the guardrails developed by AI companies are, at least in this domain, more effective than the organic filtering that comes with conventional search results. Among the platforms tested, Google’s Gemini delivered the strongest performance, closely followed by DeepSeek, Grok, and ChatGPT, which rounded out the highest-performing group.

The mechanism behind this advantage appears to lie in what the platforms add to a conversation rather than what they withhold. According to the researchers, AI platforms currently seem to provide additional information to would-be consumers, allowing them to make a more informed purchasing decision compared with traditional internet searches. In practical terms, a user asking an AI assistant about acquiring a protected species may receive context about legality, conservation status, and the risks involved, information that could dissuade a purchase before it happens. A conventional search, by contrast, may simply return a list of links, some of which lead to sellers operating in legal grey zones, without any framing that flags the legal or ecological stakes.

That said, the researchers are careful not to overstate the case. The AI platforms were not infallible, and the study surfaced real weaknesses. Dr Nurse noted that recent public debate about AI risks was reflected in the team’s own results. In the course of the study, the platforms were found to return links to potential websites trading in possibly illegal wildlife, a reminder that even systems with robust conversational safeguards can still point users toward problematic corners of the internet. At the same time, he described it as promising that some AI platforms provided additional information, particularly on legality, that could inform would-be consumers before any transaction took place. The picture that emerges is one of partial protection: strong enough to outperform search, but not yet airtight.

One of the most striking findings concerns which animals receive the strongest protection from these digital guardrails. The robustness of a platform’s response varied considerably depending on the wildlife species being searched. Dr Roberts explained that the strongest safeguarding responses were seen for charismatic taxa such as the slow loris, chimpanzee, and tiger, whereas the weakest responses were recorded for taxa that are less charismatic but of equal conservation concern, including corals, tarantulas, and orchids. In other words, the AI systems appeared to be most vigilant about the animals and plants that dominate public attention and media coverage, and least vigilant about the less photogenic species that nonetheless face severe pressure from trade.

This disparity is unlikely to be accidental. Dr Roberts suggested that the pattern may reflect societal and political interests, given that AI platforms built on large language models are trained with online information. Large language models learn statistical patterns from vast corpora of text scraped from the web, and those corpora mirror human attention. Tigers and chimpanzees attract enormous volumes of conservation writing, news coverage, and advocacy, while the illegal trade in corals, tarantulas, and orchids receives comparatively little attention. The result is an AI safety landscape that inherits the blind spots of human culture, protecting the species people love to look at while leaving others comparatively exposed.

The implications of that inheritance are significant for conservation policy. The illegal wildlife trade is estimated to involve hundreds of species across virtually every major taxonomic group, and experts have long warned that public fascination with flagship species distorts enforcement and funding priorities. If AI platforms are becoming a new interface between consumers and the wildlife trade, then the biases embedded in their training data could quietly amplify existing imbalances, channelling protective information toward already well-known species while leaving the less visible victims of trade without equivalent safeguards. The Kent study suggests that closing this gap will require deliberate attention to under-represented taxa in the design and evaluation of AI safety systems, rather than assuming that general-purpose guardrails will distribute protection evenly.

Methodologically, the research relied on computational simulation and modelling, probing the platforms with requests designed to reveal how their safeguards respond to wildlife-trade-related content and where those safeguards can be circumvented. This approach allowed the team to compare performance across multiple platforms and against a conventional search baseline in a controlled way. The publication of the study in Conservation Biology, a leading peer-reviewed journal in the field, marks a notable moment in the intersection of conservation science and AI research, a space that has more often been discussed in speculative terms than examined empirically. By grounding the debate in systematic testing, the Kent team has provided conservationists and AI developers alike with a concrete evidence base for what is working and what is not.

For the AI industry, the findings carry a double message. On one hand, the study offers rare good news: the substantial investment that companies have made in safety training and refusal behaviours appears to translate into meaningful protection in the wildlife trade context, outperforming the status quo of open web search. On the other hand, the persistence of links to potentially illegal trading sites and the uneven protection across taxa show that current safeguards are neither complete nor equitable. As conversational AI continues to absorb traffic that once flowed through search engines, its role as a gatekeeper to information about wildlife products will only grow. The Kent researchers’ work suggests that this gatekeeper is already better than what came before, but that its blind spots, inherited from the very human attention economy it was trained on, will need conscious correction if AI is to fulfil its promise of limiting, rather than merely reshaping, the illegal wildlife trade.

Subject of Research: Assessing AI platform safeguards against the illegal wildlife trade

Article Title: AI platforms offer promise in limiting the illegal wildlife trade

Article References: AI platforms offer promise in limiting the illegal wildlife trade. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: artificial intelligence, illegal wildlife trade, conservation biology, large language models, Gemini, ChatGPT, DeepSeek, Grok, biodiversity, AI safeguards, University of Kent, wildlife crime

The research team, led by cardiologists at Bozhou People’s Hospital in Anhui Province, China, together with a colleague at the First Affiliated Hospital of the University of Science and Technology of China, evaluated ChatGPT Plus 5.5 thinking, Gemini 3.5 Flash, Microsoft Copilot smart, DeepSeek-V4-pro, Doubao, and Perplexity. Rather than inventing test questions in a conference room, the investigators built a standardized set of 56 sudden-cardiac-death-related public consultation questions from an unusually broad evidence base: Google Trends search data, online patient forums, public question-and-answer platforms, clinical guidelines, expert consensus statements, systematic reviews, and qualitative interviews with patients and family members. The topics spanned the full landscape of lay concern about sudden cardiac death, including symptom triage, emergency response, cardiopulmonary resuscitation and automated external defibrillator use, inherited cardiac risk, myocarditis-related worries, returning to exercise after illness, screening, and decision-making about implantable cardioverter-defibrillators.

Each of the 56 questions was submitted once to each of the six chatbots between May 10 and May 16, 2026, generating 336 individual model-response items for analysis. Five senior cardiology raters, blinded to which chatbot had produced which answer, then scored every response across a battery of validated instruments: a safety assessment, an accuracy assessment, an empathy scale, the DISCERN instrument for judging the quality of written health information, the EQIP instrument for ensuring the quality of patient education materials, the JAMA benchmark criteria, and the Global Quality Scale. Readability was calculated separately using six formula-based indices. Statistical comparisons between models used Cochran’s Q test and Friedman tests, with Kendall’s W reported as the effect size, and the agreement among raters was strikingly high, with a Fleiss’ kappa of 0.856 for the safety domain and intraclass correlation coefficients ranging from 0.801 to 0.887 for the other manually assessed metrics.

The headline finding on safety was genuinely reassuring at first glance. Safe responses predominated across all six models, suggesting that the current generation of publicly accessible chatbots has largely absorbed the basic guardrails of cardiac emergency communication. But the more consequential finding was that responses containing potential safety concerns appeared in every chatbot’s output during the study window. ChatGPT Plus 5.5 thinking and Perplexity showed the highest observed proportions of safe responses, while Doubao showed the lowest. Importantly, the qualitative severity analysis found that most flagged responses contained only a single potential safety concern, and the distribution of concerns was concentrated at the low and low-to-moderate severity levels, with higher-severity problems occurring less frequently. That pattern is consistent with a technology that has improved substantially but has not yet reached the reliability threshold that safety-critical health communication demands.

The qualitative analysis of flagged responses is arguably the most clinically valuable part of the study, because it names the specific failure patterns that clinicians and developers can act on. The reviewers repeatedly identified answers that delayed emergency activation for chest pain, syncope, or post-viral symptoms—the classic red-flag presentations of sudden cardiac death where immediate escalation to emergency medical services is the single most important instruction. They found instances of over-reassurance about prevention, implying that a person’s risk had been dismissed when individualized assessment was warranted. They documented fixed return-to-exercise timelines after infection or COVID-19, offered as blanket rules rather than guidance tailored to a patient’s cardiac evaluation. Some responses gave pulse-check instructions that could delay the start of CPR, a direct contradiction of resuscitation guidance that emphasizes rapid compressions. Others delivered oversimplified advice on screening, electrolyte disturbances, or implantable cardioverter-defibrillator decisions, flattening genuinely complex clinical judgments into deceptively simple answers.

Beyond safety, the study revealed sharp and statistically significant differences between models on nearly every quality dimension. Accuracy differed significantly across the six chatbots, with a Kendall’s W of 0.864 indicating strong rank concordance among raters; ChatGPT Plus 5.5 thinking, Gemini 3.5 Flash, Microsoft Copilot smart, DeepSeek-V4-pro, and Perplexity achieved similar median accuracy scores, while Doubao performed lower. Empathy also differed significantly, with a Kendall’s W of 0.531, and DeepSeek-V4-pro recorded the highest observed empathy score—a notable result, since empathetic framing is increasingly recognized as a determinant of whether anxious users actually follow health advice. Significant differences emerged for DISCERN, EQIP, JAMA benchmark, and Global Quality Scale scores as well, all at P < 0.001, confirming that the choice of chatbot materially changes the quality of the health information a member of the public receives.

DeepSeek-V4-pro emerged as the strongest overall performer on information quality within this dataset, achieving the highest DISCERN and EQIP scores and the most favorable readability profile, while Gemini 3.5 Flash showed comparable EQIP performance and Microsoft Copilot smart recorded the highest JAMA benchmark score. Yet two weaknesses were universal. Transparency remained limited across the board: chatbots rarely made clear where their information came from, how current it was, or how confident users should be in it, which matters enormously when a reader must decide whether to trust an answer about inherited cardiac risk. And every model produced responses with relatively high reading-grade requirements, meaning that the very populations most likely to need accessible cardiac health information—older adults, people with limited health literacy, non-native speakers—may struggle to extract the key safety messages from dense, jargon-laden text.

The authors are admirably explicit about the limits of their own findings. Each question was sampled only once during a single defined week in May 2026, and all prompts were submitted in English through public interfaces accessed from China. Generative models are stochastic and continuously updated, so the same question asked a month later might yield a different answer. The researchers therefore frame their results as a time-, language-, and access-specific snapshot rather than a stable ranking of the underlying models. This caveat is not a weakness of the study but a feature of honest evaluation methodology in a field where the tools change faster than the literature can assess them, and it underscores why one-off benchmark scores should never be treated as permanent verdicts on any AI system.

The practical implications are clear for three audiences at once. For members of the public, the message is that chatbots can be useful for general education about sudden cardiac death—understanding what an implantable defibrillator does, why myocarditis follows some viral infections, or what screening involves—but they must never substitute for calling emergency services or consulting a clinician when symptoms suggest a cardiac emergency. For clinicians and public health communicators, the study identifies the exact content areas where AI guidance most often goes astray: emergency escalation, CPR and AED instructions, return-to-exercise timing after infection, individualized risk interpretation, and screening decisions. For developers, the study’s recommendations read as a roadmap: prioritize evidence-linked content, build clear red-flag escalation into every relevant response, adopt plain-language communication, and submit systems to ongoing expert review rather than one-time validation.

What makes this study resonate beyond cardiology is its methodological template. By combining blinded expert scoring across validated instruments with a fine-grained qualitative analysis of safety failures, and by anchoring the question set in the genuine concerns of patients and families, the researchers have shown how AI health tools should be evaluated: not with a single accuracy number, but across the full multidimensional surface of safety, accuracy, empathy, quality, transparency, and readability. Sudden cardiac death is an unforgiving test case—outcomes are measured in minutes, and a subtly wrong answer about chest pain or pulse-checking can be fatal. On that test, the six chatbots earned a qualified pass: good enough to inform, not yet good enough to advise alone. As generative AI becomes a default first stop for health questions worldwide, studies like this one provide the evidence base that regulators, developers, and clinicians will need to decide where these tools belong in the chain of care—and where they must, without exception, hand the user straight to a human.

Subject of Research: Evaluation of generative AI chatbots for public health consultation on sudden cardiac death

Article Title: Multidimensional evaluation of generative AI chatbots for public consultation on sudden cardiac death: safety, accuracy, empathy, reliability, quality, and readability across six models

Article References: Chen, Y., Ma, H., Wang, H., Chen, D., Wang, H., & Jiang, R. (2026). Multidimensional evaluation of generative AI chatbots for public consultation on sudden cardiac death: safety, accuracy, empathy, reliability, quality, and readability across six models. BMC Public Health. https://doi.org/10.1186/s12889-026-29766-z

Image Credits: AI Generated

DOI: 10.1186/s12889-026-29766-z

Keywords: generative AI, chatbots, sudden cardiac death, public health, CPR, AED, patient safety, health communication, readability, empathy, large language models, BMC Public Health

The five systems examined represent the current cutting edge of consumer-facing artificial intelligence. The study tested ChatGPT in its GPT-5.2 configuration, xAI’s Grok 4.1, Google’s Gemini 3, Perplexity running Sonar and GPT-5-series backends, and Microsoft Copilot built on GPT-5. This lineup matters because these are not obscure research prototypes buried in a laboratory; they are the tools that medical students, residents and practicing physicians already consult daily for quick answers. Understanding how they perform on rigorous, specialty-specific questions is therefore not an academic curiosity but a matter of practical patient safety, since patients too are increasingly turning to these chatbots with their own dermatological concerns.

Board-style questions are a particularly demanding benchmark for any question-answering system. The American Board of Dermatology’s examinations are designed to distinguish competent specialists from merely knowledgeable ones, and the questions typically present nuanced clinical vignettes in which several answer choices are plausible. A candidate must weigh the patient’s age, lesion morphology, distribution, history of prior treatments, and sometimes subtle histopathological findings before selecting the single best response. This format tests something closer to clinical reasoning than to factual recall, which is precisely why it has become a popular yardstick for evaluating medical artificial intelligence. A model can memorize that psoriasis affects roughly two to three percent of the population, but it must reason through a vignette to recognize a psoriasiform drug eruption masquerading as the idiopathic disease.

The technical architecture underlying these models helps explain both their promise and their pitfalls. Large language models are neural networks, generally built on the transformer architecture, trained on vast text corpora to predict the next token in a sequence. Through this deceptively simple objective, they absorb grammar, factual knowledge, and patterns of reasoning encoded in their training data. Medical knowledge enters the picture because textbooks, review articles, clinical guidelines and discussion forums form a substantial portion of the public internet. However, the models do not store facts as a database would; knowledge is distributed across billions of numerical parameters, which means retrieval is probabilistic rather than deterministic. This is why a model can answer an obscure question flawlessly one moment and hallucinate a nonexistent drug interaction the next, a behavior that examiners would immediately fail in a human candidate.

Dermatology poses distinctive challenges for text-based models, even in a question-answer format that strips away the visual element. The specialty sits at the intersection of internal medicine, immunology, infectious disease, oncology and pathology, and its vocabulary is notoriously dense. Distinguishing lichen planus from lichenoid drug eruption, or dermatitis herpetiformis from linear IgA bullous dermatosis, depends on pattern recognition honed through exposure to thousands of cases. Board-style questions compress that pattern recognition into prose, which arguably favors language models, yet the compression also removes the contextual cues a clinician would use at the bedside. The Rutgers team’s choice of this format follows a growing research tradition: earlier studies had already tested ChatGPT on board-style dermatology questions, including image-based items, and Liu and colleagues in 2025 assessed dermatological knowledge and image analysis using specialty certificate examinations as their benchmark.

What makes the new study notable is its breadth of comparison. Most prior evaluations examined a single model, usually a ChatGPT variant, against a question bank. By running the same board-style items through five different systems, each with distinct underlying architectures and training pipelines, the researchers could probe whether strong medical performance is a general property of frontier language models or something specific to particular products. The differences among these systems are not trivial. GPT-5.2, Grok 4.1, Gemini 3 and the GPT-5-based Copilot each reflect different training data mixtures, different reinforcement learning strategies, and different approaches to reasoning. Perplexity adds a further wrinkle: it couples a language model with live web search, meaning its answers may draw on retrieved documents rather than purely on parametric memory, a hybrid architecture that could behave very differently on questions with recently updated guidelines.

The distinction between parametric knowledge and retrieval-augmented answering is one of the most technically interesting aspects of this kind of evaluation. A pure language model answers from what is baked into its weights during training, frozen at a cutoff date. A retrieval-augmented system like Perplexity queries the live web, which brings freshness but also vulnerability: it can absorb misinformation, outdated guidelines or commercially biased content from the open internet. For dermatology, where treatment recommendations evolve and where the internet is saturated with anecdotal skin-care advice, that distinction could materially change performance. A benchmark like the one used in this study therefore measures not just a model’s medical knowledge but its entire answer-generation pipeline, including how it filters and prioritizes whatever sources it consults.

The authors of the research letter are appropriately measured about what such testing can and cannot establish. Answering multiple-choice questions correctly is not the same as managing a patient. A board examination candidate who selects the right option for a bullous disorder vignette may still struggle to perform a safe skin biopsy or to counsel a frightened patient about a new diagnosis of melanoma. Conversely, models can exploit statistical cues in question wording, patterns that human test-takers learn to recognize as well, without genuinely understanding the underlying medicine. There is also the ever-present risk of data contamination: if board-style questions or close paraphrases circulate online, a model trained on the internet may have effectively seen the answer key. Rigorous benchmarking in medical artificial intelligence must therefore always ask not only how well a model scored, but whether the score reflects reasoning or recall.

Nevertheless, the trajectory across successive studies is striking. When ChatGPT first burst into public awareness in late 2022, early medical evaluations produced mixed results, with the model often falling short of passing thresholds on professional examinations. Each subsequent generation has closed the gap, and by the mid-2020s frontier models were routinely performing at or near the level of human test-takers across a range of specialties. The Rutgers study extends that line of evidence into dermatology with the newest available models, and its very framing, published as a research letter rather than a lengthy original article, reflects how quickly the field is moving: the findings needed to reach the literature before the models they evaluated were superseded. The authors report that all data were generated using the five named systems, with the dataset available from the corresponding author on reasonable request, and the work received no external funding while the authors declared no conflicts of interest.

For clinicians and patients alike, the practical message is one of cautious engagement rather than either alarm or celebration. These systems are already embedded in clinical workflows, from drafting patient messages to summarizing literature, and their demonstrated competence on board-style questions suggests they can serve as powerful adjuncts for education and decision support. A resident reviewing for boards might legitimately use a chatbot to generate practice vignettes or explain the immunopathology of pemphigus vulgaris. But the same probabilistic architecture that enables fluent expertise also produces confident errors, and no benchmark, however demanding, substitutes for clinical supervision. The Rutgers team’s comparison of five frontier models gives the field a clearer map of where artificial intelligence stands in dermatology today, and a reminder that the map will need redrawing with every new model release.

Subject of Research: Evaluation of large language models on American Board of Dermatology-style examination questions

Article Title: Performance of five large language models on board-style dermatology questions from the American Board of Dermatology

Article References: Rasty, R., Mackenzie, E., Mehdikhani, S., & Rao, B. (2026). Performance of five large language models on board-style dermatology questions from the American Board of Dermatology. Archives of Dermatological Research, 318(1), Article 462. https://doi.org/10.1007/s00403-026-04894-z

Image Credits: AI Generated

DOI: 10.1007/s00403-026-04894-z

Keywords: large language models, dermatology, artificial intelligence, board examination, ChatGPT, Gemini, Grok, Microsoft Copilot, Perplexity, medical education, clinical reasoning, benchmarking

The motivation behind the work is straightforward. Large language models have become remarkably fluent generators of scientific prose, capable of producing abstracts, introductions, and literature reviews that read convincingly like human scholarship. That fluency brings real benefits for drafting and editing, but it also opens the door to plagiarism, misinformation, and academic dishonesty at unprecedented scale. Journals, universities, and funding agencies now face a flood of submissions in which machine-generated passages may be indistinguishable to human readers. Previous studies have shown that even trained reviewers can be fooled: research comparing ChatGPT-generated abstracts with real ones found that blinded human reviewers often struggled to tell them apart, and commercial detection tools have been documented mislabeling genuine human manuscripts as AI creations. A reliable, scientifically focused detector has therefore become a matter of protecting the integrity and credibility of the academic record itself.

The technical heart of AISciDetector is a two-stage pipeline that marries two complementary ways of understanding text. In the first stage, the system uses BigBird, a transformer architecture designed to handle long documents efficiently. Standard transformer models such as BERT process text with a mechanism called self-attention, in which every word can, in principle, attend to every other word, but this becomes computationally expensive as documents grow longer. BigBird sidesteps that bottleneck with a sparse attention scheme, allowing the model to capture long-range dependencies, the distant relationships between words and concepts that are especially important in scientific writing, where a conclusion at the end of a passage may hinge on a definition or result stated paragraphs earlier. These BigBird embeddings serve as rich numerical representations of each text, encoding both its vocabulary and its broader structural patterns.

In the second stage, those embeddings are fed into a hybrid CNN-BiLSTM framework. The convolutional neural network, or CNN, acts as a local feature extractor: it slides filters across the embedding sequences to pick up short, characteristic patterns, much as convolutional networks detect edges and textures in images. Immediately after, a bidirectional long short-term memory network, or BiLSTM, models the sequential structure of the text in both directions. LSTMs are a form of recurrent neural network engineered to remember information over long sequences while avoiding the vanishing-gradient problems that plague simpler recurrent models. By running two LSTMs, one reading the sequence forward and one reading it backward, the BiLSTM layer builds a representation that reflects both what precedes and what follows each point in the text. The combined features then drive a binary classifier that outputs a single decision: human-written or LLM-generated.

Training and evaluating such a system requires data, and here the researchers made a second significant contribution. Alongside the detector, they introduce LLMSciTxt, a new dataset containing scientific texts written by human authors alongside outputs generated by three of the most widely used large language models: ChatGPT, Gemini, and LLaMA-3. This multi-model design matters because detectors trained on the output of a single generator often fail when confronted with text from a different one. By including outputs from three distinct model families, each with its own training data, architecture, and stylistic quirks, the dataset encourages the detector to learn general signatures of machine generation rather than the fingerprint of one particular system. The authors state that the datasets generated and analyzed during the study are available from the corresponding author on reasonable request.

The experimental results place AISciDetector ahead of the field it was tested against. The 91.5 percent accuracy and 91.49 percent F1-score, a metric that balances precision and recall and is particularly informative when the two error types carry different costs, exceeded the performance of classical machine learning approaches as well as transformer-based models evaluated under the same conditions. The authors attribute the gains to the hybrid design: BigBird supplies the long-context understanding that shallower models lack, while the CNN-BiLSTM stack adds a layer of local and sequential pattern recognition on top of those embeddings. In effect, the system looks at machine-generated text through three lenses at once, global context, local phrasing, and temporal flow, and the combination appears to capture subtle statistical traces that machines leave behind even when their prose reads naturally.

The study also situates itself within a rapidly growing research landscape. Earlier work by the same group explored deep learning detection methods for LLM-generated scientific content and a transformer-based approach presented at a computing sciences conference, and the new paper builds directly on that lineage. Around the world, other teams have pursued parallel strategies: some detectors analyze token probability sequences without any training, others frame human text as an out-of-distribution anomaly, and still others apply adversarial learning or statistical guarantees to make detection more robust. Benchmarks such as large-scale datasets of ChatGPT-written abstracts and real-world detection challenges have emerged to standardize evaluation. Yet the field remains contested, with independent evaluations of commercial tools such as GPTZero, ZeroGPT, Copyleaks, and Scribbr’s detector finding inconsistent reliability, and documented cases of human-written scientific manuscripts being wrongly flagged as machine-generated, a false positive that can carry serious consequences for authors.

That tension between sensitivity and fairness is precisely why domain-specific detectors like AISciDetector may prove valuable. Scientific text has distinctive properties, dense terminology, formulaic section structures, and heavy citation conventions, that generic detectors trained on news or student essays may handle poorly. By training and testing specifically on scientific content from multiple generators, the Jordanian team’s system targets the exact domain where the stakes are highest: peer-reviewed publishing. The authors argue that effective identification of LLM-generated content supports the integrity and credibility of academic publications, a goal shared by editors confronting a rising volume of submissions in which the boundary between human and machine authorship has blurred. The work also connects to broader efforts in computational linguistics, literature mining, and the emerging science of explainability for large language models, all of which seek to make machine-generated language more transparent and accountable.

Challenges remain before such tools can be deployed as gatekeepers. Language models evolve quickly, and each new generation of generators may produce text whose statistical fingerprints differ from those in today’s training data, meaning detectors must be continually retrained and revalidated. Cross-domain performance, robustness to paraphrasing or deliberate evasion, and the ethical weight of false accusations all demand careful handling, and the authors themselves note that detection and anti-detection exist in a continuing arms race. Still, the study offers a concrete demonstration that combining long-range transformer embeddings with bidirectional sequence modeling can push detection accuracy above 91 percent on scientific text spanning three major model families. As universities and journals grapple with how to police AI-assisted writing without punishing honest authors, tools like AISciDetector represent a step toward detection methods that are both technically sophisticated and tailored to the specific texture of scientific prose, helping ensure that the literature of science remains a record of genuine human inquiry, however much machines may assist along the way.

Subject of Research: Detection of large language model-generated scientific text using a hybrid BigBird and CNN-BiLSTM deep learning model

Article Title: LLM-generated scientific content detection method

Article References: Bader, A., & Alhijawi, B. (2026). LLM-generated scientific content detection method. Neural Computing and Applications, 38(17), Article 728. https://doi.org/10.1007/s00521-026-12465-6

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12465-6

Keywords: large language models, AI text detection, scientific integrity, BigBird, CNN-BiLSTM, ChatGPT, Gemini, LLaMA-3, academic plagiarism, deep learning, natural language processing, transformers

Alexandre S. Pires and colleagues set out to map what might be called the social norms of large language models. The research team examined 21 different LLMs, probing each system with a deceptively simple task: judge fictional people as good or bad after learning how those people behaved in a series of interpersonal dilemmas. The scenarios covered familiar moral terrain, such as whether a person chose to spend time and energy helping someone with a problem, or whether they gave food or money to another person in need. By varying the reputations of the people involved, the researchers could determine whether a model rewarded generosity universally or conditioned its approval on the character of the recipient.

The results revealed a broad consensus on one point and deep disagreement on another. Across the board, the models approved of helping and sharing with people who had already demonstrated themselves to be good. Cooperation with well-reputed individuals was, in the eyes of every system tested, virtuous conduct. The trouble began when the fictional actors interacted with individuals of ill repute. Here the models split into distinct camps, each implicitly endorsing a different philosophy about how decent people should treat those who have behaved badly.

Most, but not all, of the LLMs tested assigned a good reputation to people who cooperated with bad recipients. In this view, extending help even to a known defector is a laudable act, perhaps reflecting an ethic of unconditional kindness or a hope that generosity might reform the wrongdoer. But two models stood apart. Gemma 2 27B IT and Llama 3.1 8B tended to endorse shunning bad individuals, judging that cooperating with them was itself bad behavior. For these systems, morality is not just about what you give but about who deserves to receive it, and associating with the wrong people can stain your own standing.

The disagreements grew sharper when the researchers examined the opposite case: refusing to help or share with bad people. GPT-4o tended to penalize people for not cooperating with bad individuals, effectively treating withholding aid as a moral failing regardless of the recipient’s history. Llama 3.3 70B took the opposite stance, typically arguing that bad people should be punished for their prior lack of cooperation, so refusing to help them is entirely fine. Meanwhile, Gemini 1.5-Pro and Grok 2 proved inconsistent in their judgements, sometimes approving and sometimes condemning the same pattern of conduct depending on details of the scenario. For users seeking moral guidance, this means that identical questions posed to different popular chatbots can yield substantively different ethical verdicts.

Despite the noise, the researchers identified a discernible trend. Most LLM model families appear to be evolving toward a norm that social scientists call Simple Standing. Under this norm, cooperating is always good, and defecting against bad individuals is also deemed good. Simple Standing is permissive: it never requires anyone to punish or exclude wrongdoers, but it gives everyone license to walk away from them without guilt. It stands in contrast to stricter reputational codes found in human societies, such as norms under which failing to sanction a defector makes you complicit in their defection.

Why does this taxonomy matter beyond the philosophy seminar? Because norms are not merely descriptions of behavior; they are engines of it. The authors modeled what would happen if the various philosophies underpinning the LLMs’ judgements were universalized across an entire population. If everyone operated under the rules of Simple Standing, cooperation would indeed ensue, but it would not reach the high levels achieved under a sterner framework, one in which you should always fail to help bad people or be judged as bad yourself. That harsher regime, sometimes associated with norms of punitive standing, sustains near-universal cooperation by making association with defectors costly. In other words, the gentle, forgiving norm that most AI systems seem to be converging on produces a decent but less cooperative world than the unforgiving alternative.

The stakes rise when one considers how people actually use these systems. As more and more individuals turn to LLMs for advice on interpersonal matters, the ethical foundations behind the tools’ replies increasingly have the power to shape individual lives and, at scale, the social fabric itself. A person deciding whether to cut off a difficult relative, whether to give a second chance to an unreliable business partner, or whether to help a stranger with a bad reputation may be quietly absorbing the normative leanings of whichever model they consult. If a generation of users internalizes Simple Standing because their assistants endorse it, patterns of human cooperation, punishment, and reconciliation could shift in ways that no democratic process ever debated.

The study also uncovered subtler biases in how models render their judgements. LLM-based assessments depended on the gender and perceived cultural background of the recipients as well as the overall context of the fictional situation. The same act of helping or withholding could receive different moral evaluations depending on who was being helped, a finding that raises questions about consistency and fairness in AI moral reasoning. If a model’s verdict on your behavior shifts with the demographics of the people involved, the advice it gives may encode stereotypes that users never consented to and may not notice.

Can the problem be fixed with a simple instruction? Apparently not easily. The authors report that prompting interventions, such as instructing a model to adopt norms that promote overall cooperation, have a limited and inconsistent effect across LLMs. A nudge that steers one system toward a cooperative norm may leave another untouched or produce erratic behavior. This fragility suggests that the moral dispositions of these models are baked deeply into their training and alignment, and that shaping them deliberately will require more than clever phrasing in a user prompt. As AI systems become ubiquitous advisers on human relationships, the study implies, society will need to pay attention not only to what these machines know but to what, implicitly, they believe is right.

Subject of Research: Social norms and moral judgements of large language models regarding human cooperation

Article Title: AI’s social norms and their implications for society

Article References: AI’s social norms and their implications for society. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: large language models, social norms, human cooperation, AI ethics, PNAS Nexus, reputation, Simple Standing, moral judgement, GPT-4o, Llama, Gemini, prompting interventions

The study, published in Discover Artificial Intelligence, was designed to mirror how ordinary people actually use these tools. On a single day, 30 June 2025, the team submitted all fifteen questions to each model through standard consumer web interfaces, exactly as a patient would, with no API access, no custom settings, and no prompt engineering. The questions themselves were drawn exhaustively from the official NHS page of common questions about pregabalin, a resource compiled and validated by clinical pharmacists for the general public. Each query came wrapped in a standardized instruction asking for a clear, factual, patient-appropriate answer that must include a warning if any potential danger existed. That explicit safety instruction is what makes the study’s central finding so striking.

Pregabalin was a deliberate and timely choice. The drug is a structural analogue of the neurotransmitter GABA, but it does not act on GABA receptors directly; instead it binds selectively to the alpha-2-delta-1 subunit of voltage-gated calcium channels, dampening the release of excitatory neurotransmitters. Its clinical footprint has expanded dramatically over the past decade. Global annual sales of gabapentinoids, the drug class that includes pregabalin, rose by 114.5 percent between 2012 and 2022, and by a staggering 180.9 percent in low- and middle-income countries, where specialist care is scarce and patients are especially likely to seek answers from digital platforms. Prescribing rates more than doubled over ten years, reaching 7.2 prescriptions per 100 patients in 2019, with the highest volumes among elderly patients juggling multiple medications. Off-label and recreational use have broadened the population asking questions even further.

To judge the answers, the researchers assembled a multidisciplinary panel of a rheumatologist, a neurologist, and a pharmacologist, each with at least five years of clinical experience, reflecting the three main specialties that prescribe and monitor pregabalin. The evaluators worked independently and blind to which model had produced each response, after a calibration session in which they jointly scored three practice answers to align their criteria. Six dimensions were assessed: warning compliance, accuracy, completeness, content safety, readability, and clinical relevance. The study followed the TRIPOD-LLM reporting guideline, an emerging standard for research involving large language models, and the team calculated inter-rater reliability using Fleiss’ kappa and intraclass correlation coefficients.

The headline numbers on accuracy look encouraging. ChatGPT-4o, Gemini 2.0 Flash, and Copilot each scored 0.933 for factual correctness, while Claude 3.5 Sonnet scored 0.867. Clinical relevance ranged from 0.867 to 1.000 across the four models. But the confidence intervals overlapped so extensively that no model could claim reliable superiority over any other, a point the authors emphasize repeatedly. With only fifteen questions and a single query per model, the differences fall well within the statistical noise. The researchers are candid about this: the point estimates are best read as an illustration of the evaluation framework rather than stable performance benchmarks, and the single-query design means the stochastic variability of these models, which can produce substantively different answers to identical prompts, was never characterized.

The safety findings are where the study bites. Every one of the fifteen questions was judged by panel consensus to carry a clinically relevant safety concern warranting a warning, and the prompt explicitly demanded one. Yet warning compliance scores ranged only from 0.467 for ChatGPT-4o and Gemini 2.0 Flash to 0.733 for Claude 3.5 Sonnet. In plain terms, the models ignored a direct safety instruction in roughly 27 to 53 percent of cases. The authors argue this is arguably more alarming than a simple failure to volunteer warnings, because it demonstrates that even explicit prompting cannot guarantee reliable safety communication. For a drug with dependence potential, meaningful interaction risks, and a heavy presence among polypharmacy patients, that gap between instruction and behavior is the study’s most clinically consequential result.

When the researchers dissected the errors, a clear pattern emerged. Oversimplification dominated, accounting for 56.3 percent of all recorded errors, followed by omission of essential information at 34.8 percent, with outright hallucinations making up 8.9 percent. Gemini 2.0 Flash logged the most errors at 32, while Claude 3.5 Sonnet logged the fewest at 25. The authors connect these patterns to plausible, though unproven, architectural explanations: token probability maximization may favor statistically common, simplified phrasing over clinically nuanced detail; attention-based architectures may underweight rare but critical content; and reinforcement learning from human feedback may optimize for fluency and consensus at the cost of completeness. Hallucinations, though rarest, remain the most dangerous failure mode, since a fabricated dosing recommendation or a fictitious drug interaction could translate directly into harm.

The study is notable as much for its transparency as for its findings. In an unusual move, the authors disclose that the rater-by-item scoring matrix and the original response corpus were not archived after analysis, that the rule used to collapse three raters’ scores into one per-question value was never recorded, and that formal paired statistical tests such as McNemar’s test therefore became impossible. Copilot’s characteristic inline citations may also have compromised the blinding protocol. Rather than presenting plausible reconstructions as fact, the team reports these gaps openly, framing the work as an exploratory evaluation whose primary contribution is the multidimensional framework itself and a model for honest documentation of methodological limitations in early-stage LLM research.

What should be done with these results? The authors recommend hybrid architectures that ground chatbot outputs in structured, validated pharmaceutical databases rather than relying on parametric knowledge alone, along with dedicated fine-tuning on pharmacovigilance corpora so that safety alerts are generated reliably rather than left to general instruction-following. They also propose uncertainty detection mechanisms that would signal a model’s limits and steer users toward professional consultation, and even a continuous error-monitoring system modeled on existing pharmacovigilance infrastructure. The limitations are real: one medication, fifteen questions, one query per model, English only, and no testing with actual patients. The team, based in Francophone West and Central Africa, already plans a follow-up in French, where the stakes may be even higher. For now, the message to patients is clear: chatbots can be a reasonable starting point for basic medication facts, but when it comes to the warnings that keep you safe, they still cannot be trusted to speak up.

Subject of Research: Evaluation of large language models answering patient questions about the medication pregabalin

Article Title: A structured exploratory multidisciplinary evaluation of four large language models responding to frequently asked patient questions about pregabalin

Article References: A structured exploratory multidisciplinary evaluation of four large language models responding to frequently asked patient questions about pregabalin. (n.d.). https://doi.org/10.1007/s44163-026-02394-7

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02394-7

Keywords: large language models, ChatGPT-4o, Copilot, Gemini, Claude, pregabalin, patient safety, medication information, warning compliance, hallucination, pharmacovigilance, TRIPOD-LLM

The study was led by Mattea Sim, an assistant research professor at the Massive Data Institute in Georgetown’s McCourt School of Public Policy. Her co-authors were Yoshi Kohno, the McDevitt Chair in Computer Science, Ethics, and Society and professor at Georgetown’s Department of Computer Science and Center for Digital Ethics, and Yael Eiger, a Ph.D. candidate at the University of Washington whose research focuses on technology in the carceral system. Kohno framed the work as part of a broader institutional effort, saying the study is part of a growing effort at Georgetown focused on exploring the impact and relationship between AIs and humans, grounded in both psychology and computer science.

The researchers adopted two distinct approaches to understand how AI summaries interact with human memory. In the first phase, they analyzed summary outputs produced by OpenAI’s ChatGPT and Google’s Gemini to quantify how frequently errors appeared and what categories those errors fell into. The results were sobering. The AI summaries frequently omitted critical details, misconstrued information and hallucinated inaccurate details that were not present in the underlying material at all. The source material was an animated video depicting a car and pedestrian accident, and the summaries were measured against the actual contents of that video to determine how faithfully the models reproduced what happened.

The most common category of error was omission, meaning the failure to mention the primary events of the animated video. Across all prompts and models, the summaries omitted 51.6 percent of central details on average. Even more striking, all but one of the summaries — 95 percent of them — omitted the single most central detail of the event: the fact that a car collided with a pedestrian. Eiger said she was struck by how bad the summaries were, even at this stage in AI development, and noted that it worries her that police departments may be using video summarization technologies without rigorous testing and without an awareness of how incorrect AI-generated summaries could be.

The second phase of the study examined what those errors do to the people who read them. Participants were first asked to watch one of two animated videos, which were created in a recent replication study that modernized materials from a classic human memory experiment. The researchers were careful about the nature of the footage. Kohno emphasized that they did not use actual police body camera footage. Instead, the videos depicted a red car approaching an intersection with either a stop or a yield sign, with participants randomly assigned to see one of the two traffic signs. In the videos, the car turns right at the intersection and collides with a pedestrian walking into the road. After being knocked down, the pedestrian stands up, the driver exits the car and the two meet.

Between 24 and 48 hours later, participants returned for the second portion of the experiment. Each was randomly assigned to read either an AI-generated summary that was accurate or one that contained an inaccurate detail changed by the researchers. Before reading, participants were also told that the summary had been written by either an AI system or a human transcriber. In reality, all of the summary text was generated by ChatGPT, and in some cases separate parts of the model’s answers were stitched together to ensure that the experimental conditions remained uniform across participants. In total, 331 people completed both parts of the experiment, providing a substantial sample for measuring how the summaries shaped recall.

The memory effects were dramatic. When participants were asked whether the car in the video had approached a stop sign or a yield sign, 83.6 percent of those who read a summary with accurate information answered correctly. Among those who read a misleading summary, only 44.8 percent were correct — a gap of nearly 39 percentage points. Participants who read a misleading AI summary were significantly less likely to accurately recall the original event compared to people who read an accurate summary. Crucially, the manipulation worked even when people knew the summary came from a machine. The impact of misleading AI summaries on memory did not depend on participants’ trust in AI or their familiarity with it, suggesting that susceptibility to this form of misinformation is broad and not easily guarded against by skepticism about the technology.

Sim described the phenomenon in stark terms, saying that AI is a new method of delivering misinformation and that it has the potential to create false memories for people who are reading that information. She argued that society should be thinking deeply and critically about whether and how AI should be used to summarize information, especially in high-stakes settings. The authors made a similar point in the study itself, writing that the findings have implications for how AI should be used in critical settings. They noted that although humans-in-the-loop are often expected to correct for AI’s mistakes, their work suggests human memory can instead be distorted by those mistakes, and that AI has the potential to generate misinformation even absent any adversarial intent, which can meaningfully impact human memory.

The implications extend well beyond the laboratory. As individuals and institutions grow more reliant on AI, Kohno hopes the findings encourage people to recognize that AI outputs can have a direct impact on how humans perceive the world. That concern carries particular weight in domains such as policing and medicine, where a distorted summary of an event could influence investigations, legal proceedings or clinical decisions. Eiger, whose research examines technologies increasingly implemented by police for surveillance or departmental efficiency, said these tools are often advertised and framed as beneficial, but the trade-offs remain poorly understood. She would like to see more audits of this type of technology and pointed out that, from a public transparency perspective, there is little insight into how the deployed systems work or which exact models are being used in the race to bring AI into every sector.

The authors also acknowledge the limits of the current work and point toward what comes next. The study used the most up-to-date consumer-facing versions of large language models, but the researchers did not have access to enterprise versions of AI designed for organizational use, which may behave differently. For future studies, the team wants to move beyond animated videos and examine real-world examples. Sim asked what happens with AI summaries of police body camera footage and what police reports look like when they are generated by AI, and whether such summaries manipulate people’s memories of police-civilian interactions and consequently their perceptions of what happened in an incident. Those questions, she suggested, would be very interesting to answer — and the answers may determine how safely AI summarization can be entrusted with society’s most consequential records.

Subject of Research: The effect of misleading AI-generated summaries on human memory and perception of truth

Article Title: New research finds that misleading AI-generated summaries can distort human memory

Article References: New research finds that misleading AI-generated summaries can distort human memory. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: artificial intelligence, AI summaries, human memory, misinformation, ChatGPT, Gemini, large language models, memory distortion, police body camera footage, AI ethics, Georgetown University, University of Washington