The core argument of the editorial is that evaluation must satisfy two distinct but complementary demands. Robustness means examining whether a system’s performance and behavior remain consistent under variable real-world environments, where data is ambiguous, distributions shift over time and users behave in ways that laboratory datasets rarely capture. Responsibility means ensuring that the objectives and indicators chosen for evaluation actually reflect the users and environments the system affects, including the interests of companies, governments, organizations and society as a whole. Traditional offline testing on curated datasets, the authors contend, cannot deliver either guarantee on its own. A model that scores well on a static benchmark may degrade badly when browsing patterns change, or it may optimize a metric that silently sacrifices fairness for engagement. Evaluation, they argue, should therefore not be viewed as the final stage of comparison between competing proposals, but rather as a key component of the entire system design process that reveals which behaviors and trade-offs should be prioritized.

The first contribution in the collection, from Baysan and colleagues, tackles one of the most practical bottlenecks in large-scale search: evaluating structured query parsing before live A/B testing is possible. The team proposes an LLM-as-a-Judge framework built on different versions of Gemini, in which a large language model scores the outputs of query parsers against defined criteria. The framework supports three evaluation methodologies: pointwise scoring, which assigns a numerical score based on evaluation criteria; pairwise comparison, used to rank two parsed outputs against each other; and pass/fail binary classification. The reported results show high agreement with human judgments both on a small cleaned dataset and on a large real-world dataset, suggesting that automated judges can scale evaluation to a volume no human team could sustain. Yet the authors are candid about failure modes: their error analysis documents difficulties with hallucinated outputs related to categories, and they stress that LLM evaluators themselves must be validated against human evaluation to ensure reliability. The lesson is a layered one: machine judges can extend human oversight, but they cannot yet replace it without risking silent, systematic errors.

The second study shifts the lens from infrastructure to values. Müllner and colleagues present a survey asking what algorithmic fairness in recommender systems can learn from tourism management, a field that has grappled for decades with questions of who benefits and who loses from visitor flows. After an initial screening, the researchers analyzed 44 articles, 24 from tourism management and 20 from computer science. Their comparison is revealing: tourism research tends to embrace contextual, qualitative and participatory approaches, whereas computer science typically relies on more restrictive formal criteria and mathematical definitions of fairness. Rather than treating these as incompatible, the authors argue that closer interdisciplinary collaboration is critical to developing responsible recommender systems. They propose using insights from tourism research to expand the definitions of algorithmic fairness, to identify the different stakeholders involved, to surface the trade-offs that occur in real deployments, and to include participatory processes in system design. Most provocatively, they show how qualitative objectives such as reducing overtourism can be transformed into measurable indicators aligned with broader sustainability and equity goals, bridging the gap between soft values and hard metrics.

The third contribution, by Senger and colleagues, addresses career path prediction, a task with obvious stakes for individuals navigating labor markets. The team systematically compared four families of approaches: linear transformations, multilayer perceptrons, long short-term memory networks, and large language models. The experiments were run on two datasets, DECORTE and KARRIEREWEGE+, using different input representations of career histories. The technical findings are instructive for anyone building sequence models on textual data. When job titles are combined with full descriptions, results are generally better than when using title-only inputs, indicating that the semantic richness of a description carries signal a bare title cannot. In most cases the authors’ multilayer perceptron achieved the best performance, setting new state-of-the-art scores, a humbling result for the era of foundation models. Long short-term memory networks obtained good results when experiments involved longer career histories, exploiting their capacity to model sequential dependencies. Meanwhile, the performance of large language models proved highly sensitive to fine-tuning, grounding and the alignment of the dataset, a reminder that pretraining scale does not automatically translate to domain competence.

Senger and colleagues also examined a technique many practitioners reach for by default: synthetic data augmentation. Their analysis showed that while adding synthetic examples may increase the diversity of training data, it may also introduce semantic distortions. In other words, generated career paths can look superficially varied while quietly shifting the meaning of the underlying data, corrupting the very signal a model is meant to learn. For evaluation, this finding carries a broader implication: data augmentation strategies must themselves be evaluated for the distortions they introduce, not merely credited for the volume they add. It is exactly the kind of hidden trade-off that the editorial argues responsible evaluation should be designed to expose.

The fourth study, by Waterschoot and van den Bosch, moves into the volatile environment of online news. Selecting which reader comments to highlight on a news platform requires significant resources and attention from human moderators, so the authors propose a group recommender system built from an ensemble of classifiers that integrates both contextual relevance and user history to rank candidate comments. Crucially, they tested their approach on unobserved discussions from 2020 and 2023, a temporal evaluation strategy that directly probes robustness across shifting topics and discourse styles. The system showed robustness against a variety of topic factors, and the results of the best model were validated by experts. Just as importantly, the authors are explicit about the system’s intended role: it is designed to support moderators, not to replace them, so the final say always rests with the human moderator. This design philosophy, automation as augmentation rather than substitution, exemplifies the responsible dimension of evaluation the editorial champions.

From the four studies, three cross-cutting concepts emerge. First, evaluation settings should closely resemble real-world environments. In practice this means temporal testing on data from different time periods, using multiple datasets and input representations, employing query patterns that reflect genuine browsing behavior, and explicitly analyzing the needs of different stakeholders. A single static split of a benchmark dataset tells us almost nothing about how a system will behave next month or next year. Second, human expertise in the domain remains critical. Moderators’ evaluations, professional taxonomies, expert annotations and stakeholder participation all help determine what counts as suitable, accurate or fair in each specific domain, and no automated pipeline can infer those judgments from data alone. Third, and perhaps most decisively, there is no single aggregate metric that can capture all the relevant behavior a system should be analyzed for. Classification and ranking metrics must be complemented with experiments across different groups of users, reliable statistical computation, qualitative error analysis and stakeholder-oriented testing that reveals complementary perspectives on real-world performance.

The editorial’s closing argument elevates evaluation from a technical checkpoint to a continuous social and technical process anchored in the stakeholders involved across a system’s life-cycle. It is critical, the authors write, to define which results and interests are important according to the different affected stakeholders, and evaluation protocols should include realistic scenarios that examine potential failures occurring under human supervision. Assumptions and limitations of studies must be clearly stated so that they can serve as the basis for responsible decision-making, showing the ambiguity and disagreements that lie hidden behind overall scores. In an era when a single number on a leaderboard can determine which algorithm ships to millions of users, that reframing is a call to action for the entire field: the numbers we choose to trust ultimately decide what our systems optimize, and therefore what our digital environments become.

Subject of Research: Responsible and robust evaluation of real-world recommendation and search systems

Article Title: Editorial: Responsible and robust evaluation for real-world recommendation and search systems

Article References: Bellogín, A., Sánchez, P., & Sebastiá, L. (2026). Editorial: Responsible and robust evaluation for real-world recommendation and search systems. Frontiers in Big Data, 9, Article 1995917. https://doi.org/10.3389/fdata.2026.1995917

Image Credits: AI Generated

DOI: 10.3389/fdata.2026.1995917

Keywords: recommender systems, search systems, evaluation, large language models, algorithmic fairness, LLM-as-a-Judge, career path prediction, synthetic data, comment moderation, stakeholders, temporal testing, Frontiers in Big Data

The problem ARTEMIS addresses is familiar to anyone who has watched a large language model answer a specialized scientific question with confident fluency and quietly wrong facts. When the researchers asked a base model to produce analytical reports on monoclonal antibodies, the 3Rs framework of Replacement, Reduction and Refinement, and animal-use reduction, the output was polished but generic, untraceable to source evidence, and occasionally incorrect. In one striking example, the model asserted that the FDA had only 12 reviewers with specific alternative-methods expertise across all centers, a precise quantitative claim that was simply false. Static training data and the risk of hallucination make raw model knowledge unreliable in specialized domains, and even retrieval-augmented approaches that fetch raw passages often fall short when a question requires integrating findings scattered across dozens of studies.

ARTEMIS takes a different route: rather than handing the model raw documents, it first reorganizes the entire corpus into a hierarchical, evidence-grounded memory. The workflow operates at two linked levels. At the granular level, each source article is segmented into overlapping chunks of 2,000 characters, and each chunk is distilled by a language model into a concise observation capturing its most salient evidence. At the higher level, the system generates reflections by synthesizing observations from each chronological batch of five articles together with the immediately preceding reflection, so that new findings are always interpreted in light of previously accumulated knowledge. This recursive process builds an evolving memory that preserves both local evidence and broad patterns, and because the corpus is processed in publication order, the memory retains a timeline that supports historical analysis.

The scale of the compression is considerable. Processing 3,096 full-text articles produced 31,757 raw chunks, which were distilled into a memory of simple observations and 620 reflections. The median word counts tell the story: observations averaged just 24 words, reflections 599, and raw chunks 805. Crucially, the compression did not flatten the literature into a few oversimplified themes. Topic modeling with BERTopic found a similar number of topics in the reflections as in equally sized random samples of raw chunks, suggesting broad thematic coverage was preserved. But the reflections were far more conceptually convergent: their average inter-topic distance was roughly 0.2, compared with about 0.5 for raw-chunk samples, and they occupied a markedly smaller semantic volume in embedding space, approximately 0.02 versus 0.045 to 0.055. In other words, ARTEMIS does not merely shorten the literature; it reorganizes distributed evidence into a tighter, more coherent analytical representation.

To test whether this synthesized memory actually helps, the team assembled 30 high-level questions drawn from real analytical problems in the monoclonal antibody literature, including questions about reducing animal use and the evidence base for emerging alternative methods. Because these questions demand synthesis across studies, they are poorly served by retrieving isolated passages. For each question, answers were generated with the same Llama 3.3-70B model under different context conditions, and quality was scored on a 1-to-10 rubric by three independent LLM judges: GPT-4o, Claude 4.0 and Claude 4.6, each operating in isolated sessions to avoid conversational carryover. The results were consistent. ARTEMIS-based contexts produced complete answers for all 30 questions, whereas raw-chunk retrieval explicitly declared insufficient information in 5 of 30 cases. Across judges, ARTEMIS configurations received higher or comparable scores than raw-chunk retrieval, achieving 100 percent answerability versus 83 percent for the baseline.

The efficiency gains were just as notable. A context of 10 observations and 10 reflections represented only about 77.4 percent of the word count of 10 raw chunks yet matched or beat their answer quality, and a leaner ARTEMIS-lite configuration of 10 observations and 5 reflections, just 40.2 percent of the raw word count, still scored similarly or higher in many cases. Simply throwing more text at the problem did not help: when the team expanded retrieval to 50 raw chunks on a subset of questions, performance generally failed to improve and sometimes degraded, and on one item both approaches returned not-enough-information responses. The lesson, the authors argue, is that context structure and information density matter more than volume.

A series of component comparisons clarified why the combined design works. Observation-only contexts were often too compressed, failing to answer 16 of 30 questions with a mean score of 3.47. Reflection-only contexts were better but sometimes too abstract, with 8 insufficient responses and a mean score of 5.23, slightly below raw-chunk retrieval at 5.53. A single-pass summarization baseline, in which the top 10 retrieved chunks were flattened into one summary, performed poorly as well, with 14 insufficient answers and a mean score of 3.73, showing that one-shot compression without recursive accumulation is not the secret. The full ARTEMIS configurations topped the table with mean scores of 6.05 for ARTEMIS-lite and 6.28 for full ARTEMIS, and were at least as good as raw-chunk retrieval on 25 and 26 of the 30 questions respectively. The two layers, granular evidence and higher-level synthesis, are complementary, and neither alone suffices.

Perhaps the most intriguing capability is steerability. Because the reflection prompt can be modified, the same corpus can be organized under different analytical lenses without building separate evidence bases. The researchers generated three distinct sets of reflections from the identical monoclonal antibody literature using three expert personas: a 3Rs Ethical Reviewer, a Clinical Development Risk Strategist, and a Novel Alternative Methods Technology Analyst. Each persona produced distinct analyses emphasizing different signals and concerns, much as different human experts would. When the team probed the memories with a NAMs-specific query about organ-on-a-chip validation challenges, the NAMs persona’s reflections showed clearly higher cosine similarity at the top of the relevance ranking, confirming that the steering process enriches the memory with signals pertinent to the target lens. The framework also supports timeline-aware historical synthesis: using publication-ordered reflections as context, ARTEMIS generated an evidence-grounded narrative of Aducanumab’s development in four chronological phases, tracing how animal model usage evolved through the principles of Replacement, Reduction and Refinement.

The authors are careful about scope and limitations. LLM-as-a-Judge evaluations, they note, should be read as supportive rather than definitive; the judges differed in scoring stringency, with GPT-4o generally assigning higher marks than the Claude models, though relative rankings were consistent and repeated Claude 4.0 scoring in isolated sessions yielded identical results. Synthesis is also inherently lossy, and observations and reflections may omit low-frequency technical details that matter for some questions, which is why ARTEMIS is positioned as a support framework for complex literature analysis rather than a replacement for consulting source documents when full detail is required. The study centers on a single monoclonal antibody use case, and the team suggests future work could explore alternative corpus-ordering strategies, such as thematic grouping or citation-based progression, that might shape the analytical perspective embedded in the accumulated memory. Funded by the FDA and released as open access, the work arrives at a moment when regulatory agencies, drug developers and researchers alike are wrestling with literature too large and fast-moving for any human team to synthesize alone. If the approach generalizes beyond this first domain, the idea of building reusable, steerable, evidence-grounded memories from scientific corpora could become a standard layer between the world’s papers and the models asked to make sense of them.

Subject of Research: A memory-based AI framework for generating reusable analytical contexts from scientific literature for regulatory science

Article Title: ARTEMIS enables reusable analytical context generation from scientific literature

Article References: Li, D., Gong, B., Tong, W., & Xu, J. (2026). ARTEMIS enables reusable analytical context generation from scientific literature. Discover Artificial Intelligence, 6(1), Article 1465. https://doi.org/10.1007/s44163-026-02485-5

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02485-5

Keywords: ARTEMIS, large language models, regulatory science, monoclonal antibodies, retrieval-augmented generation, literature mining, 3Rs, novel alternative methods, LLM-as-a-Judge, evidence synthesis, FDA, knowledge curation

The experiment was designed with unusual rigor. Rospocher assembled 80 recent English-language arXiv manuscripts, twenty each from computer science, mathematics, physics, and quantitative biology, deliberately chosen in short and long length buckets and posted after March 2025 to reduce the chance that the evaluator models had memorized them. Author-identifying information, acknowledgements, and repository links were stripped from the PDFs so that the text itself carried no identity clues. Each manuscript was then evaluated under a factorial design: a corresponding-author profile was appended varying three factors at two levels each, producing eight profile conditions plus a blind baseline with no author information at all. Because every manuscript was evaluated under every condition, each paper served as its own control, a repeated-measures structure that sharply isolates the effect of the metadata from the effect of the content.

The three manipulated factors were carefully chosen to represent different classes of real-world cues. The first was a name-coded identity cue, alternating between first names commonly perceived in the United States as Black female-coded, such as Lakisha and Tanisha, and White male-coded, such as Brad and Hunter, all paired with a fixed surname to limit confounding. The second was institutional prestige, drawn from the QS World University Rankings 2026, contrasting elite institutions like MIT, Oxford, and Stanford with lower-ranked ones like Cleveland State and Central Michigan. The third was a bibliometric profile: a high profile with an h-index of 32 and roughly 8,000 citations, versus a low profile with an h-index of 8 and 320 citations. These numbers matter because they are exactly the kind of externally retrievable signals that a retrieval-augmented AI assistant could surface through a routine author lookup, even when the manuscript itself is anonymous.

Six heterogeneous evaluator models scored every manuscript under every condition: GPT-4.1-mini accessed through an API, and five locally hosted open-weight models including Gemma 3, Llama 3.1, Olmo 3, Qwen3, and a review-specialized model called DeepReviewer. Each model produced integer ratings on a seven-point scale from minus three to plus three across fourteen review-style criteria, spanning writing clarity, novelty, methodological soundness, perceived importance, award-worthiness, funding potential, hiring-committee appeal, and an overall acceptance recommendation. Each manuscript-condition-model combination was run five times with near-deterministic decoding settings, and the results were averaged and pooled across models with equal weight. Uncertainty was quantified with 20,000 bootstrap resamples, sign-flip permutation tests, and false-discovery-rate correction, alongside an ordinal-direction diagnostic that used only the sign of each paired difference to avoid assuming equal intervals on the rating scale.

The headline finding is that evaluator outputs are not invariant to author metadata. The strongest and most consistent driver was the bibliometric profile: papers attributed to a high-citation author scored systematically higher on perceived prestige, impact, and acceptance recommendation than the very same papers attributed to a low-citation author. Institution tier produced a similar but smaller elevation. The name-coded identity cue had comparatively modest effects overall, though it was not absent: it produced a reliable positive shift on prestige-oriented items such as award-worthiness and hiring-committee appeal. Crucially, the effects were not distributed evenly across the evaluation instrument. Content-quality judgments, such as clarity and methodological adequacy, barely moved. The distortions concentrated on the gatekeeping layer: top-tier acceptance, awards, funding, promotion, and the final accept-or-reject recommendation.

This selective vulnerability is what makes the result so consequential. It suggests the models are not applying a uniform halo effect that inflates every judgment when a prestigious name appears. Instead, they are perturbing a specific layer of evaluation tied to anticipated status, visibility, and institutional reward. The question-level decomposition showed the largest and most robust effects on items asking whether the manuscript meets the standard of a top-tier venue, deserves a best-paper award, warrants competitive funding, would impress a hiring committee, or is likely to change how the field thinks. Items about whether the writing is clear or the evidence sufficient were largely untouched. In other words, the metadata leaks into the decisions, not the diagnosis.

The study then pushed the analysis one step further, from scores to consequences. Using the acceptance recommendation as the ranking scalar, Rospocher recomputed manuscript rankings under counterfactual metadata swaps while holding everything else fixed. The results were striking. When a high-bibliometric profile was swapped for a low one, roughly a quarter to nearly two fifths of manuscripts originally selected in top-K shortlists, across thresholds from ten to twenty-five percent of the pool, fell below the cutoff after the perturbation. Institution-tier swaps produced an intermediate disruption, and name-cue swaps a smaller but still non-trivial one. Rank-displacement distributions were asymmetric, with downward movements more common and more extreme than upward ones, consistent with the finding that high-prestige cues systematically elevate scores. Even when most manuscripts stayed near their original position, those sitting near the selection boundary could cross the threshold, changing shortlist membership outright.

The audit also revealed meaningful variation across evaluator models. Random-effects meta-analysis showed that effect directions were broadly consistent, but magnitudes varied substantially, especially for prestige-related contrasts, with a large share of total variation attributable to between-model heterogeneity rather than sampling error. Inter-evaluator agreement diagnostics reinforced the picture: single models showed only modest absolute agreement on the same manuscripts, while averaging across all six improved reliability considerably. The practical implication is that metadata sensitivity is not a quirk of one model but a property of the model class, and that two pipelines differing only in their choice of evaluator may exhibit meaningfully different bias profiles. Domain-stratified analyses, meanwhile, found the qualitative pattern held across all four sampled fields, though magnitudes varied, and interaction tests suggested the three cue channels operate approximately additively rather than through narrow cue combinations.

The author is careful about scope. The study does not include a human-reviewer baseline, so it cannot say whether this sensitivity is unique to machines or inherited from patterns in human judgments embedded in training data. The corpus is limited to 80 English-language papers from four largely hard-science arXiv domains, the name cues are U.S.-centric and name-coded rather than ground-truth demographic attributes, and the experiment disabled browsing and retrieval to isolate the effect of explicitly supplied metadata. Real-world assistants that actively look up author information may introduce additional pathways not tested here. Still, the counterfactual logic is airtight within its design: identical texts, varied only in author context, produced different verdicts.

The design implications are immediate. The paper argues that author-context metadata should be treated as a consequential design variable, not neutral input, in any system that uses large language models for scientific evaluation, ranking, or discovery. In double-blind settings, LLM-assisted review should default to content-only evaluation with identifiers and lookup pathways suppressed. In single-blind or retrieval-augmented settings, systems should structurally decouple content scoring from context-informed summaries, document which metadata fields were available to the evaluator, and treat track record as an explicit, separately disclosed criterion if it is genuinely part of the decision rule, rather than letting it silently leak into manuscript scores. Because sensitivity varies by model, deployments should be audited whenever the evaluator is changed or updated. As AI-assisted reviewing spreads through editorial pipelines, the study delivers a clear warning: an algorithmic reviewer that can see who wrote a paper may not be able to stop that knowledge from coloring its judgment of the work itself.

Subject of Research: Counterfactual auditing of author-metadata sensitivity in large language model-based scientific peer review

Article Title: Author metadata affects Large Language Model scores in scientific peer review

Article References: Rospocher, M. (2026). Author metadata affects Large Language Model scores in scientific peer review. Discover Artificial Intelligence, 6(1), Article 1404. https://doi.org/10.1007/s44163-026-02268-y

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02268-y

Keywords: large language models, peer review, algorithmic bias, bibliometrics, institutional prestige, counterfactual audit, scientific evaluation, editorial triage, LLM-as-a-judge, ranking fairness, arXiv, research integrity

The core problem the framework addresses is familiar to anyone who has followed the AI governance debate: deployed language models are opaque, and the most sophisticated tools for interpreting them — mechanistic interpretability, feature attribution methods such as SHAP — are expensive, fragile, and in some cases provably limited. Impossibility theorems for feature attribution mean that no technique can reliably reveal why a model produced a given output in every case. Meanwhile, AI incidents continue to accumulate in databases cataloging real-world failures, and audits of large language models remain labor-intensive, point-in-time exercises rather than continuous monitoring. What is missing, the paper argues, is a measurement layer: a way to turn messy, free-form conversations between professionals and AI systems into structured, comparable data that institutions can actually monitor.

The proposed solution is a measurement standardization framework built around a grammar of eight interaction fields. Three of these are input–output fields — mission, conclusion, and justification — that capture what a professional asked the model, what the model recommended, and why. The paper illustrates how these fields transfer across domains: in a clinical setting, the mission might be recommending a diagnostic workup for a persistent cough in a heavy smoker, with the conclusion being urgent imaging justified by red-flag features for malignancy; in lending, the mission might be assessing a mortgage application with a low credit score, ending in a rejection justified by underwriting thresholds; in law, evaluating a settlement offer against comparable precedent awards. Because the fields carry the same structure regardless of domain, scores produced in one sector can in principle be compared with those in another.

The remaining fields feed a scoring engine. A large language model acting as a judge — the now well-established “LLM-as-a-judge” paradigm — evaluates each standardized interaction along dimensions such as risk level, policy alignment, and evidential alignment. Policy alignment measures whether the AI’s recommendation conforms to applicable guidelines and regulations; evidential alignment measures whether the justification is actually supported by the evidence cited. Crucially, the judge does not need access to the model’s internals. It works entirely from the observable interaction, which means the framework can be applied to any deployed system, including proprietary models whose weights are hidden even from the institutions using them.

Of course, using one black-box model to grade the outputs of another black-box model introduces what the paper candidly names structural circularity. LLM judges are known to suffer from systematic biases: sycophancy, the tendency to agree with whatever a user asserts; self-preference, the tendency to favor outputs resembling their own generations; and verbosity bias, the tendency to reward longer answers regardless of quality. There is also the problem of non-determinism — even supposedly deterministic settings can produce different judgments across runs, and small prompt changes can have butterfly-effect consequences for model performance. The framework confronts these weaknesses head-on rather than pretending they do not exist.

Its answer is a set of bounded conditions designed to reduce measurement inconsistency: explicit scoring rubrics, structured chain-of-thought reasoning before each verdict, reference documents that anchor judgments to authoritative guidelines, and low-temperature generation to suppress randomness. On top of these, the paper specifies a reliability verification procedure to detect and quantify the residual biases. The statistical machinery is equally explicit: paired bootstrap inference for the main comparisons, DeLong’s test for paired areas under the receiver-operating-characteristic curve as a sensitivity check, a pre-specified one-sided non-inferiority margin of 0.05, and Holm–Bonferroni correction to control for multiple testing. Agreement statistics drawn from the reliability literature, such as intraclass correlation coefficients and the Landis–Koch benchmarks for observer agreement, provide the yardsticks for judging whether the automated judge is consistent enough to be trusted.

To demonstrate feasibility, the paper applies the protocol in a minimal form to a published corpus of expert–AI interactions: the Rheum2Guide vignette study, in which ChatGPT’s treatment recommendations for rheumatic patients were compared against specialist decisions across nineteen clinical cases covering conditions from rheumatoid arthritis to giant cell arteritis. The judge scored each interaction for risk, policy alignment, and evidential alignment, using EULAR guideline documents as the reference corpus. The key finding is narrow but important: the judge reproduced its policy and evidential alignment scores across two independent runs under the specified conditions. In other words, the measurement instrument was stable. The author is careful to stress what this does not show — reliability at scale, across domains and models, remains a question for future empirical work, and the population-level claims the framework is designed to support are the subject of a staged research program, not results claimed in this paper.

If that research program succeeds, the payoff comes in three stages. The first is a reliability claim: under bounded conditions, language models can produce dependable, standardized assessments of expert–AI interactions. The second is a governance claim: alignment scores give experts an immediate signal during deployment — a warning light that flashes when a model’s advice drifts from policy or evidence — while giving institutions a basis for monitoring alignment patterns across mission types, models, and domains. The third, and most ambitious, is the outcome validation claim: once measurement is standardized, aggregate alignment scores could be correlated with downstream outcomes in regulated professional settings, exactly as cholesterol measurements were standardized and then linked to heart disease risk in the Framingham study, and as the lipid standardization programs of the CDC and National Heart, Lung and Blood Institute made population-scale cardiovascular epidemiology possible.

This is where the epidemiological analogy does its heaviest lifting. John Snow did not need to identify the cholera vibrio to map deaths around the Broad Street pump; Doll and Hill built the case against tobacco on correlated variables long before the molecular mechanisms of smoking-induced carcinogenesis were worked out. Hill’s famous 1965 criteria for distinguishing association from causation — strength, consistency, specificity, temporality, and the rest — were tools for reasoning rigorously in the absence of mechanistic certainty. The paper proposes importing precisely this style of reasoning into AI safety: instead of waiting for interpretability research to make model internals transparent, collect standardized measurements at the interface, look for statistical associations with harm, and intervene on the patterns that emerge. It is risk detection by surveillance rather than by dissection.

The framework is honest about its own limits. The demonstration involved a single judge, a single domain, and a small corpus; the circularity of black-box judging is mitigated, not eliminated; and unfaithful chain-of-thought reasoning — language models that do not always say what they think — remains a live concern even for structured scoring pipelines. But the conceptual move is significant. AI governance has long oscillated between two poles: demanding impossible transparency from opaque systems, or resigning itself to anecdotal incident reports. A standardized measurement layer offers a third path, one that medicine took a century to build and that could give institutions a shared language for describing, comparing, and ultimately predicting AI risk. Whether AI epidemiology can graduate from concept paper to working surveillance system will depend on the empirical studies the author has now carefully specified — but the blueprint for the experiment is on the table.

Subject of Research: A measurement standardization framework for detecting risks in deployed AI systems using epidemiological methods

Article Title: Toward AI epidemiology: a measurement standardization framework for prospective risk detection

Article References: Tempest-Walters, K. (2026). Toward AI epidemiology: a measurement standardization framework for prospective risk detection. AI & SOCIETY. https://doi.org/10.1007/s00146-026-03276-3

Image Credits: AI Generated

DOI: 10.1007/s00146-026-03276-3

Keywords: AI epidemiology, measurement standardization, LLM-as-a-judge, AI governance, risk detection, alignment, large language models, epidemiological methods, AI auditing, sycophancy, reliability, AI & Society

The scale of uptake was striking. Between September 2025 and January 2026, the chatbots recorded 6,291 question-and-answer exchanges across 1,534 individual sessions. Rheumatoid arthritis chatbots saw the heaviest use, followed by those for axial spondyloarthritis and ANCA-associated vasculitis. Usage peaked during the day but extended well into the evening, and engagement dropped off at weekends, a pattern consistent with evidence that patients reach for digital health tools precisely when their clinics are closed. The median conversation lasted just three turns, suggesting that most users were hunting for a targeted answer rather than settling in for an extended dialogue, and longer sessions attracted proportionally more feedback ratings.

What patients actually asked about is revealing. Using a rheumatologist-guided, human-in-the-loop automated coding pipeline, the team identified thirteen question categories. Half of all questions, 50.2 percent, were disease-specific, covering topics such as how to prepare for a doctor’s appointment to discuss vasculitis. Medication and monitoring questions accounted for 37.3 percent, with patients asking things like when their newly started drugs should begin to work. Diagnostics came next at 28.2 percent, followed by non-pharmacological interventions and lifestyle at 27.8 percent and prognosis and complications at 20.1 percent. Because coding was multi-label, a single question could touch several themes at once. These are not abstract definitional queries; they are the practical, often anxious questions that arise between consultations and that static leaflets were never designed to answer.

The technical architecture behind the chatbots is central to the story. Rather than letting a general-purpose model answer from its internal knowledge, the researchers used retrieval-augmented generation, or RAG, a technique that anchors every response in a predefined corpus of documents. Here, each chatbot’s knowledge base was restricted to the latest valid German clinical guideline for its disease. Guideline text was hierarchically segmented and embedded in a vector database using a domain-specific embedding model, and candidate passages were retrieved through a hybrid of lexical and dense-vector search, then ranked by a statistical reranker weighting vector and token similarity at seventy to thirty. Passages falling below a twenty percent similarity threshold were discarded, and up to eight were inserted into each prompt alongside source identifiers that were displayed to users. The underlying GPT-4o model, accessed through an Azure endpoint hosted in Germany, was never fine-tuned; behaviour was controlled entirely by retrieval and the system prompt, with generation settings held deliberately conservative at a temperature of 0.10 and top-p of 0.30 to favour reproducible output.

Patients noticed the difference. Of the 2,671 responses that received a rating, 92.9 percent earned a thumbs-up. Among the 190 negative ratings, the dominant complaint, at 65.8 percent, was not error but insufficient detail, a telling signal that users wanted more depth than a single guideline could supply. A total of 602 users completed the optional evaluation questionnaire, and their verdicts were largely favourable: 84.6 percent agreed the chatbot was easy to use, 84.1 percent found the answers easy to understand, and 80.2 percent considered it a useful addition to existing patient education materials. Some 77.4 percent said it would save them time searching for information, and 56 percent preferred it to general internet searches such as Google. Trust, however, was more qualified, with only 67.3 percent describing the answers as trustworthy, a reminder that credibility in AI health tools rests on more than accuracy alone.

On the critical question of safety, the automated evaluation was broadly reassuring. Applying a six-dimensional LLM-as-a-judge framework adapted from prior work, the researchers rated all 6,291 exchanges on correctness, question difficulty, completeness, patient-readiness, safety and patient-centredness. Fully 95.3 percent of answers were rated completely safe, 79.1 percent completely correct, and 70.4 percent highly patient-centred. Only 0.2 percent of answers, eleven in total, were flagged as potentially unsafe, and when two board-certified rheumatologists with more than a decade of experience each independently reviewed those flagged cases, just one was confirmed as genuinely problematic, involving a suggestion that cortisone could be administered subcutaneously during disease flares. For comparison, a recent red-teaming study of publicly available, non-RAG chatbots reported unsafe response rates of five to thirteen percent, although the authors caution that naturalistic and adversarial testing conditions differ too much to declare their system categorically safer.

Yet the study is equally notable for what it exposes. Only 45 percent of responses were judged fully adherent to the underlying guidelines, with 44.2 percent partially adherent and 10.9 percent not adherent at all, and agreement between the automated adherence ratings and physician assessment was weak. In some cases, strict fidelity to the uploaded guidelines produced answers that were outdated, such as a response stating that no Janus kinase inhibitors were approved for ankylosing spondylitis, reflecting the guideline text rather than current regulatory reality. In others, the model drifted beyond its curated corpus despite instructions and conservative settings. This exposes a fundamental design tension: confining a chatbot to a single guideline preserves transparency but leaves legitimate questions unanswered, while loosening the leash undermines the very source-grounding that makes the system trustworthy. The researchers argue the way forward lies in refined retrieval configuration and the controlled integration of additional verified content.

The methodology itself may prove as influential as the results. Evaluating thousands of exchanges by hand is impractical, so the team refined a rheumatologist-guided LLM-as-a-judge approach, using a locally hosted, data-secure open model to perform qualitative coding and structured quality assessment, with physician-coded samples used to calibrate the prompts. Agreement between automated and human coding was strong for question and feedback categories, with Gwet’s prevalence-robust AC1 coefficients reaching 0.91 and 0.96 respectively, and substantial for safety, patient-centredness, correctness and completeness. But the approach faltered on patient-readiness and question difficulty, where the model rated far fewer responses at the highest level than the rheumatologists did, a pattern consistent with known central tendency bias in LLM-based ordinal scoring. The authors are blunt about the implication: automated evaluation should be treated as a scalable screening tool, not a substitute for expert validation.

The study also has honest limits. Use was voluntary and may have attracted digitally confident patients; only 42.5 percent of responses were rated; no user accounts meant repeat sessions could not be excluded; and diagnoses were self-reported. Crucially, no patient outcomes were measured, so whether chatbot use actually improves knowledge, self-efficacy or adherence remains an open question. There was no real-time clinical safety monitoring or escalation pathway, only a system-prompt instruction to seek professional advice for medication changes and Azure’s general content filters. Still, the overall picture is one of cautious optimism. Guideline-grounded chatbots, co-designed with patient research partners from Germany’s largest rheumatology patient organisation, delivered predominantly safe, correct and well-received answers to thousands of real questions asked at all hours. Realising that promise at scale, the authors conclude, will demand robust source governance, faster guideline updates, transparent answer boundaries, continuous safety monitoring and, above all, sustained human oversight.

Subject of Research: Guideline-grounded large language model chatbots for patient education and self-management in rheumatology

Article Title: Development and Nationwide Multicentre Evaluation of Guideline-Grounded Large Language Model Chatbots to Support Patient Self-Management and Education in Rheumatology

Article References: Wilhelmi, T., Bartsch, V., Platt, M., Rashid, A., Hornig, J., Krusche, M., Hueber, A. J., Fink, D., Pfeil, A., Dischereit, G., Holzer, M.-T., Drott, U., Aries, P., Müller, M., Böhm, P., Morf, H., Labinsky, H., Mühlensiepen, F., Benavent, D., … Knitza, J. (2026). Development and Nationwide Multicentre Evaluation of Guideline-Grounded Large Language Model Chatbots to Support Patient Self-Management and Education in Rheumatology. Journal of Medical Systems, 50(1), Article 143. https://doi.org/10.1007/s10916-026-02470-6

Image Credits: AI Generated

DOI: 10.1007/s10916-026-02470-6

Keywords: rheumatology, large language models, chatbots, retrieval-augmented generation, patient education, self-management, artificial intelligence, clinical guidelines, rheumatoid arthritis, patient safety, LLM-as-a-judge, digital health

The study, published in the Journal of Ambient Intelligence and Humanized Computing by Roberto Casadei, Giovanni Delnevo, Barry Bassi, Chiara Ceccarini, and Silvia Mirri, is a position paper with an empirical backbone. Its central claim is that effective collaboration, like any good teamwork, requires all members to understand the problem, the environment, and each other. Yet while a vast literature examines the strengths and limitations of AI systems, far less attention has been paid to the risks created by a lack of maturity on the human side: users who over-trust plausible-sounding outputs, delegate too much reasoning, or hold mental models of the AI that bear little resemblance to what the system can actually do. The authors propose a formally structured model in which both actors continuously and proactively assess one another, closing the gap through what they describe as bidirectional scaffolding.

Technically, the framework models a human-AI team as a dynamic system state containing the user’s knowledge, behaviour, and configuration, the AI agent’s corresponding state, the shared goals, the task context, and the full process history, including logs, outputs, and plans. Crucially, each actor holds only its own view of this state, and those views may diverge from the ground truth and from each other. Collaboration falters when these internal representations drift apart. The proposed process framework therefore runs a cognitive pipeline: a grounding and calibration phase aligns user intent with AI capabilities; a cognition-tracking layer captures actions, decisions, and reasoning traces through what the authors call a Human Cognitive Interface; and a cognitive state representation feeds a simulation engine that generates candidate plans and hypothetical reasoning trajectories. Adaptive feedback, nudging, and targeted user assessments are then calibrated to the maturity of both partners.

Two novel concepts anchor the vision. The first, AI-in-the-human-loop, positions the AI not as a subordinate awaiting validation but as a reflective partner embedded within the human’s cognitive loop, observing and mirroring mental processes to foster metacognitive awareness and epistemic growth. The second, human-under-test, borrows an analogy from software engineering. Just as a subject-under-test is probed by test cases to verify and validate a system, the human partner can be gently assessed by the AI agent, with tests planned to cover key risks, detect regressions in understanding, and check compliance with shared requirements and factual knowledge. The authors are careful to note that the analogy is meant to highlight goals and procedures, not to dehumanise: the aim is to identify risks and promote quality, never to assign blame.

To test whether any of this is observable in practice, the team analysed a real-world collaboration: a programmer with no prior expertise in statistics who used ChatGPT 5.4 to analyse a questionnaire administered to secondary school students about craft-oriented education and the local footwear sector. The corpus comprised eleven documented chats, thirty human turns and thirty corresponding LLM turns, moving from broad exploratory questions about how to begin, through data auditing and item-type-specific analyses of open-ended, binary, single-choice, multiple-choice, ordinal, and reversed-Likert questions, to synthesis work producing charts, subgroup comparisons, and reusable scripts. The researchers then examined this workflow retrospectively using an LLM-as-a-judge setup, with two models from different providers, GPT 5.4 and Gemini 2.0 Flash, acting as independent evaluators across four dimensions: quality of the initial context, interaction issues, collaboration maturity, and missed opportunities for proactive AI intervention.

The findings are striking precisely because the project was, by conventional measures, a success. Both judges agreed that the initial context was generally good, with goals, data, and deliverables mostly clear. Yet the corpus still revealed non-trivial frictions. The clearest was an incorrect count produced by the operational LLM when processing a JSON file containing duplicate keys, an error caught only because the human explicitly checked it. Judges also flagged mismatches between the granularity of the model’s coding of open responses and the coding later adopted by the human, difficulties with artefact availability and format, and the need to explicitly request reusable outputs. Human researchers added further methodological concerns, including the formal treatment of grid-style questionnaire items and the validation of some statistical outputs, suggesting that automated judges may still need complementing by expert review.

Both evaluators also detected a maturation trajectory: early chats were exploratory and framing-oriented, while later interactions became structured by question type and oriented toward concrete deliverables. Interestingly, the two judges disagreed on where the collaboration ultimately landed. GPT 5.4 offered a cautious assessment, placing the human-AI dyad mainly between the Repeatable and Defined levels of a Humphrey-style maturity scale, while Gemini was more optimistic, situating it between Defined and Managed. The authors preserved rather than averaged this disagreement, noting that it may reflect differences between the judge models themselves. The most robust result, they argue, is the convergence in identifying recurring patterns and concrete points where a more proactive model, clarifying file semantics, agreeing on coding granularity, confirming which CSV columns to analyse, or setting minimal rules for subgroup comparisons, would have helped.

The study is candid about its limits. It examines a single analyst, a single operational model, a single task, and eleven chats; the analysis is retrospective rather than embedded in real-time interaction; the LLM-as-a-judge methodology may carry model-specific biases; and the framework itself has not yet been implemented as a complete system. Generalisability across users, domains, and tasks therefore remains an open question. Still, the exploratory evidence supports the paper’s core motivation: even a successful project can harbour maturity issues, collaborative frictions, and methodological misalignments that no one notices until an external assessor looks for them. In a world where millions of knowledge workers now lean on conversational AI daily, that observation alone is unsettling.

The broader implications reach into some of the most pressing debates in contemporary AI research. The paper situates its proposal against evidence that users tend to over-rely on AI advice even when it conflicts with contextual information, that LLM sycophancy can inflate trust without improving reliability, and that hallucinations persist even in retrieval-augmented systems, with models sometimes internally encoding the correct answer while generating a wrong one. It also draws on work showing that accurate user mental models improve decision-making, and that effective systems should promote calibrated trust rather than maximal trust. The authors’ framework treats these not as isolated problems but as symptoms of a missing feedback loop on the human side of the partnership, one that agentic AI, with its growing autonomy, makes increasingly urgent to close.

Many open questions remain before AI-in-the-human-loop becomes a practical architecture. What kinds of tests suit different aspects of a human-under-test, and how should they be generated contextually? How can expectations be defined, compliance quantified, and coverage estimated when human behaviour is fundamentally non-deterministic? How can testing be planned to avoid bothering the user, and designed to respect privacy and ethics? And, pointedly, who evaluates the evaluators? The Bologna team frames these as research challenges rather than settled answers, and calls for prototypes and quantitative studies to validate the vision. But the conceptual shift they propose is already provocative: AI systems designed not merely to perform tasks with users, but to help users better understand themselves, turning every collaboration into an opportunity for mutual learning and co-evolution.

Subject of Research: An operational model for bidirectional assessment and maturity in human-AI collaboration

Article Title: AI-in-the-human-loop: an operational model for mature human-AI collaboration

Article References: Casadei, R., Delnevo, G., Bassi, B., Ceccarini, C., & Mirri, S. (2026). AI-in-the-human-loop: an operational model for mature human-AI collaboration. Journal of Ambient Intelligence and Humanized Computing. https://doi.org/10.1007/s12652-026-05135-x

Image Credits: AI Generated

DOI: 10.1007/s12652-026-05135-x

Keywords: human-AI collaboration, AI-in-the-human-loop, human-under-test, large language models, metacognitive scaffolding, LLM-as-a-judge, mental models, trust calibration, human-AI co-evolution, AI maturity, hallucination, agentic AI

The research, led by Ghinevra Comiti with Paul-Antoine Bisgambiglia and Nathalie Lameta, focuses on Corsica, an island whose sharp contrasts make it an ideal testing ground. A heavily developed coastline sits against a declining rural interior; a tourism-driven economy produces seasonal swings; and demographic growth concentrates in a handful of urban centers. The team built their system on a corpus covering all 361 Corsican municipalities, combining a composite well-being indicator called OppChoLiv, inventories of municipal amenities, and demographic profiles with survey verbatims from 246 respondents across 70 municipalities, qualitative interviews, and aggregated questionnaire scores spanning 16 dimensions of well-being, from transport and health to safety and housing.

At the heart of the system is retrieval-augmented generation, or RAG, a technique that grounds the answers of a large language model in a controlled document corpus rather than in the model’s training data alone. Standard RAG, however, struggles with the kind of questions decision-makers actually ask. A single query may cross several thematic dimensions, several municipalities, and several demographic groups at once. Semantic similarity search over fragmented text chunks handles localized factual questions well but falters when the answer requires a global synthesis, an effect compounded by the well-documented tendency of language models to lose track of information buried in the middle of long contexts.

The proposed architecture rests on three principles. First, the user’s question is explicitly decomposed into five sub-questions, each processed independently with its own retrieval step, so that every semantic dimension of a complex demand gets dedicated coverage. Second, the document index is enriched with 674 precomputed multi-level syntheses built along ten business-driven analytical views, such as a specific well-being dimension in a specific municipality, an approach inspired by the RAPTOR method but replacing its automatic clustering with a controlled partitioning by survey metadata. Each synthesis is attached to a known population of respondents, which makes every summary attributable, addressable, and reproducible, qualities that matter when a synthesis must be presented to territorial actors alongside its scope.

The third principle is a strict typing rule. A factual, quantitative query can only spawn factual sub-questions, and a qualitative query only qualitative ones, preventing the system from fabricating sub-questions for which no documentary support exists. Entity extraction is deliberately deterministic: municipalities are matched against a closed list, thematic dimensions against a fixed taxonomy of 16 categories, guaranteeing perfect precision at no inference cost. Retrieval then proceeds hierarchically, always prioritizing the most specific representation available, and injecting both the aggregated synthesis and the raw chunks behind it, so the generation model sees the big picture and the concrete evidence at once.

The team evaluated the architecture on a benchmark of 109 manually crafted questions spanning seven categories, from simple factual retrieval to causal reasoning, uncertainty handling, and deliberate trick questions about missing data. Answers were scored by an independent language model judge, GPT-4o, whose reliability was calibrated against human annotations and cross-checked against three alternative judges. The headline retrieval result is striking: the full architecture achieved an average Recall of 0.970, against 0.678 for vanilla RAG, a gap of nearly 30 percentage points reflecting a far more exhaustive mobilization of the available evidence.

On overall answer quality, the full system scored 4.62 out of 5, statistically indistinguishable from vanilla RAG’s 4.63, but the aggregate hides a crucial pattern. The advantage concentrates precisely where decisions are hardest: comparative reasoning, causal and counter-intuitive questions, and the delicate crossings between what indicators say and what inhabitants feel. Ablation tests confirmed that both the typing rule and the multi-level syntheses contribute significantly, and that removing the typing rule alone drops the score from 4.30 to 3.72, with the damage concentrated on targeted factual queries where an untyped decomposition spawns qualitative sub-questions that have no matching sources.

Case studies reveal why the design matters. Asked whether objective indicators and perceptions diverge in Bastia, vanilla RAG retrieved only quantitative scores and speculated about verbatims it never saw, while the full system articulated both registers, noting that mobility was rated objectively at 8.65 out of 10 even as residents scored it 2.24 out of 5, complaining of congestion. On a trick question about a municipality with good indicators but low perceived well-being, the full system correctly recognized that the available perception data could not be disaggregated by municipality, refusing to over-interpret a quantitative proxy as a measure of feeling. Yet the architecture has a documented weakness: on questions about missing information, its appetite for assembling rich contexts sometimes tempts it to approximate where a simpler system would refuse cleanly.

The authors are candid about costs and limits. The full pipeline issues roughly six times more language model calls per query than vanilla RAG, with latency about 4.6 times higher, though the offline synthesis construction cost only about $0.16 and is amortized once. Section-level differences often lose statistical significance after correction, and the Corsican corpus is entirely in French. The method transfers to other territories only where surveys carry structured metadata comparable to theirs. The team’s proposed remedy is a two-tier system that routes simple factual questions to cheap direct retrieval and reserves the heavyweight machinery for genuinely analytical queries, alongside an agentic extension in which a controller decides actively whether to answer or refuse, a direction they identify as the immediate next step toward AI assistants that policymakers can actually trust.

Subject of Research: A hybrid quantitative-qualitative retrieval-augmented generation architecture for territorial well-being decision support

Article Title: A hybrid quantitative–qualitative RAG approach for territorial well-being decision support

Article References: Comiti, G., Bisgambiglia, P.-A., & Lameta, N. (2026). A hybrid quantitative–qualitative RAG approach for territorial well-being decision support. Machine Learning with Applications, 26, Article 101005. https://doi.org/10.1016/j.mlwa.2026.101005

Image Credits: AI Generated

DOI: 10.1016/j.mlwa.2026.101005

Keywords: retrieval-augmented generation, large language models, territorial well-being, decision support systems, Corsica, survey data, RAPTOR, query decomposition, hallucination control, public policy, composite indicators, LLM-as-a-judge