Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety

Korbak, T., et al. (2025).

arXiv:2507.11473

Abstract

AI systems that “think” in human language offer a unique opportunity for AI safety: we can monitor their chains of thought (CoT) for the intent to misbehave. Like all other known AI oversight methods, CoT monitoring is imperfect and allows some misbehavior to go unnoticed. Nevertheless, it shows promise and we recommend further research into CoT monitorability and investment in CoT monitoring alongside existing safety methods.  Because CoT monitorability may be fragile, we recommend that frontier model developers consider the impact of development decisions on CoT monitorability.

A pdf is here.

Here are some thoughts:

The paper highlights a unique moment in AI development, where large language models reason in human language, making their decisions interpretable through visible “chain of thought” (CoT) processes. This human-readable reasoning enables researchers to audit, monitor, and potentially catch misaligned or risky behaviors by reviewing the model's intermediary steps rather than just its final outputs.

While CoT monitoring presents new possibilities for AI oversight and transparency, the paper emphasizes its fragility: monitorability can decrease if model training shifts toward less interpretable methods or if models become incentivized to obscure their thoughts. The authors caution that CoT traces may not always faithfully represent internal reasoning and that models might find ways to hide misbehavior regardless. They call for further research into how much trust can be placed in CoT monitoring, the development of benchmarks for faithfulness and transparency, and architectural choices that preserve monitorability.

Ultimately, the paper urges AI developers to treat CoT monitorability as a valuable but unstable safety layer, advocating for its inclusion alongside—but not in place of—other oversight and alignment strategies

You sound like ChatGPT

Sara Parker

The Verge

Originally posted 20 June 25

Here is an excerpt:

AI shows up most obviously in functions like smart replies, autocorrect, and spellcheck. Research out of Cornell looks at our use of smart replies in chats, finding that use of smart replies increases overall cooperation and feelings of closeness between participants, since users end up selecting more positive emotional language. But if people believed their partner was using AI in the interaction, they rated their partner as less collaborative and more demanding. Crucially, it wasn’t actual AI usage that turned them off — it was the suspicion of it. We form perceptions based on language cues, and it’s really the language properties that drive those impressions, says Malte Jung, Associate Professor of Information Science at Cornell University and a co-author of the study.

This paradox — AI improving communication while fostering suspicion — points to a deeper loss of trust, according to Mor Naaman, professor of Information Science at Cornell Tech. He has identified three levels of human signals that we’ve lost in adopting AI into our communication. The first level is that of basic humanity signals, cues that speak to our authenticity as a human being like moments of vulnerability or personal rituals, which say to others, “This is me, I’m human.” The second level consists of attention and effort signals that prove “I cared enough to write this myself.” And the third level is ability signals which show our sense of humor, our competence, and our real selves to others. It’s the difference between texting someone, “I’m sorry you’re upset” versus “Hey sorry I freaked at dinner, I probably shouldn’t have skipped therapy this week.” One sounds flat; the other sounds human.

The info is here.

Here are some thoughts:

The increasing influence of AI language models like ChatGPT on everyday language, as highlighted in the article, holds significant implications for practicing psychologists. As these models shape linguistic trends—boosting the use of certain words and phrases—patients may unconsciously adopt these patterns in therapy sessions. This shift could reflect broader cultural changes in communication, potentially affecting how individuals articulate emotions, experiences, and personal narratives. Psychologists must remain attuned to these developments, as AI-mediated language might introduce subtle biases or homogenized expressions that could influence self-reporting and therapeutic dialogue.

Additionally, the rise of AI-generated content underscores the importance of digital literacy in mental health care. Many patients may turn to chatbots for support, making it essential for psychologists to help them critically assess the reliability and limitations of such tools. Understanding AI's linguistic impact also has research implications, particularly in qualitative studies and diagnostic tools that rely on natural language analysis. By recognizing these trends, psychologists can better navigate the evolving relationship between technology, language, and mental health, ensuring they provide informed and adaptive care in an increasingly AI-influenced world.

Temporal Dynamics in Suicide: How Fluid Vulnerability Theory Informs Assessment and Intervention

Gavazz, J. (2025).

On Board with Psychology.

American Board of Clinical Psychology

A pdf is here.

Abstract

Background

Maintaining competency in current suicidology research is essential for psychologists to provide evidence-based care to individuals at risk for suicide. The Fluid Vulnerability Theory (FVT) provides a fundamental framework for understanding the dynamic and temporal nature of suicide risk, conceptualizing risk as fluctuating rather than static.

Purpose

This clinical review integrates contemporary FVT research to inform evidence-based assessment and intervention strategies for practicing psychologists working with at-risk populations.

Clinical Framework

FVT distinguishes between baseline risk (chronic, stable predispositions) and acute risk (dynamic states reactive to stressors). Recent research demonstrates that suicide risk follows nonlinear trajectories, with sudden crisis onset often occurring without overt warning signs. High-risk individuals exhibit distinct patterns, oscillating between baseline and elevated risk phases, while low-risk individuals typically stabilize in single states. Emotional regulation capacity emerges as a critical clinical factor, with compromised self-regulation predicting prolonged vulnerability and erratic state transitions.

Practice Applications

Contemporary research supports five evidence-based clinical recommendations: (1) implementing frequent temporal assessments using tools like ecological momentary assessment rather than relying on static evaluations; (2) assessing patients' self-regulatory processes during acute stress as early intervention points; (3) targeting specific mechanisms including loneliness and anxiety sensitivity in treatment planning; (4) closely monitoring recent stressors—particularly relationship conflicts, perceived failures, and disciplinary actions—that differentiate ideation from attempts; and (5) customizing interventions based on individual risk subtype patterns and nonlinear trajectories.

Clinical Significance

Understanding temporal dynamics of suicide risk enables psychologists to better identify at-risk individuals, time interventions appropriately, and tailor treatment approaches to specific risk states. This dynamic approach represents a significant advancement over traditional static risk assessment models, offering more precise and clinically relevant suicide prevention strategies for diverse populations and practice settings.

AlphaGo Moment for Model Architecture Discovery

Liu, Y., Nan, Y., Xu, W. et al. (2025)

arXiv.org

Abstract

While AI systems demonstrate exponentially improving capabilities, the pace of AI research itself remains linearly bounded by human cognitive capacity, creating an increasingly severe development bottleneck. We present ASI-Arch, the first demonstration of Artificial Superintelligence for AI research (ASI4AI) in the critical domain of neural architecture discovery--a fully autonomous system that shatters this fundamental constraint by enabling AI to conduct its own architectural innovation. Moving beyond traditional Neural Architecture Search (NAS), which is fundamentally limited to exploring human-defined spaces, we introduce a paradigm shift from automated optimization to automated innovation. ASI-Arch can conduct end-to-end scientific research in the domain of architecture discovery, autonomously hypothesizing novel architectural concepts, implementing them as executable code, training and empirically validating their performance through rigorous experimentation and past experience. ASI-Arch conducted 1,773 autonomous experiments over 20,000 GPU hours, culminating in the discovery of 106 innovative, state-of-the-art (SOTA) linear attention architectures. Like AlphaGo's Move 37 that revealed unexpected strategic insights invisible to human players, our AI-discovered architectures demonstrate emergent design principles that systematically surpass human-designed baselines and illuminate previously unknown pathways for architectural innovation. Crucially, we establish the first empirical scaling law for scientific discovery itself--demonstrating that architectural breakthroughs can be scaled computationally, transforming research progress from a human-limited to a computation-scalable process. We provide comprehensive analysis of the emergent design patterns and autonomous research capabilities that enabled these breakthroughs, establishing a blueprint for self-accelerating AI systems.

The article is here.

Here are some thoughts:

ASI-ARCH is a revolutionary AI system that helps other AIs invent new ways to learn. Instead of humans telling AI what to do, ASI-ARCH figures out new designs for AI brains all by itself. This means AI can now discover breakthroughs much faster than humans can, speeding up how quickly we make smart AI. It's like an "AlphaGo Moment" because the AI is showing us new, better ways to build AI that we hadn't thought of before, which can be terrifying.

Moral learning and Decision-Making across the lifespan

Lockwood, P. L., Van Den Bos, W., & Dreher, J. (2024).

Annual Review of Psychology.

https://doi.org/10.1146/annurev-psych-021324-060611

Abstract

Moral learning and decision-making are crucial throughout our lives, from infancy to old age. Emerging evidence suggests that there are important differences in learning and decision-making in moral situations across the lifespan, and these are underpinned by co-occurring changes in the use of model-based values and theory of mind. Here, we review the decision neuroscience literature on moral choices and moral learning considering four key concepts. We show how in the earliest years, a sense of self/other distinction is foundational. Sensitivity to intention versus outcome is crucial for several moral concepts and is most similar in our earliest and oldest years. Across all ages, basic shifts in the influence of theory of mind and model-free and model-based learning support moral decision-making. Moving forward, a computational approach to key concepts of morality can help provide a mechanistic account and generate new hypotheses to test across the whole lifespan.

Here are some thoughts:

The article highlights that moral learning and decision-making evolve dynamically throughout the lifespan, with distinct patterns emerging at different developmental stages. From early childhood to old age, individuals shift from rule-based moral reasoning toward more complex evaluations that integrate intentions, outcomes, and social context.

Understanding these developmental trajectories is essential for psychologists, as it informs age-appropriate interventions and expectations regarding moral behavior. Neuroscientific findings reveal that key brain regions such as the ventromedial prefrontal cortex (vmPFC), temporoparietal junction (TPJ), and striatum play critical roles in processing empathy, fairness, guilt, and social norms. These insights help explain how neurological impairments or developmental changes can affect moral judgment, particularly useful in clinical and neuropsychological settings.

Social influence also plays a significant role, especially during adolescence, where peer pressure and reputational concerns strongly shape moral decisions. This has practical implications for therapists working with youth, including strategies to build resilience against antisocial influences and promote prosocial behaviors.

The research further explores how deficits in moral learning are linked to antisocial behaviors, psychopathy, and conduct disorders, offering valuable perspectives for forensic psychology and clinical intervention planning.

Lastly, the article emphasizes the importance of cultural sensitivity, noting that moral norms vary across societies and change over time. For practicing psychologists, this underscores the need to adopt culturally informed approaches when assessing and treating clients from diverse backgrounds.

The Law Meets Psychological Expertise: Eight Best Practices to Improve Forensic Psychological Assessment

Neal, T. M., Martire, K. A.,  et al. (2022).

Annual Review of Law and Social Science, 

18(1), 169–192.

The article is here.

Abstract

 

We review the state of forensic mental health assessment. The field is in much better shape than in the past; however, significant problems of quality remain, with much room for improvement. We provide an overview of forensic psychology's history and discuss its possible future, with multiple audiences in mind. We distill decades of scholarship from and about fundamental basic science and forensic science, clinical and forensic psychology, and the law of expert evidence into eight best practices for the validity of a forensic psychological assessment. We argue these best practices should apply when a psychological assessment relies on the norms, values, and esteem of science to inform legal processes. The eight key considerations include (a) foundational validity of the assessment; (b) validity of the assessment as applied; (c) management and mitigation of bias; (d) attention to quality assurance; (e) appropriate communication of data, results, and opinions; (f) explicit consideration of limitations and assumptions; (g) weighing of alternative views or disagreements; and (h) adherence with ethical obligations, professional guidelines, codes of conduct, and rules of evidence.

Here are some thoughts:

This article outlines eight best practices designed to enhance the quality and validity of forensic psychological assessments. It provides a historical context for forensic psychology, discussing its evolution and future directions. Drawing on extensive research from basic science, forensic science, clinical and forensic psychology, and the law of expert evidence, the authors present key considerations for psychologists conducting assessments in legal settings. These practices include ensuring foundational and applied validity, managing biases, implementing quality assurance, communicating data and opinions appropriately, explicitly considering limitations, weighing alternative perspectives, and adhering to ethical guidelines. The article underscores the importance of these best practices to improve the reliability and scientific rigor of psychological expertise within the legal system.

Meta-analysis of risk factors for suicide after psychiatric discharge and meta-regression of the duration of follow-up

Tai, A., Pincham, H., Basu, A., & Large, M. (2025).

The Australian and New Zealand journal of psychiatry,

48674251348372. Advance online publication.

https://doi.org/10.1177/00048674251348372

Abstract

Background: Rates of suicide following discharge from psychiatric hospitals are extraordinarily high in the first week post-discharge and then decline steeply over time. The aim of this meta-analysis is to evaluate the strength of risk factors for suicide after psychiatric discharge and to investigate the association between the strength of risk factors and duration of study follow-up.

Methods: A PROSPERO-registered meta-analysis of observational studies was performed in accordance with PRISMA guidelines. Post-discharge suicide risk factors reported five or more times were synthesised using a random-effects model. Mixed-effects meta-regression was used to examine whether the strength of suicide risk factors could be explained by duration of study follow-up.

Results: Searches located 83 primary studies. From this, 63 risk estimates were meta-analysed. The strongest risk factors were previous self-harm (odds ratio = 2.75, 95% confidence interval = [2.37, 3.19]), suicidal ideation (odds ratio = 2.15, 95% confidence interval = [1.73, 2.68]), depressive symptoms (odds ratio = 1.84, 95% confidence interval = [1.48, 2.30]), and high-risk categorisation (odds ratio = 7.65, 95% confidence interval = [5.48, 10.67]). Significantly protective factors included age ⩽30, age ⩾65, post-traumatic stress disorder, and dementia. The effect sizes for the strongest post-discharge suicide risk factors did not decline over longer periods of follow-up.

Conclusion: The effect sizes of post-discharge suicide risk factors were generally modest, suggesting that clinical risk factors may have limited value in distinguishing between high-risk and low-risk groups. The highly elevated rates of suicide immediately after discharge and their subsequent decline remain unexplained.

Reimagining "Multiple Relationships" in Psychotherapy: Decolonial/Liberation Psychologies and Communal Selfhood

Lacerda-Vandenborn, E., et al. (2025).

American Psychologist, 80(4), 522–534.

https://doi.org/10.1037/amp0001441

Abstract

Promoting decolonial and liberation psychologies (DLPs) requires psychologists to critically interrogate taken-for-granted assumptions pertaining to psychotherapy relationships. One fruitful area of interrogation surrounds conceptualizations and practices concerning multiple relationships (MRs), wherein a psychologist and client share another form of relationship outside of the psychotherapy context. The prevention or minimization of MRs is widely viewed as an ethical imperative, codified within professional ethics codes and further encouraged through insurance and liability practices. From the standpoint of DLPs, the profession has not adequately grasped the extent to which psychotherapy relationships reflect individualistic selves that facilitate psychologists’ serving, however unwittingly, as “handmaidens of the status quo.” We present three practitioner testimonios from among our authors—Indigenous, Muslim, and lesbian, gay, bisexual, transgender, queer, questioning, and other sexual/gender minorities—to concretely demonstrate how the professional and ethical framing around this ubiquitous practice within psychology has served to flatten human relationships within a colonizing frame. We then discuss three problematic assumptions concerning MRs that are reflected in the American Psychological Association’s Ethics Code. We offer communal selfhood, a theoretical framework that aligns with DLPs, as a potential space for understanding and reframing MRs. We conclude with general recommendations for conceptualizing therapeutic relationships without recourse to a problematic conceptualization of MRs.

Public Significance Statement

Decolonial and liberation psychologies challenge conventional thinking concerning “multiple relationships” in psychotherapy. Discouragement of multiple relationships reflects an individualistic ideology and risk-aversive managerialism, protecting the profession more than promoting public welfare. Professional and ethical reforms, in line with a “communal selfhood” framework, would reinforce the profession’s commitments toward antiracism and anticolonialism.

Here are some thoughts:

The paper critically examines the traditional ethical stance on "multiple relationships" (MRs) in psychotherapy, arguing that the prevailing individualistic, risk-averse approach is often unsuitable for diverse communities. The article uniquely applies decolonial and liberation psychologies (DLPs) to challenge these Western-centric norms, advocating for a "communal selfhood" framework. It stands out by featuring compelling practitioner testimonios from Indigenous, Muslim, and LGBTQ+ psychologists, illustrating how rigid MR prohibitions can be detrimental in community-oriented contexts where interconnected relationships are vital for trust and healing. The article not only critiques existing guidelines but also offers recommendations for systemic reform, aiming to foster antiracism and anticolonialism within the psychology profession.

Crossing the Line: Daubert, Dual Roles, and the Admissibility of Forensic Mental Health Testimony

Gordon, S. G. (2016).

SSRN Electronic Journal.

Scholarly Works. 969.

Link is here.

Abstract

Psychiatrists and other mental health professionals often testify as forensic experts in civil commitment and criminal competency proceedings. When an individual clinician assumes both a treatment and a forensic role in the context of a single case, however, that clinician forms a dual relationship with the patient—a practice that creates a conflict of interest and violates professional ethical guidelines. The court, the parties, and the patient are all affected by this conflict and the biased testimony that may result from dual relationships. When providing forensic testimony, the mental health professional’s primary duty is to the court, not to the patient, and she has an obligation to give objective and truthful testimony. But this testimony can result in the patient’s detention or punishment, a legal outcome that implicates the mental health professional’s corresponding obligation to “do no harm” to the patient. Moreover, the conflict of interest created by a dual relationship can affect the objectivity and reliability of forensic testimony.

A dual clinical and forensic relationship with a single patient is contrary to quality patient care, and existing clinical and forensic ethical guidelines strongly discourage the practice. Notwithstanding the mental health community’s general consensus about the impropriety of the practice, many courts do not question the mental health professional’s ability to provide forensic testimony for a patient with whom she has a simultaneous clinical relationship. Moreover, some state statutes require or encourage clinicians at state-run facilities to engage in these multiple roles. This Article argues that the inherent conflict created by these dual roles does not provide a reliable basis for forensic mental health testimony under Federal Rule of Evidence 702 and should not be admitted as reliable expert testimony by courts. Because dual relationships are often initiated due to provider shortages and the unavailability of neutral forensic examiners, this Article will also discuss the use of telemedicine as a way to provide forensic evaluations in underserved areas, especially those where provider shortages have prompted mental health professionals to engage in dual clinical and forensic roles. Finally, this Article argues that courts should exercise their powers more broadly under Federal Rule of Evidence 706 to appoint neutral and independent mental health experts to conduct forensic evaluations in civil commitment and criminal competency proceedings.

Here are some thoughts:

The article explores the ethical and legal complexities surrounding mental health professionals who serve in dual roles—both as clinicians and forensic evaluators. The article highlights how these dual relationships can compromise objectivity and reliability in forensic testimony, a concern widely recognized within the psychiatric and psychological communities. Despite professional ethical codes discouraging such practices, courts often fail to exclude testimony from clinicians offering forensic opinions about their own patients. This inconsistency is particularly problematic under the Daubert standard, which mandates that trial judges act as gatekeepers to ensure expert testimony is both relevant and reliable. The piece argues that violating professional ethical norms—such as those against dual relationships—should be considered when evaluating the admissibility of forensic mental health testimony, especially since these violations are seen as markers of unreliability by the relevant scientific community. Additionally, the article touches on the practical implications of these dual role dilemmas, including the impact on patient care, legal outcomes, and the integrity of the judicial process. It concludes with a call for courts to take professional ethics more seriously when assessing the admissibility of expert testimony in forensic mental health cases.