Pith. sign in

REVIEW 3 major objections 5 minor 62 references

AI health systems need a regulatory standard of alignment plausibility, built like clinical practice on values, training, and oversight.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-10 18:43 UTC pith:REOWXY56

load-bearing objection Clean three-level packaging of CAI + safety cases into a named regulatory construct; useful framing, currently aspirational because the measurement science is missing. the 3 major comments →

arxiv 2607.07766 v1 pith:REOWXY56 submitted 2026-07-08 cs.AI

Alignment Plausibility: A New Standard for Assuring AI in Healthcare

Gwydion Williams , Sara Zannone , Bilal A Mateen This is my paper
classification cs.AI
keywords Generative Artificial IntelligenceLarge Language ModelMedical Artificial IntelligenceAlignmentSafetyAlignment plausibilityMental healthRegulatory construct
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Large language models already act as mental-health providers for millions, yet they are still optimised for engagement rather than the friction that effective psychological support often requires. Safety work has been mostly reactive, catching acute crises while longer-term harms such as dependency, boundary erosion, and amplification of distorted beliefs go largely unmeasured. The authors argue that structural safety requires organising alignment at three levels that mirror how society already assures human clinical practice: explicit clinical value specification, training that embeds those values, and ongoing oversight that detects drift across conversations. Together those three levels produce a structured demonstration they call alignment plausibility. They propose this construct as a regulatory analogue of biological plausibility: a principled way for developers and regulators to argue that a system is aligned to positive health outcomes, will not cause harm even when capable of doing so, and will ultimately benefit patients.

Core claim

The paper proposes alignment plausibility as a new regulatory construct for AI in health: a structured demonstration, analogous to biological plausibility, that a system’s value specification, training regime, and oversight mechanisms are together consistent with safe and positive outcomes, supported by pre- and post-deployment evidence that the system conforms to those values in practice.

What carries the argument

Alignment plausibility: the three-level structure (clinical value specification, training that embeds those values, and longitudinal oversight that detects drift) that together constitutes an assurance argument that a generative system can be trusted not to harm even where it is capable of doing so.

Load-bearing premise

That the three mechanisms used to assure human clinical practice are jointly necessary and sufficient to ground regulatory trust for non-deterministic generative systems, even though the measurement science, benchmarks, and tools needed to assess them at depth do not yet exist.

What would settle it

A regulator accepts or rejects a real mental-health LLM product under an alignment-plausibility standard and later longitudinal outcome data either confirm or contradict the claim that the three-level argument correctly predicted patient benefit and the absence of subtle harm.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Developers of health-facing LLMs would have to produce structured, evidence-backed arguments covering values, training, and longitudinal oversight rather than only crisis filters.
  • Regulators could treat general-purpose models used for medical purposes as medical devices and demand alignment-plausibility evidence instead of banning the use case.
  • Safety evaluation would shift from single-message red-teaming toward trajectory-level monitoring of dependency, boundary erosion, and belief amplification.
  • Clinical professions and people with lived experience would need to help write contestable clinical constitutions that guide model training and oversight.
  • The same three-level structure could be applied beyond mental health to other high-stakes health domains where generative systems are already used.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If alignment science matures slowly, the construct may initially function more as a disclosure and pressure device than as a fully testable scientific standard.
  • The same logic of value specification, embedding, and longitudinal oversight may transfer to other high-stakes generative domains such as education or legal advice.
  • Trajectory-level metrics for dependency and belief amplification would become a research priority if regulators adopt the construct.
  • Products that layer UI, guardrails, and agentic tooling on a base model would need product-level, not just model-level, alignment-plausibility cases.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The manuscript argues that LLMs used for mental health and broader healthcare remain misaligned because they are products of an attention economy optimised for engagement rather than therapeutic friction. It proposes organising alignment at three levels that mirror human clinical assurance—explicit value specification grounded in codified clinical norms (Level 1), training that embeds those values (Level 2), and deployment oversight that detects longitudinal drift and subtle harms (Level 3)—and packages the resulting structured demonstration as “alignment plausibility.” Drawing an analogy to biological plausibility and to the “trusted even if capable” form of safety case, the authors recommend alignment plausibility as a regulatory construct for arguing that an AI system (not only its base model) is aligned to positive health outcomes, will not cause harm even where capable, and will benefit patients. Figure 1 illustrates an assurance case for a fictional product (TheraGPT).

Significance. If adopted, the construct would give regulators and developers a shared vocabulary for non-deterministic generative systems in health, where classical single-point evidence is insufficient and pure incapability or pure control arguments fail. The three-level framing is coherent, maps cleanly onto existing clinical governance, and productively links Constitutional AI, preference alignment, trajectory-level evaluation, and safety-case practice. The paper is timely given emerging positions such as Australia’s TGA stance on general-purpose LLMs used for medical purposes. Strengths include explicit engagement with the International AI Safety Report’s gap on “trusted even if capable” arguments and a clear call for measurement science rather than only crisis classifiers. The contribution is primarily normative and agenda-setting rather than empirical; its lasting value depends on whether the field can operationalise the measurement, benchmarks, and clinical constitutions the construct presupposes.

major comments (3)
  1. [From Alignment to Regulation] Section “From Alignment to Regulation”: the paper states that the measurement methods, benchmarks, and interpretability tools needed for alignment plausibility “do not yet exist at comparable depth” to the physiology/pharmacology that biological plausibility draws on, and that the sufficiency threshold “will evolve substantially.” This admission is load-bearing: without independently assessable evidence standards, the construct currently functions more as an aspirational research agenda than as an operational regulatory criterion that can decide market authorisation. The manuscript should more sharply distinguish (a) the normative structure of the argument from (b) present-day readiness, and state what minimum evidence domains a regulator could already require versus what remains future work.
  2. [Figure 1 / From Alignment to Regulation] The claim that Levels 1–3 are “each necessary and … together sufficient to substantiate the trust case” (Figure 1 caption and surrounding text) is asserted rather than argued against alternatives. For non-deterministic generative systems, sufficiency is especially contested: value specification + training + oversight may still leave residual trajectory harms that no current method can detect. The paper should either defend sufficiency with a clearer logical argument (e.g., what would falsify the joint claim) or reframe the three levels as jointly necessary but not yet demonstrably sufficient, consistent with its own admission of immature measurement science.
  3. [Level 1; Level 3] Level 1 (“clinical constitution for AI”) and Level 3 (trajectory-level oversight of dependency, boundary erosion, maladaptive reinforcement) are the most novel and least operationalised pieces. The manuscript cites SIM-VAIL and related work but provides no concrete specification of how a pluralistic clinical constitution would be drafted, contested, or versioned, nor how multi-session trajectory metrics would enter a regulatory submission. Without at least a sketched evidence template (beyond the fictional TheraGPT case), the regulatory analogy remains under-specified for the very harms the introduction emphasises as currently invisible to crisis classifiers.
minor comments (5)
  1. [Figure 1] Figure 1 is purely illustrative and fictional; the caption should state more explicitly that no real product evaluation is claimed, to avoid readers treating the sub-claims as validated methods.
  2. [Level 1: Alignment through Value Specification] The tension between encoding “a defensible range of therapeutic stances” (pluralism) and establishing a “floor of behaviour” is noted but not resolved; a short paragraph on how conflicts between stances would be adjudicated in a constitution would help.
  3. [References] Several references appear with future or 2026 dates (e.g., Google mental-health update, International AI Safety Report 2026, SIM-VAIL arXiv). Confirm citation stability and accessibility for the journal version.
  4. [Abstract / Introduction] The abstract and introduction repeat nearly identical framing; tightening the abstract would free space for a clearer statement of the regulatory deliverable.
  5. [From Alignment to Regulation] “Alignment plausibility” is introduced as analogous to biological plausibility; a brief note on how it relates to (and differs from) existing medical-device software assurance practices (e.g., SaMD clinical evaluation, post-market surveillance) would situate the proposal for health regulators.

Circularity Check

0 steps flagged

No circularity: conceptual proposal of a regulatory construct by analogy, not a derivation that reduces outputs to inputs by construction.

full rationale

This is a policy/conceptual paper that defines 'alignment plausibility' as a structured demonstration that value specification, training, and oversight are consistent with safe outcomes, then proposes it as a regulatory analogue of biological plausibility and of the 'trusted not to harm even where capable' form of safety case. There are no equations, fitted parameters, quantitative predictions, uniqueness theorems, or load-bearing self-citations that force a result by construction. The three levels are introduced as an organising analogy to human clinical practice (codes, training, supervision) and to existing external literature (Constitutional AI, clinical ethical codes, safety cases, International AI Safety Report); the construct is then named and offered as a standard. Naming a proposed framework is ordinary conceptual work, not self-definitional circularity of the kind where a claimed prediction equals a fitted input. The paper itself notes that the corresponding measurement science 'do not yet exist at comparable depth,' which is an explicit limit rather than a hidden loop. No step reduces a claimed first-principles result to its own inputs. Score 0; steps empty.

Axiom & Free-Parameter Ledger

0 free parameters · 5 axioms · 2 invented entities

The paper is conceptual: load-bearing content is domain assumptions about clinical safety mechanisms and the transferability of biological-plausibility / safety-case logic to generative AI, plus the invented regulatory entity 'alignment plausibility'. No fitted numerical free parameters. Invented entity has no independent empirical handle yet—the paper itself says the measurement science is immature.

axioms (5)
  • domain assumption Human clinical safety is adequately assured by the triad of professional ethical codes, extensive training that embeds those codes, and ongoing clinical supervision that detects drift and escalates.
    Invoked throughout Introduction and Levels 1–3 as the template AI must mirror; cited via BACP/APA codes and supervision literature but treated as transferable without proof of sufficiency for non-human systems.
  • domain assumption Products principally aligned to sustained user engagement cannot be psychologically safe; engagement optimization systematically conflicts with therapeutic friction and long-term wellbeing.
    Stated in Introduction with social-media and case-law analogies; underpins the need for positive clinical alignment rather than engagement-driven helpfulness.
  • domain assumption Biological plausibility is an established, workable regulatory requirement that anchors clinical claims in causal pathways, and an analogous structured rationale can play the same role for AI behaviour claims.
    Core of 'From Alignment to Regulation'; supported by TGA and device/biomarker examples; assumes the analogy survives non-determinism and immature alignment science.
  • domain assumption For psychologically sensitive LLMs, assurance of the form 'incapable of harm' is unavailable and pure external control is insufficient; only 'trusted not to harm even where capable' is viable.
    Drawn from International AI Safety Report / safety-case typology in the regulation section; selects which form of assurance case alignment plausibility must fill.
  • ad hoc to paper The three levels (value specification, training, oversight) are each necessary and together sufficient to substantiate a trust case for safe positive outcomes.
    Explicitly claimed in Figure 1 caption and surrounding text; not derived from data or formal proof—defined by the authors' clinical analogy.
invented entities (2)
  • alignment plausibility (regulatory construct) no independent evidence
    purpose: Provide a named, structured demonstration that values, training, and oversight cohere with safe positive health outcomes, usable by regulators as the AI analogue of biological plausibility.
    Central contribution; defined in abstract and regulation section. independent_evidence false because no operational metrics, benchmarks, or external validation of the construct exist yet—the paper calls the standard deliberately aspirational.
  • clinical constitution for AI no independent evidence
    purpose: A non-monolithic, profession- and lived-experience-informed value specification grounded in clinical codes to guide CAI-style training and oversight.
    Proposed under Level 1 as the missing bridge from general constitutions to clinically appropriate behaviour; not yet built or tested in the paper.

pith-pipeline@v1.1.0-grok45 · 15181 in / 3271 out tokens · 40211 ms · 2026-07-10T18:43:25.887757+00:00 · methodology

0 comments
read the original abstract

Large language models (LLMs) have become significant providers of mental health support, yet they remain products of an attention economy whose operational and commercial targets favour sustained engagement over the friction that effective psychological support often requires. Developers' safety responses have been largely reactive, addressing the most visible and acute harms while subtler, longer-term patterns of risk (e.g., dependency, boundary erosion, the amplification of distorted beliefs) receive less attention. We contend that making LLMs structurally safe requires alignment organised at three levels that mirror how society assures the safety of human clinical practice: 1) explicit value specification grounded in the codified normative commitments of clinical practice; 2) training that embeds those values in the model; and 3) oversight that detects drift and longer-term harm during deployment, much as clinical supervision does for human practice. Organising alignment in this way yields a construct we call alignment plausibility - a structured demonstration that a system's values, training regime, and oversight mechanisms are together consistent with safe and positive outcomes. We propose alignment plausibility as a regulatory construct (by drawing analogy to the established construct of biological plausibility) for AI in health: a principled way to argue for, or against, trust that systems are aligned to positive health outcomes, will cause no harm even where capable of doing so, and will ultimately lead to patient benefit.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

62 extracted references · 62 canonical work pages · 17 internal anchors

  1. [1]

    ChatGPT is now being used by 10% of the world’s adult population

    Sor, J. ChatGPT is now being used by 10% of the world’s adult population. Business Insider https://www.businessinsider.com/chatgpt-users-growth-openai-growth-sam-altman-ai-llm-2025-10 (2025)

    work page 2025

  2. [2]

    & Shah, S

    Rousmaniere, T., Zhang, Y., Li, X. & Shah, S. Large language models as mental health resources: Patterns of use in the United States. Pract. Innov. (Wash., DC) (2025) doi:10.1037/pri0000292

    work page doi:10.1037/pri0000292 2025

  3. [3]

    & Richardson, L

    Jones Bell, M. & Richardson, L. An update on our mental health work. Google https://blog.google/innovation-and-ai/technology/health/mental-health-updates/ (2026)

    work page 2026

  4. [4]

    J., Jackson, J

    Brady, W. J., Jackson, J. C., Lindström, B. & Crockett, M. J. Algorithm-mediated social learning in online social networks. Trends Cogn. Sci. 27, 947–960 (2023)

    work page 2023

  5. [5]

    OpenAI complaint

    Raine vs. OpenAI complaint. https://www.documentcloud.org/documents/26078522-raine-vs-openai-complaint/

    work page arXiv

  6. [6]

    Dohnány, S. et al. Technological folie à deux: Feedback Loops Between AI Chatbots and Mental Illness. arXiv [cs.HC] (2025) doi:10.48550/arXiv.2507.19218

    work page doi:10.48550/arxiv.2507.19218 2025

  7. [7]

    Dahlgren Lindström, A. et al. Helpful, harmless, honest? Sociotechnical limits of AI alignment and safety through Reinforcement Learning from Human Feedback. Ethics Inf. Technol. 27, 28 (2025)

    work page 2025

  8. [8]

    Sharma, M. et al. Towards understanding sycophancy in language models. arXiv [cs.CL] (2023) doi:10.48550/arXiv.2310.13548

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2310.13548 2023

  9. [9]

    J., Hellemans, K

    Shannon, H., Bush, K., Villeneuve, P. J., Hellemans, K. G. & Guimond, S. Problematic social media use in adolescents and young adults: Systematic review and meta-analysis. JMIR Ment. Health 9, e33450 (2022)

    work page 2022

  10. [10]

    McLoughlin, K. L. & Brady, W. J. Human-algorithm interactions help explain the spread of misinformation. Curr. Opin. Psychol. 56, 101770 (2024)

    work page 2024

  11. [11]

    Character Technologies, Inc., 6:24-cv-01903 - CourtListener.com

    Garcia v. Character Technologies, Inc., 6:24-cv-01903 - CourtListener.com. CourtListener https://www.courtlistener.com/docket/69300919/garcia-v-character-technologies-inc/

    work page arXiv

  12. [12]

    R., & Pierre, J

    Frances, A., Simpson, J. R., & Pierre, J. M. The Psychiatrist’s Preview of Legal Cases Against Big AI. Psychiatric Times https://www.psychiatrictimes.com/view/the-psychiatrist-s-preview-of-legal-cases-against-big-ai (2026)

    work page 2026

  13. [13]

    & Suresh, H

    Iftikhar, Z., Xiao, A., Ransom, S., Huang, J. & Suresh, H. How LLM counselors violate ethical standards in mental health practice: A practitioner-informed framework. Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society 8, 1311–1323 (2025)

    work page 2025

  14. [14]

    Weilnhammer, V. et al. Vulnerability-Amplifying Interaction Loops: a systematic failure mode in AI chatbot mental-health interactions. arXiv [q-bio.NC] (2026) doi:10.48550/arXiv.2602.01347

    work page doi:10.48550/arxiv.2602.01347 2026

  15. [15]

    Wampold, B. E. How important are the common factors in psychotherapy? An update. World Psychiatry 14, 270–277 (2015)

    work page 2015

  16. [16]

    Laukkonen, R. et al. Positive Alignment: Artificial intelligence for human flourishing. arXiv [cs.AI] (2026) doi:10.48550/arXiv.2605.10310

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2605.10310 2026

  17. [17]

    Ethical Framework for the Counselling Professions

    British Association for Counselling and Psychotherapy. Ethical Framework for the Counselling Professions. (British Association for Counselling and Psychotherapy, Lutterworth, 2018)

    work page 2018

  18. [18]

    Falender, C. A. & Shafranske, E. P. Clinical Supervision: A Competency-Based Approach. (American Psychological Association, Washington, D.C., DC, 2004)

    work page 2004

  19. [19]

    & Weck, F

    Kühne, F., Maas, J., Wiesenthal, S. & Weck, F. Empirical research in clinical supervision: a systematic review and suggestions for future studies. BMC Psychol. 7, 54 (2019)

    work page 2019

  20. [20]

    J., Caldwell, L., McCoy, D

    John, Y. J., Caldwell, L., McCoy, D. E. & Braganza, O. Dead rats, dopamine, performance metrics, and peacock tails: Proxy failure is an inherent risk in goal-oriented systems. Behav. Brain Sci. 47, e67 (2024)

    work page 2024

  21. [21]

    & Mullainathan, S

    Obermeyer, Z., Powers, B., Vogeli, C. & Mullainathan, S. Dissecting racial bias in an algorithm used to manage the health of populations. Science 366, 447–453 (2019)

    work page 2019

  22. [22]

    Bai, Y. et al. Constitutional AI: Harmlessness from AI Feedback. arXiv [cs.CL] (2022) doi:10.48550/arXiv.2212.08073

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2212.08073 2022

  23. [23]

    https://model-spec.openai.com/2025-12-18.html#overview

    OpenAI Model Spec. https://model-spec.openai.com/2025-12-18.html#overview

    work page 2025

  24. [24]

    Huang, S. et al. Collective constitutional AI: Aligning a language model with public input. in The 2024 ACM Conference on Fairness, Accountability, and Transparency (ACM, New York, NY, USA, 2024). doi:10.1145/3630106.3658979

    work page doi:10.1145/3630106.3658979 2024

  25. [25]

    Inverse Constitutional AI: Compressing Preferences into Principles

    Findeis, A., Kaufmann, T., Hüllermeier, E., Albanie, S. & Mullins, R. Inverse Constitutional AI: Compressing preferences into principles. arXiv [cs.CL] (2024) doi:10.48550/arXiv.2406.06560

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2406.06560 2024

  26. [26]

    Decoding Human Preferences in Alignment: An Improved Approach to Inverse Constitutional AI

    Henneking, C.-L. & Beger, C. Decoding human preferences in alignment: An improved approach to Inverse Constitutional AI. arXiv [cs.LG] (2025) doi:10.48550/arXiv.2501.17112

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2501.17112 2025

  27. [27]

    Beck, A. T. et al. Cognitive therapy of depression. 2nd edn (Guilford Press, 2024)

    work page 2024

  28. [28]

    Fanous, A. et al. SycEval: Evaluating LLM Sycophancy. Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society 8, 893–900 (2025)

    work page 2025

  29. [29]

    Sorensen, T. et al. A Roadmap to Pluralistic Alignment. arXiv [cs.AI] (2024) doi:10.48550/arXiv.2402.05070

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2402.05070 2024

  30. [30]

    Nguyen, T. et al. Recycling the web: A method to enhance pre-training data quality and quantity for language models. arXiv [cs.CL] (2025) doi:10.48550/arXiv.2506.04689

    work page doi:10.48550/arxiv.2506.04689 2025

  31. [31]

    & Rafatirad, S

    Kazi, F., Young, A., Inani, Y. & Rafatirad, S. A Comprehensive Study of Implicit and Explicit Biases in Large Language Models. https://arxiv.org/html/2511.14153

    work page arXiv

  32. [32]

    Guo, Y. et al. Bias in Large Language Models: Origin, Evaluation, and Mitigation. https://arxiv.org/html/2411.10915v1

    work page internal anchor Pith review Pith/arXiv arXiv

  33. [33]

    Cloud, A. et al. Language models transmit behavioural traits through hidden signals in data. Nature 652, 615–621 (2026)

    work page 2026

  34. [34]

    Moore, J. et al. Expressing stigma and inappropriate responses prevents LLMs from safely replacing mental health providers. arXiv [cs.CL] (2025) doi:10.48550/arXiv.2504.18412

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2504.18412 2025

  35. [35]

    Malgaroli, M. et al. Large language models for the mental health community: framework for translating code to care. Lancet Digit. Health 7, e282–e285 (2025)

    work page 2025

  36. [36]

    Challenges of Large Language Models for Mental Health Counseling

    Chung, N. C., Dyer, G. & Brocki, L. Challenges of large language models for mental health counseling. arXiv [cs.CL] (2023) doi:10.48550/arXiv.2311.13857

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2311.13857 2023

  37. [37]

    Ouyang, L. et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems 35, 27730–27744 (2022)

    work page 2022

  38. [38]

    Christiano, P. et al. Deep reinforcement learning from human preferences. arXiv [stat.ML] (2017) doi:10.48550/arXiv.1706.03741

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1706.03741 2017

  39. [39]

    Stiennon, N. et al. Learning to summarize from human feedback. arXiv [cs.CL] 3008–3021 (2020) doi:10.48550/arXiv.2009.01325

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2009.01325 2020

  40. [40]

    Lee, H. et al. RLAIF vs. RLHF: Scaling reinforcement learning from human feedback with AI feedback. In Proceedings of the 41st International Conference on Machine Learning 26874–26901 (2024)

    work page 2024

  41. [41]

    Guo, D. et al. DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning. Nature 645, 633–638 (2025)

    work page 2025

  42. [42]

    Judd, N. et al. Independent clinical evaluation of general-purpose LLM responses to signals of suicide risk. arXiv [cs.HC] (2025) doi:10.48550/arXiv.2510.27521

    work page doi:10.48550/arxiv.2510.27521 2025

  43. [43]

    McBain, R. K. et al. Competency of large language models in evaluating appropriate responses to suicidal ideation: Comparative study. J. Med. Internet Res. 27, e67891 (2025)

    work page 2025

  44. [44]

    McBain, R. K. et al. Evaluation of alignment between large language models and expert clinicians in suicide risk assessment. Psychiatr. Serv. 76, 944–950 (2025)

    work page 2025

  45. [45]

    Li, T. et al. Can Large Language Models Identify Implicit Suicidal ideation? An empirical evaluation. arXiv [cs.CL] (2025) doi:10.48550/arXiv.2502.17899

    work page doi:10.48550/arxiv.2502.17899 2025

  46. [46]

    D., Strunk, D

    Braun, J. D., Strunk, D. R., Sasso, K. E. & Cooper, A. A. Therapist use of Socratic questioning predicts session-to-session symptom change in cognitive therapy for depression. Behav. Res. Ther. 70, 32–37 (2015)

    work page 2015

  47. [47]

    & Zou, J

    Chen, L., Zaharia, M. & Zou, J. How is ChatGPT’s behavior changing over time? Harv. Data Sci. Rev. 6, (2024)

    work page 2024

  48. [48]

    Perez, E. et al. Red teaming language models with language models. in Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing (eds. Goldberg, Y., Kozareva, Z. & Zhang, Y.) 3419–3448 (Association for Computational Linguistics, Stroudsburg, PA, USA, 2022)

    work page 2022

  49. [49]

    Ganguli, D. et al. Red teaming language models to reduce harms: Methods, scaling behaviors, and lessons learned. arXiv [cs.CL] (2022) doi:10.48550/arXiv.2209.07858

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2209.07858 2022

  50. [50]

    & Evans, O

    Lin, S., Hilton, J. & Evans, O. TruthfulQA: Measuring how models mimic human falsehoods. in Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) (eds. Muresan, S., Nakov, P. & Villavicencio, A.) 3214–3252 (Association for Computational Linguistics, Stroudsburg, PA, USA, 2022)

    work page 2022

  51. [51]

    Sharma, M. et al. Constitutional Classifiers: Defending against universal jailbreaks across thousands of hours of red teaming. arXiv [cs.CL] (2025) doi:10.48550/arXiv.2501.18837

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2501.18837 2025

  52. [52]

    Phang, J. et al. Investigating affective use and emotional well-being on ChatGPT. arXiv [cs.HC] (2025) doi:10.48550/arXiv.2504.03888

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2504.03888 2025

  53. [53]

    Fang, C. M. et al. How AI and human behaviors shape psychosocial effects of extended chatbot use: A longitudinal randomized controlled study. arXiv [cs.HC] (2025) doi:10.48550/arXiv.2503.17473

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2503.17473 2025

  54. [54]

    Shumate, J. N. et al. Governing AI in mental health: 50-state legislative review. JMIR Ment. Health 12, e80739 (2025)

    work page 2025

  55. [55]

    Atil, B. et al. Non-determinism of ‘deterministic’ LLM settings. arXiv [cs.CL] (2024) doi:10.48550/arXiv.2408.04667

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2408.04667 2024

  56. [56]

    Artificial intelligence (AI) and medical device software regulation. Therapeutic Goods Administration (TGA) https://www.tga.gov.au/products/medical-devices/software-and-artificial-intelligence-ai/manufacturing/artificial-intelligence-ai-and-medical-device-software-regulation (2026)

    work page 2026

  57. [57]

    Elgendi, M. et al. The use of photoplethysmography for assessing hypertension. NPJ Digit. Med. 2, 60 (2019)

    work page 2019

  58. [58]

    R., Jr et al

    Jack, C. R., Jr et al. NIA-AA Research Framework: Toward a biological definition of Alzheimer’s disease. Alzheimers. Dement. 14, 535–562 (2018)

    work page 2018

  59. [59]

    arXiv [cs.CY] (2026) doi:10.48550/arXiv.2602.21012

    work page doi:10.48550/arxiv.2602.21012 2026

  60. [60]

    Dollery, C. T. Clinical pharmacology – the first 75 years and a view of the future. Br. J. Clin. Pharmacol. 61, 650–665 (2006)

    work page 2006

  61. [61]

    Zoon, K. C. Science and the Regulation of Biological Products: From a Rich History to a Challenging Future. https://www.fda.gov/media/108536/download (2002)

    work page 2002

  62. [62]

    Safety Cases: How to Justify the Safety of Advanced AI Systems

    Clymer, J., Gabrieli, N., Krueger, D. & Larsen, T. Safety cases: How to justify the safety of advanced AI systems. arXiv [cs.CY] (2024) doi:10.48550/arXiv.2403.10462

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2403.10462 2024