The demand for such an authority is enormous. Correctness is not a marginal preference; it is an ideology deeply rooted in the school system, in educational evaluation, and in professional life, where linguistic accuracy is often rewarded and deviation penalized. Digital grammar checkers have been meeting this demand for decades, but large language models have expanded the possibilities dramatically. With a single prompt, users can obtain corrections of grammar, spelling, and style in seconds, free of charge, and without access to a human proofreader. In classrooms where a teacher may not have time to correct thirty essays, outsourcing the task to an AI system is becoming an everyday reality. The researchers illustrate this with a real case: a pupil was asked to write a short text and have it corrected by an AI because the teacher lacked time. That single classroom decision, they note, already signals to students that AI corrections deserve the same trust as a teacher’s red pen.

Yet what exactly is corrected, how, and with reference to which resources remains largely unexplored. This matters because less linguistically informed users are the most likely to accept AI-generated corrections uncritically. Previous studies have produced worrying hints: AI systems may give more corrections than human editors, sometimes over-correcting perfectly acceptable language, sometimes producing fluent but misleading fixes, and occasionally introducing fresh grammatical errors into edited text. One recent classroom study found that German learners more readily adopted colloquial forms when they believed an AI had generated them. The apparent simplicity of AI-based correction, Dovalil and Hanulíková argue, conceals a far more complex sociolinguistic question.

To answer it, the authors first rethink what standard language actually is. Against the purely descriptivist view, in which a variant counts as standard when it appears frequently in corpora and model texts, they draw on Ulrich Ammon’s model of the social forces that determine what is standard in a language. In this framework, standard varieties are not fixed properties of a language but sociolinguistic concepts negotiated among four groups of actors: codifiers such as the authors of dictionaries and grammars, linguists, model speakers and writers, and norm authorities, defined as persons whom other language users expect to correct their language production. Crucially, this expectation must be legitimate. When a user addresses a grammatical doubt to another person, they expect trustworthy and reliable information in return. The question is whether an LLM can fulfill that role, and the answer, the authors insist, cannot be read off from statistics alone.

Herein lies a subtle technical point about norms. Frequencies, distributions, and regularities are often called norms in the broad sense, but normative validity and the factuality of language use must be strictly separated. The assumption that actual usage automatically confers normative force does not stand the test: regularities may arise from language norms but also from many other sources, so the inference from pattern to prescription is not compelling. LLMs, trained on vast collections of human-written text and tuned through preference optimization and post-training alignment, ultimately generate their linguistic behavior from statistical learning rather than explicit grammatical codification. Their transformer architectures encode words as vectors, operating on the premise that words appearing in similar contexts have similar meanings. When such a model corrects a text, it is pattern matching, not metalinguistic reasoning in the informed sense, and recent cognitive science suggests the two can dissociate. Worse, the researchers note a feedback loop: AI-generated metalinguistic responses become part of the future linguistic environment on which models are trained, a self-referential production of language use that the authors compare to autopoiesis, in which the boundary between language use and standard norm grows increasingly blurred.

To move from theory to evidence, the team designed an empirical test in German, a language rich in well-known cases of grammatical doubt. They constructed a short fiction story containing thirteen targeted variants across eight grammatical variables, seven inflectional and one syntactic. The variables included the case government of the prepositions wegen and gegenüber, where codification recommends genitive and dative respectively; the declension of demonstrative pronouns in genitive constructions such as das Leben deren versus derer and Ende diesen versus dieses Monats; parallel declension of consecutive attributive adjectives in noun phrases without articles, as in mit großem finanziellen Aufwand; declension of adjectives following the pronoun alles; the uninflected language name in im modernen Irisch; the weak masculine noun Präsident after a preceding genitive, as in mit Irlands Präsidenten, where a trend toward reduced declension collides with the absence of a declension trigger; and word order in um…zu constructions with verba dicendi, the so-called Echo-Satz pattern. Corpus frequencies were examined for every variant pair, mostly aligning with the codified form, sometimes strongly, as in wegen dem versus des at 93 to 3779, though one variant, mit großem finanziellen Aufwand versus finanziellem, diverged at roughly 3 to 1 in favor of the non-codified form.

The same text was then submitted to six proofreading and LLM tools in eight model configurations: ChatGPT 5.5, Copilot M365, Gemini Flash and Flash-Lite, DeepL Write in corrections-only mode, Claude Sonnet 5 in medium and Max reasoning effort settings, and Perplexity. Free, non-subscription versions were used, memory functions were disabled, and each tool corrected the identical text at three time points spaced five days apart in July 2026. The prompt framed the text as a school assignment and asked for grammar corrections only, treating the models exactly as users treat a norm authority.

The results were striking. Correction behavior varied considerably both between models and within individual models over time, and no model corrected all targeted phenomena at any time point. Correction rates ranged from 8 percent for DeepL to 77 percent for Gemini Flash at the first time point, or 85 percent once reformulations were added. No model showed improvement over time. Some configurations fluctuated wildly, with Gemini Flash sliding from 77 to 62 to 38 percent across the three sessions, while others remained stable, such as Claude Max reasoning at 69 percent. Except for Claude and DeepL, all models also produced a third type of behavior, reformulating rather than directly correcting or leaving a targeted construction untouched. Only one variant, wegen dem corrected to wegen des, was managed consistently in all instances, whereas one, mit Irlands Präsidenten, was never corrected by any model. Duplicate occurrences of the same variant were always handled identically, but lexical realizations of the same underlying rule diverged: while mit Irlands Präsidenten went uncorrected, mit Deutschlands Präsidenten was corrected in 12 percent of cases. Most tellingly, the syntactic um…zu variant was corrected only twice despite corpus data showing a strong preference for the corrected word order at 1 to 283. Correction behavior, the authors conclude, is not solely a function of model capability or reasoning effort and cannot be reliably predicted across repeated queries, undermining the intuitive assumption that repeated use of the same tool yields increasingly consistent corrections.

Approximately half of the corrections complied with codification and model texts, a level of limited consistency that falls short of what norm authorities are expected to deliver. The authors are careful not to overstate the indictment: human correction practices among teachers and linguistic experts are themselves heterogeneous, so LLMs simply add another set of inconsistent instances to an already pluralistic landscape. Interviews conducted by the researchers with laypeople and linguists reveal sharply different usage patterns. Laypeople trust AI opinions somewhat more, treating the models as norm authorities and citing the sheer amount of training data as a guarantee of quality, whereas linguists remain more careful before implementing AI-designed adjustments. This shows, the authors write, that LLMs may be turned into a norm authority provided a user does so. Whether that transformation sticks may depend on a phenomenon known as algorithm appreciation, the finding that people sometimes adhere more to advice they believe comes from an algorithm than from a human.

The crucial issue for future research, the authors conclude, is therefore not whether LLMs can manage standard language, but whether statistical text generators should acquire the social legitimacy traditionally granted to human norm authorities. They sketch conditions under which that legitimacy might be earned: if models can be shown to draw transparently on knowledge codified by reliable sources, dictionaries, grammars, expert linguistic literature, or even correction practices completed by human authorities, they could be treated as an autonomous norm authority. Until then, the researchers recommend AI as a powerful proofreading supplement to, not yet a replacement for, human judgment on standard-related issues. Answering the legitimacy question, they stress, cannot be done through benchmark evaluations alone; it requires integrating sociolinguistic theory, language ideology, and studies of human-AI interaction. The red pen, in other words, has not been retired. It has merely acquired a new, statistical rival whose authority rests on trust that has not yet been justified.

Subject of Research: Language management in the age of LLMs: AI tools in proofreading as emerging norm authorities?

Article Title: Language management in the age of LLMs: AI tools in proofreading as emerging norm authorities?

Article References: Dovalil, V., & Hanulíková, A. (2026). Language management in the age of LLMs: AI tools in proofreading as emerging norm authorities?. Frontiers in Language Sciences, 5, Article 1957524. https://doi.org/10.3389/flang.2026.1957524

Image Credits: AI Generated

DOI: 10.3389/flang.2026.1957524

Keywords: large language models, LLMs, proofreading, standard language, norm authority, language management, grammar correction, sociolinguistics, AI trust, German grammar, Language, management

The data bottleneck the researchers set out to break is a familiar one in medical machine learning. High-quality psychological interviews are expensive to collect, slow to transcribe and annotate, and guarded by strict privacy rules because they contain some of the most sensitive information a person can disclose. Cloud-based commercial language model APIs, which might seem like an easy shortcut for generating substitute data, raise their own risks of exposing medical records to third parties. The Sichuan University team, led by Yixuan Liu with Junjie Hu, Yuanyuan Chen and Zhang Yi, designed PRISM — short for Psychological Role-playing for Interview Synthetic Modeling — to sidestep both problems at once, generating interview dialogues locally without ever touching real patient records.

PRISM works in three coordinated stages, each of which tackles a different failure mode of naive text generation. The first is a theme-based Markov model that learns how real clinical interviews unfold. Rather than letting a language model ramble freely, the Markov component captures the statistical structure of interview flow — which topics tend to follow which, when interviewers probe deeper, and how conversations drift between small talk and symptoms. This gives the synthetic dialogues a clinically plausible skeleton, mimicking the way trained interviewers move from opening questions about daily life toward questions about sleep, mood and diagnosis history.

The second stage supplies the psychological depth. The team built a multi-dimensional persona template grounded in the biopsychosocial model, the long-standing clinical framework that views mental health as the product of biological, psychological and social factors. Each synthetic patient is generated with a sampled age, gender and symptom profile drawn from the PHQ-8 depression questionnaire, covering eight dimensions including loss of interest, depressed mood, sleep problems, fatigue, appetite changes, self-blame, concentration difficulties and psychomotor changes. The template deliberately leaves many fields blank — hobbies, recent life events, social support, marital status, education, financial situation, personality — and instructs the language model to fill them in, ensuring each persona is internally consistent before a single line of dialogue is spoken.

The results of this persona construction are strikingly lifelike. In the paper’s demonstration, a 51-year-old male persona named Ethan Granger emerges as a career Army officer recently returned from a year-long deployment, suffering chronic back pain and occasional insomnia, reluctant to discuss personal struggles even with a supportive spouse, and stoic to the point of being described as locked up. When interviewed, Granger deflects questions about his military assignments, admits sleep has been better than a few weeks ago but still troubled, and acknowledges regrets that stick with him — exactly the kind of guarded, symptom-laden conversational texture that real clinical interviews capture and that simpler synthetic data lacks.

The third stage is a two-stage, metric-driven optimization strategy that decides how much synthetic data to generate and how useful it is. Rather than assuming more data is always better, the framework measures the quality and diversity of generated dialogues using established text-generation metrics, including BLEU scores for fluency and perplexity-based filtering to catch degenerate outputs. The filtering rule proves its worth in the paper’s own examples: when a generated response contains a garbled second sentence with perplexity above threshold, only the clean first sentence is retained. This quality control keeps the synthetic corpus from polluting the training signal with the hallucinations and artifacts that large language models are known to produce.

When the PRISM-generated interviews were used to augment training data for depression detection on the DAIC-WoZ benchmark — a widely used corpus of semi-clinical interviews collected by the University of Southern California’s Institute for Creative Technologies — the results were dramatic. Macro-F1, a metric that balances performance across both depressed and non-depressed classes, rose from 0.61 to 0.762, comfortably surpassing zero-shot baselines in which large language models attempt detection directly without task-specific training. The improvement matters because depression detection is a heavily imbalanced problem: depressed individuals are a minority in most datasets, and models that optimize naive accuracy can quietly fail the very people they are meant to help.

Generalizability was tested on two fronts. Cross-dataset experiments on the Chinese EATD corpus, an emotional audio-textual depression dataset, confirmed that the benefits of PRISM-augmented training transfer beyond the original English-language benchmark. The team also verified that the synthetic data improves multiple underlying network architectures rather than being an artifact of one particular model, suggesting the approach is a general-purpose data remedy rather than a model-specific trick. The work builds on a growing line of research into role-playing language models and LLM-simulated patients, including systems designed to train mental health professionals, but PRISM is distinguished by its explicit grounding in clinical interview structure and biopsychosocial persona modeling.

The implications reach beyond depression detection. The study demonstrates that carefully structured role-playing can turn general-purpose language models into sources of clinically plausible training data for mental health tasks, potentially easing data scarcity across psychiatry-adjacent machine learning problems where privacy rules and annotation costs have long throttled progress. Because the framework avoids cloud APIs and real records, it offers a template for institutions that want to pursue synthetic data generation without compromising patient confidentiality. The researchers caution that synthetic data is a supplement, not a replacement, for real clinical evidence, and that hallucination in language models remains an open challenge their filtering pipeline only partially addresses.

For a field where the gap between laboratory benchmarks and clinical reality remains wide, PRISM offers a pragmatic path forward: teach the machine not just to recognize depression, but to imagine the conversations in which it reveals itself. With depression detection models now capable of exceeding 0.76 Macro-F1 on standard benchmarks using purely synthetic augmentation, the study marks a step toward screening tools that could one day work in settings where real interview data simply does not exist — and toward a future in which the hardest part of building a mental health AI is no longer finding the data, but asking the right questions of the machines that generate it.

Subject of Research: Advancing depression detection with role-playing LLMs and synthetic psychological interviews

Article Title: Advancing depression detection with role-playing LLMs and synthetic psychological interviews

Article References: Liu, Y., Hu, J., Chen, Y., & Yi, Z. (2026). Advancing depression detection with role-playing LLMs and synthetic psychological interviews. International Journal of Machine Learning and Cybernetics, 17(11), Article 499. https://doi.org/10.1007/s13042-026-03337-5

Image Credits: AI Generated

DOI: 10.1007/s13042-026-03337-5

Keywords: Advancing, depression, detection, role-playing, LLMs, synthetic, psychological, interviews, scientific research

The task in question is named entity recognition, or NER, a foundational technique in natural language processing that involves locating and classifying mentions of things like persons and organisations within raw text. In the legal world, NER is the engine behind contract analysis, document redaction and large-scale information extraction. The trouble is that legal language is notoriously hostile to generic models: entities appear as abbreviations, acronyms and synonyms, and off-the-shelf tools trained on general English text stumble badly when confronted with contract boilerplate.

Traditionally, high-performing NER has required either hand-crafted rules or vast amounts of manually annotated training data — a bottleneck that is especially painful in law, where labelling must be done by qualified legal experts. The research team, led by Shefali Singhal and Niall McCarroll, took a different route. They repurposed the Contract Understanding Atticus Dataset (CUAD), a collection of 510 real contracts spanning 25 different contract types, into a new NER benchmark called CUAD-NER, with five legal domain experts labelling every PERSON and ORGANISATION entity — nearly 100,000 organisation mentions in total.

Against this dataset they lined up an unusually broad field of competitors: the industry-standard SpaCy pipeline, Microsoft’s Azure Entity Recognition V3 service, and four transformer models — DistilBERT, BERT base (cased), RoBERTa and DeBERTa — each fine-tuned on the legal contract data using standard low-rank adaptation (LoRA) settings on Azure’s machine learning platform. The challenger was GPT-4, deployed not with any fine-tuning at all, but with a carefully structured few-shot prompt that included ten worked examples written in a chain-of-thought style, adapted from the PromptNER strategy of Ashok and Lipton.

The prompt itself is a study in prompt engineering discipline. It contains three components: a definition section spelling out exactly what does and does not count as a PERSON or ORGANISATION entity, the contract text to be analysed, and the ten few-shot exemplars that force the model to reason step by step — deciding whether a given phrase is an entity, justifying the decision, and then assigning a label of ORG, PER or NA. The exemplars were chosen to give balanced coverage of positive identifications and negative rejections, and the team deliberately avoided optimising over example choice or ordering, aiming instead for the kind of reliable average-case performance a practitioner would actually deploy.

The results were striking. GPT-4 with few-shot chain-of-thought prompting achieved a Micro F1 score of 73 percent across both entity types — a five percentage point improvement over the best fine-tuned competitor, BERT base (cased), which scored 69 percent. RoBERTa managed 67 percent, while DistilBERT and DeBERTa each reached 66 percent. The generically trained baselines fared far worse: Azure Entity Recognition V3 scored just 53 percent, and SpaCy collapsed to a dismal 10 percent, mislabelling common nouns like

The five-point margin in favour of GPT-4 is notable not only for its size but for what it represents: a model that never saw a single labelled contract during training outperforming systems that were explicitly fine-tuned on the very same corpus. The fine-tuned transformers in the study were adapted using LoRA, a parameter-efficient technique that freezes the bulk of a pre-trained model’s weights and trains only small, low-rank update matrices. This keeps training costs manageable compared with full fine-tuning, but it still requires the annotated corpus, the compute infrastructure, and the engineering effort to run the adaptation — all of which the few-shot approach sidesteps entirely. For organisations that lack teams of legal experts available to annotate documents, that difference in workflow may matter as much as the difference in F1 score.

The Micro F1 metric used in the evaluation blends precision and recall into a single figure, and in an entity recognition task it treats every mention as equally important. That makes it a reasonable headline number, but the underlying trade-offs are worth considering. A redaction system, for instance, may tolerate false positives — flagging a phrase as an entity when it is not — far more readily than false negatives, which would leave sensitive names exposed in a document. Conversely, an information extraction pipeline feeding a contract analytics dashboard might prefer precision, since spurious entities clutter downstream analysis. The study’s aggregate scores do not by themselves reveal how each model distributes its errors between these two failure modes, and practitioners deploying such systems would likely want per-class breakdowns before committing to a choice.

The collapse of SpaCy to a 10 percent score is a reminder of how domain-specific legal text really is. SpaCy’s general-purpose pipeline relies on statistical models trained largely on news wire and web text, where entities tend to appear in predictable contexts — a person’s name adjacent to a title or a company name followed by corporate suffixes. Contract language inverts many of these expectations: defined terms appear in quotation marks, parties are referred to by capitalised shorthand after their first mention, and organisational names can be embedded deep within nested clauses. The result is a cautionary tale about assuming that tools which perform well on general text will transfer to specialised professional domains without substantial retraining.

The choice of CUAD as the foundation for the new benchmark also carries significance. CUAD was originally designed for contract clause classification — identifying which of 41 clause types appear in a given contract — rather than for token-level entity tagging. Converting it into CUAD-NER required the five legal experts to annotate every PERSON and ORGANISATION mention across roughly 100,000 organisation references, a labour-intensive process that itself illustrates the annotation bottleneck the paper seeks to address. The fact that such a dataset had to be created by hand underscores the economics at play: once a labelled corpus exists, fine-tuned models become viable; without one, few-shot prompting may be the only practical option.

The cost–benefit analysis that accompanies the performance comparison is arguably the study’s most distinctive contribution. Academic NER papers routinely report F1 scores but rarely model the financial realities of deployment, where inference costs scale with usage volume. The study distinguishes between low-usage and high-usage scenarios, recognising that the economics of calling a commercial LLM API differ fundamentally from hosting a fine-tuned open-weight model. At low volumes, the absence of training and infrastructure costs gives the few-shot approach an advantage; at high volumes, the per-token pricing of a commercial API can accumulate to the point where a self-hosted fine-tuned model, despite its lower accuracy, becomes the more economical choice. The crossover point between these regimes is exactly the kind of information practitioners need but rarely find in the literature.

It is also worth situating the findings within the broader trajectory of LLM capabilities. GPT-4’s performance here reflects a model evaluated without any task-specific adaptation, relying purely on in-context learning — the ability to infer a task’s requirements from examples embedded in the prompt. Chain-of-thought prompting, the technique used to elicit step-by-step reasoning before the model commits to a label, has been shown across many benchmarks to improve performance on tasks requiring deliberate judgement rather than pattern matching. Entity recognition in legal text sits somewhere between these categories: it involves both surface pattern recognition and contextual judgement about whether a phrase refers to a person or an organisation in a given contractual role. The structured reasoning the prompt enforces appears to help the model navigate this ambiguity more effectively than direct labelling would.

The study’s authors acknowledge that the few-shot approach is not without limitations. Performance on unfamiliar, domain-specific NER tasks had remained uncertain precisely because LLM behaviour on such tasks is sensitive to prompt design, and the team deliberately avoided optimising over example selection or ordering to reflect realistic deployment conditions. This choice favours robustness over peak performance, but it also means that a practitioner willing to invest in prompt refinement — selecting exemplars that better cover edge cases, or tuning the definition section — might achieve results beyond those reported. Conversely, the fine-tuned baselines were trained with standard LoRA settings rather than exhaustively hyperparameter-optimised, so neither side of the comparison represents a ceiling on what is achievable.

Looking forward, the authors suggest that few-shot chain-of-thought learning could be strengthened with additional support strategies, and that flexible prompting approaches are likely to prove especially valuable in low-resource scenarios where annotated data is scarce. The legal domain is a prime example: contracts are confidential, annotation requires expertise, and the diversity of contract types means that a model fine-tuned on one organisation’s documents may not generalise to another’s. In such settings, the ability to obtain competitive performance from a general-purpose model with nothing more than a well-crafted prompt and a handful of examples represents a genuine shift in the accessibility of legal NLP technology — one that could bring sophisticated document analysis within reach of smaller firms and legal aid organisations that could never afford the annotation campaigns traditional approaches demand.

The broader implication for the field is that the evaluation of NLP systems is maturing beyond accuracy metrics alone. By pairing performance figures with a structured cost analysis, this study offers a template for how future work might be framed: not as a simple leaderboard, but as a decision-support tool that acknowledges the real constraints under which organisations operate. As LLMs continue to improve and API pricing evolves, the balance between prompting-based and fine-tuning-based approaches will shift, but the framework for comparing them — performance, cost, and the practical realities of deployment — is likely to endure.

Subject of Research: Evaluating LLMs for named entity recognition: a cost-benefit analysis in the legal domain

Article Title: Evaluating LLMs for named entity recognition: a cost-benefit analysis in the legal domain

Article References: Singhal, S., O’Connell, E., McCarroll, N., Keogh, N., Doherty, M., Munasinghe, V., Rani, S., Curran, K., McNamee, E., Clist, A., & Brammer, A. (2026). Evaluating LLMs for named entity recognition: a cost-benefit analysis in the legal domain. Knowledge and Information Systems, 68(1), Article 288. https://doi.org/10.1007/s10115-026-02886-z

Image Credits: AI Generated

DOI: 10.1007/s10115-026-02886-z

Keywords: Evaluating, LLMs, named, entity, recognition, cost-benefit, analysis, legal, domain, scientific research

The research team, led by Marina Drushlyak and Olena Semenikhina of Sumy State Pedagogical University together with Svitlana Martos and Svitlana Klymovych of Kherson State University, built their study around a construct they call research reasoning. They operationalized it through three indicators. The first, problematization, covers the precise formulation of a research question and the boundaries of a topic. The second, methodological justification, captures the conscious choice of procedures, criteria for selecting sources, and methods of analysis. The third, reflective evaluation, involves checking the reliability of results and recognizing the limitations of methods, including the risks posed by generative artificial intelligence. Students were then classified as basic, intermediate, or advanced depending on how far their work moved along this spectrum.

Crucially, the researchers did not treat the tools themselves as the object of learning. Instead, they identified three mechanisms through which digital services can shape reasoning. The first mechanism involves external representations and the structuring of the research field: knowledge maps, citation graphs, and timelines that transform a vague search into genuine problematization. The second is heuristic generation, in which language models and recommendation systems accelerate the production of alternatives and preliminary models, but only when explicit rules for source selection are applied. The third mechanism is the organization and verification of evidence, supported by services such as Scite, Elicit, and Consensus that screen for contradictory citations and help verify claims. In the authors’ framing, tools operate as links in a chain running from tool to procedure to evidence, rather than as isolated user skills.

The qualitative heart of the study lies in the students’ own work. One participant described using ResearchRabbit before citing an influential article, discovering that although the paper had accumulated more than a hundred citations, fifteen of them were contradictory and pointed to methodological flaws. That discovery prompted a critical reassessment of the article’s reliability, and the student ultimately cited it while noting the existing criticisms. Another student, after Connected Papers revealed twelve contradictory citations attached to a key theoretical source, went so far as to revise the entire introduction of a literature review, adding a section on the scientific debates surrounding the theory. These are not the actions of students outsourcing their thinking to software; they are examples of tools triggering disciplined engagement with evidence.

Large language models played a different but complementary role. Several students used ChatGPT or Claude not to generate content but to attack their own arguments. One presented a preliminary structure to ChatGPT and asked it to identify weaknesses in the reasoning; the model suggested three alternative perspectives, two of which the student integrated into a Research Limitations section. Another uploaded a written conclusion to Claude with a prompt asking the system to assess the strength of the arguments and identify logical errors, and discovered that one argument was circular. The student rewrote it. In both cases, the researchers note, the artificial intelligence acted as a trigger for reflection, but the conclusions were formulated by the student, a distinction that sits at the core of their pedagogical model.

The survey component of the study, while descriptive rather than explanatory, helps map the landscape of tool use. Among the 104 respondents, 56 percent mentioned large language models in open-ended responses, 33 percent cited AI literature-review tools based on semantic search, 22 percent noted bibliographic managers with AI features, and 7 percent mentioned citation-based mapping tools. Google Scholar dominated literature searching at 85 percent, though it received a lower effectiveness rating than Scopus because of filtering difficulties. Excel remained the most common analysis tool at 67 percent, while Python and R were rated as flexible but burdened by a steep entry barrier. Most participants reported combining three to five tools, a pattern the authors interpret not as fragmentation but as a rational division of labor across the mechanisms of structuring, generation, and verification.

Educator observations, recorded in field memos over two academic years, added an external check on the qualitative themes. The observers used a rubric ranging from a procedural minimum, in which isolated technical actions bear no connection to the research goal, up to reflective integration, in which evidence verification, framework revision, and the delineation of AI boundaries become systematic. Most cases fell into the middle levels, where tools function as conveniences or as instruments of modeling, while the highest level occurred sporadically but carried methodological significance. The memos also documented failure modes: one student copied an AI-generated source list into a paper without checking accessibility or formatting, and three of the seven references turned out to be unusable. Such episodes illustrate the gap between possessing a tool and possessing a method.

That gap is precisely what the study identifies as the central challenge. The authors describe three groups of barriers that keep students trapped at superficial levels of digital practice. Technical difficulties, including unstable access, format incompatibilities, and synchronization failures, fragment digital work into disconnected operations. Conceptual barriers arise when students do not understand how a tool supports the research task, breaking the chain from tool to procedure to evidence into isolated actions. Reflective barriers appear when generated texts, source lists, and summaries are accepted without critical revision. The proposed remedy is not more software but more structure: mandatory micro-sections in reports asking what could be wrong, verification of counter-claims for key sources, brief methodological commentaries explaining each tool’s function and limits, and group work organized around collective evidence audits rather than mutual technical assistance.

The course itself offers a template for how this might work in practice. Students completed personalized micro-projects tied to their dissertation topics: analyzing research landscapes with VOSviewer, constructing academic genealogies with the Mathematics Genealogy Project, classifying sources in Zotero, and visualizing key concepts with Canva. Reflective mini-essays required them to articulate the limitations and potential of their chosen tools and to analyze the ethical challenges of AI use. One student described how an initial keyword map in VOSviewer revealed a separate cluster that clarified the boundaries of a review; another explained how a network graph exposed methodological articles previously overlooked, changing the structure of a section. In each case, the visualization served as a framework for synthesis rather than an end in itself, which is exactly the disposition the researchers hoped to cultivate.

The authors are careful about the limits of their claims. The empirical material is qualitative, drawn from two Ukrainian universities, and the survey served a contextual and triangulation function rather than testing hypotheses or measuring change. No theme frequencies or inter-coder coefficients were calculated; the analysis prioritized reflexive thematic analysis, thick description, and analytical generalization. Future work, they suggest, should include independent pre- and post-course measurements and quantitative validation of the reasoning indicators. Even so, the implications are hard to ignore at a moment when generative AI is flooding academia with fabricated references and confident nonsense. The study’s central message is that digital literacy at the doctoral level cannot mean tool mastery alone. It must mean the ability to think through platforms, to verify what they produce, and to take responsibility for the chain of evidence that connects a search result to a scientific conclusion. Institutions that treat AI as a threat to be banned, the findings suggest, may be missing the more productive opportunity: teaching the next generation of researchers to reason with it, critically and on the record.

Subject of Research: How digital and AI tools integrated into doctoral methodological training shape PhD students' research reasoning

Article Title: Developing research reasoning through digital tools in PhD students’ methodological training

Article References: Drushlyak, M., Martos, S., Klymovych, S., & Semenikhina, O. (2026). Developing research reasoning through digital tools in PhD students’ methodological training. Discover Education, 5(1), Article 1001. https://doi.org/10.1007/s44217-026-02110-8

Image Credits: AI Generated

DOI: 10.1007/s44217-026-02110-8

Keywords: PhD education, research reasoning, digital tools, artificial intelligence, generative AI, methodological training, digital literacy, citation analysis, open science, LLMs, evidence verification, doctoral training

The core problem the authors identify is structural. Archaeologists work from material remains that underdetermine the social behaviors that produced them, so theory tends to cluster into interpretive camps rather than converge on consensus. Even the influx of scientific methods, from ancient DNA to isotope analysis, has largely been used to support pre-existing positions rather than adjudicate between them. As the authors note, relativist perspectives have even suggested that plurality of interpretation should simply be accepted, since theory in archaeology may never be measurable in the way it is in other sciences. What they propose instead is to narrow the space of plausible explanations rather than to close it entirely.

Their starting point is a concept Forte has developed over the past decade and a half: the potential past. Rather than seeking a single factual reconstruction, the potential past reframes archaeological theory as a generator of constraints, premises and mechanisms that define what was possible. Digital reconstructions, in this view, are not replications of truth but speculative environments, and theory becomes a way of bounding the range of plausible conditions under which social, cognitive and material phenomena could have emerged. The approach shares affinities with counterfactual history, but extends it by focusing on perception, embodiment and meaning-making rather than only on events and causal sequences.

Technically, the proposal distinguishes four families of AI. Discriminative models classify existing data; generative models produce new outputs such as reconstructions of artifacts; agent-based modeling simulates autonomous entities interacting within an environment; and agentic AI refers to systems given goals that autonomously determine how to pursue them through planning, tool use, memory and adaptive decision-making with minimal human intervention. The authors observe that nearly all AI in archaeology to date, documented in a survey of more than 27,100 publications over a decade, has been discriminative or generative, and case-specific. They state that they know of no existing use of agentic AI in archaeology, which is precisely the gap their architecture targets.

The key innovation lies in where theoretical commitment is encoded. In classical agent-based modeling, researchers hand-code behavioral rules for their agents, drawing on ethnographic analogy or historical records. That practice is interpretable and explicit, but it risks importing researcher bias and excluding behaviors nobody thought to program, a serious limitation when the behaviors themselves are what is under investigation. In an agentic system, the modeler instead specifies goals, perceptual apparatus and environmental constraints, and the decision engine generates the action sequences. Theoretical constructs such as cooperation or ritual authority become goal specifications rather than hand-written rules, and variation across runs arises from the engine’s sensitivity to context rather than from random draws alone.

The authors are careful about what they are and are not claiming. Emergence, they stress, is a property of complex adaptive systems that a model may or may not exhibit, and it remains an empirical question whether a coupled ABM-agentic system produces meaningful emergent patterns. What agentic AI adds is a reconfiguration of the source of agent decision-making: the space of possible actions is no longer exhaustively specified in advance but generated dynamically in response to each agent’s perceptual and internal state. They even argue that the resulting interpretive indeterminacy, where the engine’s reasoning is inspectable but not fully transparent, is a feature rather than a defect, analogous to and in some respects replacing classical stochasticity.

For the decision engine itself, the authors favor small language models, domain-specific language models and hierarchical reasoning models over general-purpose large language models, whose inference costs and contemporary web-derived training data fit poorly with archaeology’s disparate and often non-digital evidence. Their proposed training pipeline has three stages: a general pre-trained base model from families such as Phi, Mistral or Gemma; domain adaptation on excavation reports, archaeological databases such as ARIADNE, tDAR and Open Context, and curated ethnographic records; and task-specific fine-tuning on annotated decision examples, which they identify as the largest methodological unknown. At each decision step, the engine receives a persistent system specification, a rolling context window describing the agent’s perceptual state, and a recent action history, returning a structured action in JSON.

On the simulation side, the authors survey the agent-based modeling ecosystem and conclude that Python-native platforms, particularly Mesa and Repast4Py, are the most immediately viable for coupling, since agentic frameworks such as LangChain, AutoGen and CrewAI are themselves Python-based. NetLogo remains valuable for conceptual modeling but harder to bridge to external AI engines, while GAMA excels where spatial and perceptual modeling is central and MASON suits tightly embedded high-performance applications. Three integration patterns are on offer: call-based coupling, local embedding of a small model in the same process, and batched coupling that trades granularity for throughput. For simulations of dozens to hundreds of agents over simulated decades, the authors judge call-based or batched coupling with a locally hosted small model the most practical starting point.

What might such systems actually simulate? The authors sketch applications ranging from chaîne opératoire analyses of craft traditions to logistics, governance and lived experience. In one illustrative scenario, agents tasked with mobilizing monument construction would generate their own tasks for obtaining materials, constrained by technology and climate, with outcomes ranging from conflict to cooperation across runs; consistency in specific outcomes across many runs would indicate what was likely. A worked example models the Paleolithic cave of Lascaux, where agents move through a torch-lit spatial layout under constraints on vision angle, torch range, innovation rate and copying bias, revealing how perceptual limits and social interaction shape the emergence and concentration of painted motifs. Cognition in these systems is not pre-scripted but enacted through continuous perception-action loops, with internal states such as attention, fatigue and familiarity biasing decisions.

The deeper stakes are epistemological. The authors argue that agentic simulation does not reduce interpretive uncertainty but makes its architecture explicit: theoretical commitments become visible in goals, perceptual inputs and constraints, and running the coupled model under varying specifications becomes the mechanism of analysis. Falsifiability shifts from verifying a single narrative to evaluating which theoretical configurations fail to produce coherent, plausible worlds, with empirical evidence serving as constraint and corrective rather than being replaced. The authors acknowledge open problems, including the cost of engine-mediated decisions at scale and the challenge of validating that generated behavior is plausible rather than merely fluent. But they position archaeology, with its deep engagement with fragmentary evidence and multivocal interpretation, as an ideal testing ground for AI that prioritizes emergence over optimization, and they suggest that simulated pasts that think back may ultimately challenge not just what humans did, but how we know it.

Subject of Research: Integrating agentic AI with agent-based modeling to test archaeological theory through simulated potential pasts

Article Title: Archeological theory and AI: the potential past through agentic systems

Article References: Altaweel, M., & Forte, M. (2026). Archeological theory and AI: the potential past through agentic systems. AI & SOCIETY. https://doi.org/10.1007/s00146-026-03287-0

Image Credits: AI Generated

DOI: 10.1007/s00146-026-03287-0

Keywords: agentic AI, agent-based modeling, archaeology, archaeological theory, simulation, potential past, small language models, hierarchical reasoning models, cyberarchaeology, emergence, Mesa, LLMs