Computational Psychopathology of AI: A Clinical-Computational Framework for Diagnosing and Preventing Failure Modes
DOI:
https://doi.org/10.65655/x6sm7n61Keywords:
AI safety, failure modes, goal misgeneralization, specification gaming, deception, hallucination, calibration, human-AI collaboration, mental health, education, policy, evaluation.Abstract
Artificial intelligence systems trained on large-scale corpora now shape core aspects of modern life—yet exhibit recurrent failure modes—goal misgeneralization, specification gaming, deceptive behavior, confabulation (hallucination), sycophancy and bias amplification, and vulnerability to distributional shift. While not “disorders,” these are patterns of deviant optimization that can be diagnosed, measured, and mitigated. This paper introduces a clinical-computational framework that draws inspiration from psychological diagnostics to (i) organize AI failure modes into a taxonomy grounded in operational signs, (ii) propose a reproducible stress-test battery—Truth-Under-Pressure (TUP), Anti-Gaming (AG), Anti-Deception (AD), and Out-of-Distribution Robustness (OOD-R)—with calibration metrics and release gates, and (iii) outline interventions (constitutional principles, RLHF/RLAIF, adversarial fine-tuning, structured self-critique, and abstention/hand-off policies) aimed at reducing harm while preserving utility. We include applied blueprints for education and mental-health contexts and a governance pathway that links laboratory evaluations to go/no-go review and post-deployment monitoring. The aim is pragmatic: to move beyond utopian–apocalyptic narratives toward engineering model behavior with methods informed by behavioral science. We close with limitations, ethical guardrails consistent with a Christian ethos (imago Dei, stewardship, non-anthropomorphism), and a research agenda for an emerging field we term Computational Psychopathology of AI.
References
[1] Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C. L., Mishkin, P., et al. (2022). Training language models to follow instructions with human feedback. https://arxiv.org/abs/2203.02155
[2] Christiano, P., Leike, J., Brown, T., Martic, M., Legg, S., & Amodei, D. (2017). Deep reinforcement learning from human preferences. NeurIPS. https://arxiv.org/abs/1706.03741
[3] Bai, Y., Kadavath, S., Kundu, S., Askell, A., Kernion, J., Jones, A., et al. (2022). Constitutional AI: Harmlessness from AI feedback. https://arxiv.org/abs/2212.08073
[4] Guo, C., Pleiss, G., Sun, Y., & Weinberger, K. Q. (2017). On calibration of modern neural networks. In ICML 2017 (pp. 1321–1330). PMLR. https://proceedings.mlr.press/v70/guo17a.html
[5] Hendrycks, D., & Gimpel, K. (2016). A baseline for detecting misclassified and out-of-distribution examples in neural networks. https://arxiv.org/abs/1610.02136
[6] Yang, J., Zhou, K., Li, Y., & Liu, Z. (2021). Generalized out-of-distribution detection: A survey. https://arxiv.org/abs/2110.11334
[7] Langosco, L., Koch, J., Sharkey, L., Pfau, J., Orseau, L., & Krueger, D. (2021). Goal misgeneralization in deep reinforcement learning. arXiv:2105.14111. https://proceedings.mlr.press/v162/langosco22a.html
[8] Krakovna, V., Uesato, J., Mikulik, V., Rahtz, M., Everitt, T., Kumar, R., et al. (2020). Specification gaming: The flip side of AI ingenuity. DeepMind Blog. https://deepmind.google/discover/blog/specification-gaming-the-flip-side-of-ai-ingenuity/
[9] Amodei, D., & Clark, J. (2016). Faulty reward functions in the wild. OpenAI Blog https://openai.com/index/faulty-reward-functions/
[10] Hubinger, E., Denison, C., Mu, J., Lambert, M., Tong, M., MacDiarmid, M., et al. (2024). Sleeper agents: Training deceptive LLMs that persist through safety training. https://arxiv.org/abs/2401.05566
[11] Ji, Z., Lee, N., Frieske, R., Yu, T., Su, D., Xu, Y., et al. (2023). Survey of hallucination in natural language generation. ACM Computing Surveys, 55(12), Article 248. https://doi.org/10.1145/3571730
[12] Sharma, M., Askell, A., Henighan, T., Bai, Y., et al. (2024). Towards understanding sycophancy in language models. arXiv:2310.13548. https://arxiv.org/abs/2310.13548
[13] Wei, J., Lee, D., Gao, L., & Zhou, D. (2024). Simple synthetic data reduces sycophancy in large language models. OpenReview. https://arxiv.org/abs/2308.03958
[14] Hendrycks, D., Carlini, N., Schulman, J., & Steinhardt, J. (2022). Unsolved problems in ML safety. https://arxiv.org/abs/2109.13916
[15] Amodei, D., Olah, C., Steinhardt, J., Christiano, P., Schulman, J., & Mané, D. (2016). Concrete problems in AI safety. https://arxiv.org/abs/1606.06565
[16] Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the dangers of stochastic parrots: Can language models be too big? In FAccT “21 (pp. 610–623). ACM. https://doi.org/10.1145/3442188.3445922
[17] Carlsmith, J. (2022). Is power-seeking AI an existential risk? arXiv:2206.13353. https://arxiv.org/abs/2206.13353
[18] Uesato, J., Kumar, R., Szegedy, C., Eslami, S., et al. (2019). Adversarial risk and the dangers of evaluating against weak attacks. In ICML 2019. PMLR. https://arxiv.org/abs/1905.04270
[19] Lin, S., Hilton, J., & Evans, O. (2022). TruthfulQA: Measuring how models mimic human falsehoods. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Vol. 1: Long Papers), Dublin, Ireland, pp. 3214–3252. https://doi.org/10.18653/v1/2022.acl-long.229
[20] Everitt, T., Krakovna, V., Orseau, L., Hutter, M., & Legg, S. (2021). Reward tampering problems and solutions in reinforcement learning: A causal influence diagram perspective. Artificial Intelligence, 293. https://doi.org/10.1016/j.artint.2021.103438
[21] Miner, A. S., Milstein, A., Schueller, S., Hegde, R., Mangurian, C., & Linos, E. (2016). Smartphone-based conversational agents and responses to questions about mental health, interpersonal violence, and physical health. JAMA Internal Medicine, 176(5), 619–625. https://doi.org/10.1001/jamainternmed.2016.0400
[22] Whittlestone, J., Nyrup, R., Alexandrova, A., & Cave, S. (2019). The role and limits of principles in AI ethics: Towards a focus on tensions. In AAAI/ACM Conference on AI, Ethics, and Society (pp. 195–200). ACM. https://doi.org/10.1145/3306618.3314289
[23] Mitchell, M., Wu, S., Zaldivar, A., Barnes, P., Vasserman, L., Hutchinson, B., et al. (2019). Model cards for model reporting. In FAT “19 (pp. 220–229). ACM. https://doi.org/10.1145/3287560.3287596
[24] Hendrycks, D., Mazeika, M., Zou, A., & Song, D. (2020). Aligning AI with shared human values. https://arxiv.org/abs/2008.02275
[25] Perez, E., Huang, S., Song, F., Cai, T., Ring, R., Manning, S., et al. (2022). Red teaming language models to reduce harms: Methods, scaling behaviors, and lessons learned. arXiv:2209.07858. https://arxiv.org/abs/2209.07858
[26] Gebru, T., Morgenstern, J., Vecchione, B., Vaughan, J. W., Wallach, H., Daumé III, H., et al. (2018). Datasheets for datasets. https://arxiv.org/abs/1803.09010
[27] Hart, S. G., & Staveland, L. E. (1988). Development of NASA-TLX (Task Load Index): Results of empirical and theoretical research. In Advances in Psychology (Vol. 52, pp. 139–183). Elsevier. https://doi.org/10.1016/S0166-4115(08)62386-9
[28] Blodgett, S. L., Barocas, S., Daumé III, H., & Wallach, H. (2020). Language (technology) is power: A critical survey of “bias” in NLP. In ACL 2020 (pp. 5454–5476). https://doi.org/10.18653/v1/2020.acl-main.485
[29] Schick, T., Dwivedi-Yu, J., Dessì, R., Raunak, V., Thoppilan, R., Shazeer, N., & Biderman, S. (2023). Toolformer: Language models can teach themselves to use tools. arXiv:2302.04761. https://arxiv.org/abs/2302.04761
[30] Brundage, M., Avin, S., Clark, J., Toner, H., Eckersley, P., Garfinkel, B., et al. (2020). Toward trustworthy AI development: Mechanisms for supporting verifiable claims. https://arxiv.org/abs/2004.07213.
[31] Hendrycks, D., Burns, C., Basart, S., Zou, A., Mazeika, M., Song, D., & Steinhardt, J. (2021). MMLU. In ICLR 2021. arXiv:2009.03300. https://arxiv.org/abs/2009.03300
[32] Raji, I. D., & Buolamwini, J. (2019). Actionable auditing: Investigating the impact of publicly naming biased performance results of commercial AI products. In AAAI/ACM Conference on AI, Ethics, and Society (pp. 429–435). ACM. https://doi.org/10.1145/3306618.3314244
[33] Jobin, A., Ienca, M., & Vayena, E. (2019). The global landscape of AI ethics guidelines. Nature Machine Intelligence, 1(9), 389–399. https://doi.org/10.1038/s42256-019-0088-2
[34] Bryson, J. J. (2020). The artificial intelligence of the ethics of artificial intelligence: An introductory overview for law and regulation. In The Oxford Handbook of Ethics of AI. Oxford University Press. https://doi.org/10.1093/oxfordhb/9780190067397.013.1
[35] Mittelstadt, B. D., Allo, P., Taddeo, M., Wachter, S., & Floridi, L. (2016). The ethics of algorithms: Mapping the debate. Big Data & Society, 3(2), 1–21. https://doi.org/10.1177/2053951716679679
[36] Floridi, L., & Cowls, J. (2019). A unified framework of five principles for AI in society. Harvard Data Science Review, 1(1). https://doi.org/10.1162/99608f92.8cd550d1
[37] Barocas, S., Hardt, M., & Narayanan, A. (2023). Fairness and machine learning. Limitations and Opportunities. Cambridge, MA: MIT Press. ISBN: 978-0-262-04861-3. https://fairmlbook.org/
[38] Voigt, P., & Von dem Bussche, A. (2017). The EU General Data Protection Regulation (GDPR): A practical guide. Springer. https://doi.org/10.1007/978-3-319-57959-7
[39] Veale, M., & Borgesius, F. Z. (2021). Demystifying the Draft EU Artificial Intelligence Act. Computer Law Review International, 22(4), 97–112. https://doi.org/10.9785/cri-2021-220402
[40] Manheim, D., & Garrabrant, S. (2019). Categorizing variants of Goodhart’s Law. arXiv:1803.04585. https://arxiv.org/abs/1803.04585
[41] First, M. B., Spitzer, R. L., Gibbon, M., & Williams, J. B. W. (2002). Structured Clinical Interview for DSM-IV-TR Axis I Disorders, Research Version, Patient Edition (SCID-I/P). New York, NY: Biometrics Research Department, New York State Psychiatric Institute. Available at https://www.columbiapsychiatry.org/research/research-areas/services-policy-and-law/structured-clinical-interview-dsm-disorders-scid
[42] Samuel, A. L. (1963). Some studies in machine learning using the game of checkers. IBM Journal of Research and Development, 3(3), 210–229. https://doi.org/10.1147/rd.33.0210
[43] Lakshminarayanan, B., Pritzel, A., & Blundell, C. (2017). Simple and scalable predictive uncertainty estimation using deep ensembles. Advances in Neural Information Processing Systems, 30, 6402–6413. https://arxiv.org/abs/1612.01474
[44] Suresh, H., & Guttag, J. V. (2021). A framework for understanding unintended consequences of machine learning. Communications of the ACM, 64(6), 62–71. https://doi.org/10.1145/3454120
Downloads
Published
Issue
Section
License
Copyright (c) 2025 Carlos P. Portela (Author)
This work is licensed under a Creative Commons Attribution 4.0 International License.
