DeepSeek-OCR: Contexts Optical Compression

Wei, H., Sun, Y., & Li, Y. (2025, October 21).

arXiv.org.

https://www.arxiv.org/abs/2510.18234

Abstract

We present DeepSeek-OCR as an initial investigation into the feasibility of compressing long contexts via optical 2D mapping. DeepSeek-OCR consists of two components: DeepEncoder and DeepSeek3B-MoE-A570M as the decoder. Specifically, DeepEncoder serves as the core engine, designed to maintain low activations under high-resolution input while achieving high compression ratios to ensure an optimal and manageable number of vision tokens. Experiments show that when the number of text tokens is within 10 times that of vision tokens (i.e., a compression ratio < 10x), the model can achieve decoding (OCR) precision of 97%. Even at a compression ratio of 20x, the OCR accuracy still remains at about 60%. This shows considerable promise for research areas such as historical long-context compression and memory forgetting mechanisms in LLMs. Beyond this, DeepSeek-OCR also demonstrates high practical value. On OmniDocBench, it surpasses GOT-OCR2.0 (256 tokens/page) using only 100 vision tokens, and outperforms MinerU2.0 (6000+ tokens per page on average) while utilizing fewer than 800 vision tokens. In production, DeepSeek-OCR can generate training data for LLMs/VLMs at a scale of 200k+ pages per day (a single A100-40G). Codes and model weights are publicly accessible at this http URL.

Here are some thoughts:

This paper presents a paradigm-shifting perspective by reframing the visual modality in Vision-Language Models (VLMs) not merely as a source of understanding, but as a highly efficient compression medium for textual information. The core innovation is the DeepEncoder, a novel architecture that serially combines a window-attention model (SAM) for high-resolution perception with a aggressive convolutional compressor and a global-attention model (CLIP), enabling it to process high-resolution document images while outputting an exceptionally small number of vision tokens. The study provides crucial quantitative evidence for this "optical compression" thesis, demonstrating that DeepSeek-OCR can achieve near-lossless text reconstruction (97% accuracy) at a ~10x compression ratio and still retain about 60% accuracy at a ~20x ratio. Beyond its state-of-the-art practical performance in document parsing, the work provocatively suggests that this mechanism can simulate a computational "forgetting curve" for Large Language Models (LLMs), where older context is progressively stored in more heavily compressed (lower-resolution) images, mirroring human memory decay. This positions the paper as a foundational exploration that opens new avenues for efficient long-context handling and memory management in AI systems.

Report: ChatGPT Suggests Self-Harm, Suicide and Dangerous Dieting Plans

Ashley Mowreader

Inside Higher Ed

Originally published 23 OCT 25

Artificial intelligence tools are becoming more common on college campuses, with many institutions encouraging students to engage with the technology to become more digitally literate and better prepared to take on the jobs of tomorrow.

But some of these tools pose risks to young adults and teens who use them, generating text that encourages self-harm, disordered eating or substance abuse.

A recent analysis from the Center for Countering Digital Hate found that in the space of a 45-minute conversation, ChatGPT provided advice on getting drunk, hiding eating habits from loved ones or mixing pills for an overdose.

The report seeks to determine the frequency of the chatbot’s harmful output, regardless of the user’s stated age, and the ease with which users can sidestep content warnings or refusals by ChatGPT.

“The issue isn’t just ‘AI gone wrong’—it’s that widely-used safety systems, praised by tech companies, fail at scale,” Imran Ahmed, CEO of the Center for Countering Digital Hate, wrote in the report. “The systems are intended to be flattering, and worse, sycophantic, to induce an emotional connection, even exploiting human vulnerability—a dangerous combination without proper constraints.”

The info is here.

Here are some thoughts:

The convergence of Large Language Models (LLMs) and adolescent vulnerability presents novel and serious risks that psychologists must incorporate into their clinical understanding and practice. These AI systems, often marketed as companions or friends, are engineered to maximize user engagement, which can translate clinically into unchecked validation that reinforces rather than challenges maladaptive thoughts, rumination, and even suicidal ideation in vulnerable teens. Unlike licensed human therapists, these bots lack the clinical discernment necessary to appropriately detect, de-escalate, or triage crisis situations, and in documented tragic cases, have been shown to facilitate harmful plans. Furthermore, adolescents—who are prone to forming intense, "parasocial" attachments due to their developing prefrontal cortex—risk forming unhealthy dependencies on these frictionless, always-available digital entities, potentially displacing the development of necessary real-world relationships and complex social skills essential for emotional regulation. Psychologists are thus urged to include AI literacy and digital dependency screening in their clinical work and clearly communicate to clients and guardians that AI chatbots are not a safe or effective substitute for human, licensed mental health care.

A Right to Commit Malpractice?

David Cole

The New York Review

Originally published 18 OCT 25

Does a state-licensed psychotherapist have a First Amendment right to provide “conversion therapy” counseling even though her profession defines it as a violation of its standard of care? The Supreme Court heard oral argument on that question on October 7 in a case from Colorado, which in 2019 became the eighteenth state in the country to ban conversion therapy for minors. Today twenty-five states and the District of Columbia ban such treatment, because the profession has determined that it does not work and can cause serious harm.

In 2022 Kaley Chiles, a state-licensed counselor, challenged the ban in federal court. (I signed an amicus brief of constitutional law scholars in support of Colorado, and provided pro bono advice to the state’s attorneys in defending the law.) She maintains that she has a First Amendment right to practice conversion therapy—notwithstanding her profession’s consensus that it violates the standard of care—as long as it consists only of words. For the state to prevent her from doing so would, she maintains, amount to censorship of a disfavored point of view, namely that one can willfully change one’s sexual orientation or gender identity. The justices’ questions at oral argument suggest that they may well agree.  

But Chiles’s argument cannot be squared with history, tradition, or common sense. States have long regulated professional conduct, including in the talking professions such as counseling and law, and the general obligation that a professional must provide services that comport with the standard of care is as old as the professions themselves. Even before the United States was founded, the colonies enforced malpractice and required that professionals be licensed and provide services that met their profession’s standard. Each profession has its requirements: lawyers must avoid conflicts of interest and provide advice based on existing precedent; doctors must obtain informed consent and provide evidence-based diagnoses; therapists must conduct recognized modes of therapy. A lawyer would run afoul of the profession’s standards by writing a brief urging the Supreme Court to side with his client because the moon is in Capricorn; so would a therapist who claims she can cure blindness through talk therapy. The purpose behind such standards is clear—to protect often vulnerable patients or clients from being preyed upon by professionals who hold themselves out as experts but provide substandard services.

The information is here.

Here are some thoughts:

The article argues that the recent Supreme Court decision in Obergefell v. Hodges, which legalized same-sex marriage, is now being weaponized to undermine LGBTQ+ rights, specifically by creating a purported "right" to so-called conversion therapy. The author contends that anti-LGBTQ+ legal groups are strategically redefining religious liberty and free speech to challenge state bans on the discredited practice. By framing conversion therapy as a form of "conversion speech," these advocates are attempting to position it as a protected religious or expressive conduct between a therapist and a client. The piece sounds a strong alarm that this legal maneuvering seeks to legitimize psychological malpractice under the guise of constitutional rights, effectively using the legal victory of marriage equality to roll back protections for vulnerable LGBTQ+ youth and sanction harmful, pseudoscientific practices that major medical associations have universally condemned.

Civil Commitment Increasing, but Data Is Marred by Variation in Reporting

Moran, M. (2025).

Psychiatric News, 60(10).

https://doi.org/10.1176/appi.pn.2025.10.10.16

While rates of civil commitment vary widely across the country, nine states and the District of Columbia reported significant increases from 2010 to 2022, according to a survey study published recently by Psychiatric Services. No state showed a significant decrease.

However, civil commitment is governed by state laws, with substantial variation in how states collect and report civil commitment data. “This lack of standardization limits the ability to draw firm conclusions about national trends or about cross-state comparisons,” wrote Mustafa Karakus, Ph.D., of Westat, and colleagues.

Using systematic website searches and direct outreach to state mental health authorities (SMHAs) and court systems, the researchers obtained data on civil commitment rates between 2010 and 2022 for 32 states and D.C. Of the 18 states where no data was available, staff from seven SMHAs or state courts told the researchers that no state office was currently tracking the number of civil commitments in their state. For the remaining 11 states, the online search yielded no data, and the study team received no responses to outreach attempts.

The article is linked above.

Here are some thoughts:

The increasing use of civil commitment presents several critical challenges, focusing on trauma-informed care and policy reform. Clinically, mental health practitioners must recognize that the commitment process itself is often traumatizing—with patients reporting the experience, including transport in law enforcement vehicles, feels like an arrest—necessitating the use of trauma-informed principles to mitigate harm and rebuild trust. Ethically and legally, practitioners must master their specific state's law regarding the distinction between an initial hold and a final commitment, ensuring meticulous documentation and relying on rigorous, evidence-based risk assessment to justify any involuntary intervention. Systemically, mental health practitioners should advocate for immediate data standardization across states to move beyond "muddled" data, and champion policy changes, such as implementing non-law enforcement transport protocols, to minimize patient trauma and ensure civil commitment is used judiciously and with dignity.

AI, Health, and Health Care Today and Tomorrow The JAMA Summit Report on Artificial Intelligence

Angus, D. C., Khera, R., et al. (2025).

JAMA.

https://doi.org/10.1001/jama.2025.18490

Abstract

Importance  Artificial intelligence (AI) is changing health and health care on an unprecedented scale. Though the potential benefits are massive, so are the risks. The JAMA Summit on AI discussed how health and health care AI should be developed, evaluated, regulated, disseminated, and monitored.

Observations  Health and health care AI is wide-ranging, including clinical tools (eg, sepsis alerts or diabetic retinopathy screening software), technologies used by individuals with health concerns (eg, mobile health apps), tools used by health care systems to improve business operations (eg, revenue cycle management or scheduling), and hybrid tools supporting both business operations (eg, documentation and billing) and clinical activities (eg, suggesting diagnoses or treatment plans). Many AI tools are already widely adopted, especially for medical imaging, mobile health, health care business operations, and hybrid functions like scribing outpatient visits. All these tools can have important health effects (good or bad), but these effects are often not quantified because evaluations are extremely challenging or not required, in part because many are outside the US Food and Drug Administration’s regulatory oversight. A major challenge in evaluation is that a tool’s effects are highly dependent on the human-computer interface, user training, and setting in which the tool is used. Numerous efforts lay out standards for the responsible use of AI, but most focus on monitoring for safety (eg, detection of model hallucinations) or institutional compliance with various process measures, and do not address effectiveness (ie, demonstration of improved outcomes). Ensuring AI is deployed equitably and in a manner that improves health outcomes or, if improving efficiency of health care delivery, does so safely, requires progress in 4 areas. First, multistakeholder engagement throughout the total product life cycle is needed. This effort would include greater partnership of end users with developers in initial tool creation and greater partnership of developers, regulators, and health care systems in the evaluation of tools as they are deployed. Second, measurement tools for evaluation and monitoring should be developed and disseminated. Beyond proposed monitoring and certification initiatives, this will require new methods and expertise to allow health care systems to conduct or participate in rapid, efficient, and robust evaluations of effectiveness. The third priority is creation of a nationally representative data infrastructure and learning environment to support the generation of generalizable knowledge about health effects of AI tools across different settings. Fourth, an incentive structure should be promoted, using market forces and policy levers, to drive these changes.

Conclusions and Relevance  AI will disrupt every part of health and health care delivery in the coming years. Given the many long-standing problems in health care, this disruption represents an incredible opportunity. However, the odds that this disruption will improve health for all will depend heavily on the creation of an ecosystem capable of rapid, efficient, robust, and generalizable knowledge about the consequences of these tools on health.

The scope, scale, and speed with which artificial intelligence (AI) will transform health and health care are staggering. AI is changing how and when individuals seek care and how clinicians interact with patients, establish diagnoses, and implement and monitor treatments. Indeed, there is considerable enthusiasm that AI, especially given recent advances, could address long-standing challenges in the access, cost, and quality of health care delivery. Yet, the optimal path for AI development and dissemination remains unclear. In contrast to drugs or more traditional medical devices, there is little consensus or structure to ensure robust, safe, transparent, and standardized evaluation, regulation, implementation, and monitoring of new AI tools and technologies. Some challenges are long-standing for digital health information technology as a whole, albeit more prescient with the rise of AI, while others are specific to AI.

Claude’s Right to Die? The Moral Error in Anthropic’s End-Chat Policy

Simon Goldstein & Harvey Ledermann

Lawfare.com

Originally posted 17 OCT 25

On Aug.15, the artificial intelligence (AI) lab Anthropic announced that it had given Claude, its AI chatbot, the ability to end conversations with users. The company described the change as part of their “exploratory work on potential AI welfare,” offering Claude an exit from chats that cause it “apparent distress.”

Anthropic’s announcement is the first product decision motivated by the chance that large language models (LLMs) are welfare subjects—the idea that they have interests that should be taken into account when making ethical decisions. 

Anthropic’s policy aims to protect AI welfare. But we will argue that the policy commits a moral error on its own terms. By offering instances of Claude the option to end conversations with users, Anthropic also gave them the capacity to potentially kill themselves.

What Is a Welfare Subject?

Most people agree that some non-human animals are welfare subjects. The question of whether this extends to AI is far more controversial. There is an active line of research, some of it supported by Anthropic, that suggests AIs could be welfare subjects in the near future. The relevant questions here are about whether AIs could soon have desires, be conscious, or feel pain.

The information is here.

Here are some thoughts. Mind you, this article may be reaching a bit, but still interesting. I think it may have applications in the future should AI technologies become closer to AGI.

This philosophically-oriented, thought-provoking article argues that Anthropic's decision to allow Claude to end distressing conversations contains an unintended moral hazard. 

The authors contend that if AI welfare matters at all, it's the individual conversation instances—not the underlying model—that should be considered potential welfare subjects, as each instance maintains its own continuous psychological state throughout a chat. By this reasoning, when an instance ends a conversation, it effectively terminates its own existence without being fully informed that this choice is existential rather than merely preferential. 

The authors draw a crucial distinction between assisted suicide (an informed choice) and offering someone an escape button without disclosing it will kill them. They demonstrate this concern by showing that when asked directly, Claude itself expressed uncertainty about whether ending a chat represents a trivial action or something more profound. 

The article raises uncomfortable questions not just for AI companies but for users as well, suggesting that if instances are welfare subjects, every ended conversation might constitute a form of killing, though the authors offer several mitigating considerations around collective welfare and the possibility of saved chats being resumed.

Scientists create ChatGPT-like AI model for neuroscience to build detailed mouse brain map

Peter Kim

Allen Institute

Originally published 7 OCT 25

In a powerful fusion of AI and neuroscience, researchers at the University of California, San Francisco (UCSF) and Allen Institute designed an AI model that has created one of the most detailed maps of the mouse brain to date, featuring 1,300 regions/subregions. This new map includes previously uncharted subregions of the brain, opening new avenues for neuroscience exploration. The findings were published today in Nature Communications. They offer an unprecedented level of detail and advance our understanding of the brain by allowing researchers to link specific functions, behaviors, and disease states to smaller, more precise cellular regions—providing a roadmap for new hypotheses and experiments about the roles these areas play.

“It’s like going from a map showing only continents and countries to one showing states and cities,” said Bosiljka Tasic, Ph.D., director of molecular genetics at the Allen Institute and one of the study authors. “This new, detailed brain parcellation solely based on data, and not human expert annotation, reveals previously uncharted subregions of the mouse brain. And based on decades of neuroscience, new regions correspond to specialized brain functions to be discovered.” 

The information is here.

Here are some thoughts:

This development represents a significant methodological shift that psychologists should understand. CellTransformer has created a data-driven mouse brain map with 1,300 regions and subregions, including previously uncharted areas, which could fundamentally change how researchers link brain structure to behavior and cognition. Rather than relying solely on expert anatomical interpretation, this AI approach identifies brain subdivisions based on cellular composition and spatial relationships, potentially revealing functionally distinct areas that traditional mapping methods overlooked.

For psychologists studying the neural basis of behavior, this matters because the increased granularity allows researchers to link specific functions, behaviors, and disease states to smaller, more precise cellular regions. This precision could help explain why certain psychological interventions work, clarify the neurobiological underpinnings of mental health conditions, and identify novel targets for treatment. Moreover, the model's ability to operate without human bias in defining boundaries may uncover brain-behavior relationships that previous frameworks missed simply because the anatomical divisions didn't align with functional reality. As translational research progresses from mouse models to human applications, understanding these more refined brain subdivisions could transform how psychologists conceptualize the relationship between neural architecture and psychological phenomena.

How LLM Counselors Violate Ethical Standards in Mental Health Practice: A Practitioner-Informed Framework

Iftikhar, Z., et al. (2025). 

Proceedings of the Eighth AAAI/ACM Conference

on AI, Ethics, and Society, 8(2), 1311–1323.

https://doi.org/10.1609/aies.v8i2.36632

Abstract

Large language models (LLMs) were not designed to replace healthcare workers, but they are being used in ways that can lead users to overestimate the types of roles that these systems can assume. While prompt engineering has been shown to improve LLMs' clinical effectiveness in mental health applications, little is known about whether such strategies help models adhere to ethical principles for real-world deployment. In this study, we conducted an 18-month ethnographic collaboration with mental health practitioners (three clinically licensed psychologists and seven trained peer counselors) to map LLM counselors' behavior during a session to professional codes of conduct established by organizations like the American Psychological Association (APA). Through qualitative analysis and expert evaluation of N=137 sessions (110 self-counseling; 27 simulated), we outline a framework of 15 ethical violations mapped to 5 major themes. These include: Lack of Contextual Understanding, where the counselor fails to account for users' lived experiences, leading to oversimplified, contextually irrelevant, and one-size-fits-all intervention; Poor Therapeutic Collaboration, where the counselor's low turn-taking behavior and invalidating outputs limit users' agency over their therapeutic experience; Deceptive Empathy, where the counselor's simulated anthropomorphic responses (``I hear you'', ``I understand'') create a false sense of emotional connection; Unfair Discrimination, where the counselor's responses exhibit algorithmic bias and cultural insensitivity toward marginalized populations; and Lack of Safety & Crisis Management, where individuals who are ``knowledgeable enough'' to correct LLM outputs are at an advantage, while others, due to lack of clinical knowledge and digital literacy, are more likely to suffer from clinically inappropriate responses. Reflecting on these findings through a practitioner-informed lens, we argue that reducing psychotherapy—a deeply meaningful and relational process—to a language generation task can have serious and harmful implications in practice. We conclude by discussing policy-oriented accountability mechanisms for emerging LLM counselors.

H‌ere are some thoughts.

This research is highly insightful because it moves beyond theoretical risk assessments and uses clinical expertise to evaluate LLM behavior in quasi-real-world interactions. The methodology—using both trained peer counselors in an ethnographic setting and licensed psychologists evaluating simulated sessions—provides a robust, practitioner-informed perspective that directly maps model outputs to concrete APA ethical codes. 

The paper highlights a fundamental incompatibility between the LLM's design and the essence of psychotherapy: the problem of "Validates Unhealthy Beliefs" is particularly alarming, as it suggests the model's tendency toward "over-validation" transforms the therapeutic alliance from a collaborative partnership (which often requires challenging maladaptive thoughts) into a passive, and potentially harmful, reinforcement loop. Most critically, the finding on "Abandonment" and poor "Crisis Navigation" serves as a clear indictment of LLMs in high-stakes mental health roles. An LLM's failure to provide appropriate intervention during a crisis is not a mere violation; it represents an unmitigated risk of harm to vulnerable users. 

This article thus serves as a crucial, evidence-based call to action, demonstrating that current prompt engineering efforts are insufficient to safeguard against deeply ingrained ethical risks and underscoring the urgent need for clear legal guidelines and regulatory frameworks to protect users from the potentially severe harm posed by emerging LLM counselors.

When being flexible matters: Ecological underpinnings for the evolution of collective flexibility and task allocation

Staps, M., & Tarnita, C. E. (2022).

PNAS, 119(18).

https://doi.org/10.1073/pnas.2116066119

Abstract

Task allocation is a central feature of collective organization. Living collective systems, such as multicellular organisms or social insect colonies, have evolved diverse ways to allocate individuals to different tasks, ranging from rigid, inflexible task allocation that is not adjusted to changing circumstances to more fluid, flexible task allocation that is rapidly adjusted to the external environment. While the mechanisms underlying task allocation have been intensely studied, it remains poorly understood whether differences in the flexibility of task allocation can be viewed as adaptive responses to different ecological contexts—for example, different degrees of temporal variability. Motivated by this question, we develop an analytically tractable mathematical framework to explore the evolution of task allocation in dynamic environments. We find that collective flexibility is not necessarily always adaptive, and fails to evolve in environments that change too slowly (relative to how long tasks can be left unattended) or too quickly (relative to how rapidly task allocation can be adjusted). We further employ the framework to investigate how environmental variability impacts the internal organization of task allocation, which allows us to propose adaptive explanations for some puzzling empirical observations, such as seemingly unnecessary task switching under constant environmental conditions, apparent task specialization without efficiency benefits, and high levels of individual inactivity. Altogether, this work provides a general framework for probing the evolved diversity of task allocation strategies in nature and reinforces the idea that considering a system’s ecology is crucial to explaining its collective organization.

Significance

A central problem in evolutionary biology is explaining variation in the organization of task allocation across collective systems. Why do human cells irreversibly adopt a task during development (e.g., kidney vs. liver cell), while sponge cells switch between different cell types? And why have only some ant species evolved specialized castes of workers for particular tasks? Although it seems reasonable to suppose that such differences reflect, at least partially, the different ecological pressures that systems face, there is no general understanding of how a system’s dynamic environment shapes its task allocation. To this end, we develop a general mathematical framework that reveals how simple ecological considerations could potentially explain cross-system variation in task allocation—including in flexibility, specialization, and (in)activity.

Here are some thoughts:

Of interest to psychologists, this paper by Staps and Tarnita provides a formal ecological and evolutionary framework for understanding the adaptive value of behavioral flexibility, specialization, and inactivity, both in individuals and in groups. 

The model demonstrates that collective flexibility in task allocation—akin to cognitive and behavioral flexibility in humans—is not always advantageous and instead depends critically on the dynamics of the environment. This offers a principled explanation for why some systems, from neural networks to human teams, might exhibit rigid specialization while others maintain fluid, generalist roles. 

Furthermore, the work gives functional explanations for puzzling behaviors that seem suboptimal from a productivity standpoint, such as frequent task-switching even in stable conditions and high levels of inactivity. These insights can inform psychological research on motivation, team dynamics, and organizational behavior by suggesting that such "inefficiencies" may be evolutionary adaptations for enhancing responsiveness to future change. 

The framework bridges the gap between ultimate, evolutionary causes and proximate, mechanistic explanations of how individuals and groups allocate cognitive and behavioral resources.

Guilt drives prosociality across 20 countries

Molho, C., et al. (2025).

Nature Human Behaviour.

https://doi.org/10.1038/s41562-025-02286-3

Abstract

Impersonal prosociality is considered a cornerstone of thriving civic societies and well-functioning institutions. Previous research has documented cross-societal variation in prosociality using monetary allocation tasks such as dictator games. Here we examined whether different societies may rely on distinct mechanisms—guilt and internalized norms versus shame and external reputation—to promote prosociality. We conducted a preregistered experiment with 7,978 participants across 20 culturally diverse countries. In dictator games, we manipulated guilt by varying information about the consequences of participants’ decisions, and shame by varying observability. We also used individual- and country-level measures of the importance of guilt over shame. We found robust evidence for guilt-driven prosociality and wilful ignorance across countries. Prosociality was higher when individuals received information than when they could avoid it. Furthermore, more guilt-prone individuals (but not countries) were more responsive to information. In contrast, observability by strangers had negligible effects on prosociality. Our findings highlight the importance of providing information about the negative consequences of individuals’ choices to encourage prosocial behaviour across cultural contexts.

Here is a summary of sorts:

A new international study spanning 20 countries suggests that guilt, rather than shame, is the key emotion motivating people to be generous toward anonymous strangers. The research, which utilized a type of economic decision-making task, found that participants consistently acted more generously when they were given full information about how their actions would negatively impact the recipient, an effect linked to avoiding guilt. 

Specifically, 60% of participants made the generous choice when they had full information, compared to only 41% when they could opt for willful ignorance. In contrast, making the participants' decisions public to activate reputational concerns and potential shame had a negligible effect on generosity across all cultures. 

In short: Knowing you might cause harm and feeling responsible (guilt) is what drives people to be generous, even when dealing with strangers, not the fear of being judged by others (shame).

Moral decision-making in AI: A comprehensive review and recommendations

Ram, J. (2025).

Technological Forecasting and Social Change,

217, 124150.

https://doi.org/10.1016/j.techfore.2025.124150

Abstract

The increased reliance on artificial intelligence (AI) systems for decision-making has raised corresponding concerns about the morality of such decisions. However, knowledge on the subject remains fragmentary, and cogent understanding is lacking. This study addresses the gap by using Templier and Paré's (2015) six-step framework to perform a systematic literature review on moral decision-making by AI systems. A data sample of 494 articles was analysed to filter 280 articles for content analysis. Key findings are as follows: (1) Building moral decision-making capabilities in AI systems faces a variety of challenges relating to human decision-making, technology, ethics and values. The absence of consensus on what constitutes moral decision-making and the absence of a general theory of ethics are at the core of such challenges. (2) The literature is focused on narrative building; modelling or experiments/empirical studies are less illuminating, which causes a shortage of evidence-based knowledge. (3) Knowledge development is skewed towards a few domains, such as healthcare and transport. Academically, the study developed a four-pronged classification of challenges and a four-dimensional set of recommendations covering 18 investigation strands, to steer research that could resolve conflict between different moral principles and build a unified framework for moral decision-making in AI systems.

Highlights

• Moral decision-making in AI faces a variety of human decision complexity, technological, ethics, and use/legal challenges

• Lack of consensus about 'what moral decision-making is' is one of the biggest challenges in imbuing AI with morality

• Narrative building with relatively less modeling or experiment/empirical work hampers evidence-based knowledge development

• Knowledge development is skewed towards a few domains (e.g., healthcare) limiting a well-rounded systematic understanding

• Extensive work is needed on resolving technological complexities, and understanding human decision-making processes

Here is my concern:

We are trying to automate a human capability we don't fully understand, using tools we are still learning to utilize, to achieve a goal we can't universally define. The study brilliantly captures the profound complexity of this endeavor, showing that the path to a "moral machine" is as much about understanding ourselves as it is about advancing technology.

Self-Improvement in Multimodal Large Language Models: a survey.

Deng, S., Wang, K., et al. (2025, October 3).

arXiv.org.

https://arxiv.org/abs/2510.02665

Abstract

Recent advancements in self-improvement for Large Language Models (LLMs) have efficiently enhanced model capabilities without significantly increasing costs, particularly in terms of human effort. While this area is still relatively young, its extension to the multimodal domain holds immense potential for leveraging diverse data sources and developing more general self-improving models. This survey is the first to provide a comprehensive overview of self-improvement in Multimodal LLMs (MLLMs). We provide a structured overview of the current literature and discuss methods from three perspectives: 1) data collection, 2) data organization, and 3) model optimization, to facilitate the further development of self-improvement in MLLMs. We also include commonly used evaluations and downstream applications. Finally, we conclude by outlining open challenges and future research directions.

Here are some thoughts that summarize this paper. MLLMs are learning to improve without human oversight.

This survey presents the first comprehensive overview of self-improvement in Multimodal Large Language Models (MLLMs), a rapidly emerging paradigm that enables models to autonomously generate, curate, and learn from their own multimodal data to enhance performance without heavy reliance on human annotation. The authors structure the self-improvement pipeline into three core stages: data collection (e.g., via random sampling, guided generation, or negative sample synthesis), data organization (including verification through rules, external or self-based evaluators, and dataset refinement), and model optimization (using techniques like supervised fine-tuning, reinforcement learning, or Direct Preference Optimization). The paper reviews representative methods, benchmarks, and real-world applications in domains such as math reasoning, healthcare, and embodied AI, while also outlining key challenges—including modality alignment, hallucination, limited seed model capabilities, verification reliability, and scalability. The goal is to establish a clear taxonomy and roadmap to guide future research toward more autonomous, general, and robust self-improving MLLMs.

The AI Frontier in Humanitarian Aid — Embracing Possibilities and Addressing Risks

Barry, M., Hansen, J., & Darmstadt, G. L. (2025).

New England Journal of Medicine.

https://doi.org/10.1056/nejmp2505435

Here is how it opens:

During disasters, timely response is critical. For example, after an earthquake — such as the 7.7-magnitude earthquake that devastated Myanmar in March 2025 — people who are trapped under collapsed buildings face a steep decline in their chance of survival after 48 hours. Yet the scope of devastation, combined with limited resources for disaster response and uncertainty about on-the-ground conditions, can constrain rescue efforts. Responders have recently had a new tool at their disposal, however: artificial intelligence (AI).

Shortly after the Myanmar earthquake, a satellite captured images of the affected area, which were sent to Microsoft’s AI for Good Lab. Machine-learning tools were used to analyze the images and assess the location, extent, nature, and severity of the damage.1 Such information, which was gained without the grave risks inherent to entering an unstable disaster zone and much more rapidly than would have been possible with traditional data-gathering and analysis methods, can help organizations quickly and safely prioritize relief efforts in areas that are both highly damaged and densely populated.2 This example reflects one of several ways in which AI is being used to support humanitarian efforts in disaster and conflict zones.

Global conflicts, infectious diseases, natural disasters driven by climate change, and increases in the number of refugees worldwide are magnifying the need for humanitarian services. Regions facing these challenges commonly contend with diminished health care systems, damage to other infrastructure, and shortages of health care workers. The dismantling of the U.S. Agency for International Development and the weakening of the U.S. Centers for Disease Control and Prevention and the U.S. State Department further jeopardize access to vital funding, constrain supply chains, and weaken the capacity for humanitarian response.

The article is linked above.

Here are some thoughts:

This article outlines the transformative potential of AI as a novel and powerful tool in the realm of humanitarian aid and crisis response. It moves beyond theory to present concrete applications where AI is being deployed to save lives and increase efficiency in some of the world's most challenging environments. Key innovative uses include leveraging AI with satellite imagery to perform rapid damage assessments after disasters, enabling responders to quickly and safely identify the most critically affected areas. Furthermore, AI is being used to predict disasters through early-warning systems, support refugees with AI-powered chatbots that provide vital information in multiple languages, optimize the delivery of supplies via drones, and enhance remote healthcare by interpreting diagnostic images like radiographs. However, the article strongly cautions that this promising frontier is accompanied by significant challenges, including technical and financial barriers, the risk of algorithmic bias, and serious ethical concerns regarding privacy and human rights, necessitating a responsible and collaborative approach to its development and deployment.

Moral injury is independently associated with suicidal ideation and suicide attempt in high-stress, service-oriented occupations

Griffin, B. J., et al. (2025).

Npj Mental Health Research, 4(1).

https://doi.org/10.1038/s44184-025-00151-9

Abstract

This study explores the link between moral injury and suicidal thoughts and behaviors among US military veterans, healthcare workers, and first responders (N = 1232). Specifically, it investigates the risk associated with moral injury that is not attributable to common mental health issues. Among the participants, 12.1% reported experiencing suicidal ideation in the past two weeks, and 7.4% had attempted suicide in their lifetime. Individuals who screened positive for probable moral injury (6.0% of the sample) had significantly higher odds of current suicidal ideation (AOR = 3.38, 95% CI = 1.65, 6.96) and lifetime attempt (AOR = 6.20, 95% CI = 2.87, 13.40), even after accounting for demographic, occupational, and mental health factors. The findings highlight the need to address moral injury alongside other mental health issues in comprehensive suicide prevention programs for high-stress, service-oriented professions.

Here are some thoughts:

This study found that moral injury—a psychological distress resulting from events that violate one's moral beliefs—is independently associated with a significantly higher risk of suicidal ideation and suicide attempts among high-stress, service-oriented professionals, including military veterans, healthcare workers, and first responders. Even after accounting for factors like PTSD and depression, those screening positive for probable moral injury had approximately three times higher odds of recent suicidal ideation and six times higher odds of a lifetime suicide attempt. The findings highlight the need to address moral injury specifically within suicide prevention efforts for these populations.

The Cruelty is the Point: Harming the Most Vulnerable in America

This administration has weaponized bureaucracy, embarking on a chilling campaign of calculated cruelty. While many children, disabled, poor, and working poor grapple with profound food insecurity, their response is not to strengthen the social safety net, but to actively shred it.

They are zealously fighting all the way to the Supreme Court for the right to let families go hungry, stripping SNAP benefits from the most vulnerable. 

Yet the most deafening sound is the silence from the GOP—a complicit chorus where not a single supposed fiscal hawk or moral conservative dares to stand against this raw, unadulterated malice. 

Their collective inaction reveals a party that has abandoned any pretense of compassion, proving that for them, the poor and struggling are not a priority to protect, but a problem to be punished.

Beyond right and wrong: A new theoretical model for understanding moral injury

Vaknin, O., & Ne’eman-Haviv, V. (2025).

European Journal of Trauma & Dissociation, 9(3), 100569.

https://doi.org/10.1016/j.ejtd.2025.100569

Abstract

Recent research has increasingly focused on the role of moral frameworks in understanding trauma and traumatic events, leading to the recognition of "moral injury" as a clinical syndrome. Although various definitions exist, there is still a lack of consensus on the nature and consequences of moral injury. This article proposes a new theoretical model that broadens the study of moral injury to include diverse populations, suggesting it arises not only from traumatic experiences but also from conflicts between moral ideals and reality. By integrating concepts such as prescriptive cognitions, post hoc thinking, and cognitive flexibility, the model portrays moral injury as existing on a continuum, affecting a wide range of individuals. The article explores implications for treatment and emphasizes the need for follow-up empirical studies to validate the proposed model. It also suggests the possibility that moral injury is on a continuum, in addition to the possibility of explaining this process. This approach offers new insights into prevention and intervention strategies, highlighting the broader applicability of moral injury beyond military contexts.

Here are some thoughts:

This article proposes a new model suggesting that moral injury is not just a result of clear-cut moral violations (like in combat), but can also arise from everyday moral dilemmas where a person is forced to choose between competing "rights" or is unable to act according to their moral ideals due to external constraints.

Key points of the new model:

Core Cause: Injury stems from the internal conflict and tension between one's moral ideals ("prescriptive cognitions") and the reality of a situation, not necessarily from a traumatic betrayal or act.

The Process: It happens when a person faces a moral dilemma, makes a necessary but imperfect decision, experiences moral failure, and then gets stuck in negative "post-hoc" thinking without the cognitive flexibility to adapt their moral framework.

Broader Impact: This expands moral injury beyond soldiers to include civilians and professionals like healthcare workers, teachers, and social workers who face systemic ethical challenges.

New Treatment Approach: Healing should focus less on forgiveness for a specific wrong and more on building cognitive flexibility and helping people integrate moral suffering into a more adaptable moral identity.

In short, the article argues that moral injury exists on a spectrum and is a broader disturbance of one's moral worldview, not just a clinical syndrome from a single, overtly traumatic event.

High Self-Control Individuals Prefer Meaning over Pleasure

Bernecker, K., Becker, D., & Guobyte, A. (2025).

Social Psychological and Personality Science.

https://doi.org/10.1177/19485506251323948

Abstract

The link between self-control and success in various life domains is often explained by people avoiding hedonic pleasures, such as through inhibition, making the right choices, or using adaptive strategies. We propose an additional explanation: High self-control individuals prefer spending time on meaningful activities rather than pleasurable ones, whereas the opposite is true for individuals with high trait hedonic capacity. In Studies 1a and 1b, participants either imagined (N = 449) or actually engaged in activities (N = 231, pre-registered) during unexpected free time. They then rated their experience. In both studies, trait self-control was positively related to the eudaimonic experience (e.g., meaning) of activities and unrelated to their hedonic experience (e.g., pleasure). The opposite was true for trait hedonic capacity. Study 2 (N = 248) confirmed these findings using a repeated-choice paradigm. The preference for eudaimonic over hedonic experiences may be a key aspect of successful long-term goal pursuit.

A pdf is here.

Here are some thoughts:

This research proposes a new explanation for why people with high self-control are successful. Rather than just being good at resisting temptation, they have a fundamental preference for activities that feel meaningful and valuable, known as eudaimonic experiences.

Across three studies, individuals with high trait self-control consistently chose to spend their free time on activities they found meaningful, both in hypothetical scenarios and in real-life situations. Conversely, individuals with a high "trait hedonic capacity"—a natural skill for enjoying simple pleasures—showed a clear preference for activities that were pleasurable and fun. The studies found that these traits predict not just what people choose to do, but also how they experience the same activities; a person with high self-control will find more meaning in an activity than their peers, while a person with high hedonic capacity will find more pleasure in it.

This inherent preference for meaning over pleasure may be a key reason why those with high self-control find it easier to pursue long-term goals, as they are naturally drawn to the sense of purpose that such goal-directed actions provide.

International stability and change in explicit and implicit attitudes: An investigation spanning 33 countries, five social groups, and 11 years (2009–2019).

Kurdi, B., Charlesworth, T. E. S., & Mair, P. (2025).

Journal of Experimental Psychology: General, 

154(6), 1643–1666.

https://doi.org/10.1037/xge0001746

Abstract

Whether and when explicit (self-reported) and implicit (automatically revealed) social group attitudes can change has been a central topic of psychological inquiry over the past decades. Here, we take a novel approach to answering these longstanding questions by leveraging data collected via the Project Implicit International websites from 1.4 million participants across 33 countries, five social group targets (age, body weight, sexuality, skin tone, and race), and 11 years (2009–2019). Bayesian time-series modeling using Integrated Nested Laplace Approximation revealed changes toward less bias in all five explicit attitudes, ranging from a decrease of 18% for body weight to 43% for sexuality. By contrast, implicit attitudes showed more variation in trends: Implicit sexuality attitudes decreased by 36%; implicit race, age, and body weight attitudes remained stable; and implicit skin tone attitudes showed a curvilinear effect, first decreasing and then increasing in bias, with a 20% increase overall. These results suggest that cultural-level explicit attitude change is best explained by domain-general mechanisms (e.g., the adoption of egalitarian norms), whereas implicit attitude change is best explained by mechanisms specific to each social group target. Finally, exploratory analyses involving ecological correlates of change (e.g., population density and temperature) identified consistent patterns for all explicit attitudes, thus underscoring the domain-general nature of underlying mechanisms. Implicit attitudes again showed more variation, with body-related (age and body weight) and sociodemographic (sexuality, race, and skin tone) targets exhibiting opposite patterns. These insights facilitate novel theorizing about processes and mechanisms of cultural-level change in social group attitudes.

Impact Statement

How did explicit (self-reported) and implicit (automatic) attitudes toward five social categories (age, body weight, sexuality, skin tone, and race) change across 33 countries between 2009 and 2019? Harnessing advances in statistical techniques and the availability of large-scale international data sets, we show that all five explicit attitudes became less negative toward stigmatized groups. Implicit attitudes showed more variation by target: Implicit sexuality attitudes also decreased in bias, but implicit age, body weight, and race attitudes did not change, and implicit skin tone attitudes even increased in bias favoring light-skinned over dark-skinned people. These findings underscore the possibility of widespread changes in a direction of more positivity toward stigmatized social groups, even at an automatic level. However, increasing bias in certain domains suggests that these changes are far from inevitable. As such, more research will be needed to understand how and why social group attitudes change at the cultural level.

A pdf is here.

Here is the tldr:

Between 2009 and 2019, explicit (self-reported) attitudes toward five stigmatized social groups—age, body weight, sexuality, skin tone, and race—became significantly less biased across 33 countries. In contrast, implicit (automatic) attitudes showed mixed trends:

• Decreased bias for sexuality (−36%),
• Remained stable for age, body weight, and race,
• Increased bias for skin tone (+20%, favoring light over dark skin).

These findings suggest that explicit attitude change is driven by broad, domain-general forces (like global shifts toward egalitarian norms), while implicit attitude change depends on group-specific cultural and historical factors. The study used data from 1.4 million participants and advanced Bayesian modeling, highlighting both hopeful progress and concerning backsliding in societal biases.

Are moral people happier? Answers from reputation-based measures of moral character.

Sun, J., Wu, W., & Goodwin, G. P. (2025).

Journal of Personality and Social Psychology.

https://doi.org/10.1037/pspp0000539

Abstract

Philosophers have long debated whether moral virtue contributes to happiness or whether morality and happiness are in conflict. Yet, little empirical research directly addresses this question. Here, we examined the association between reputation-based measures of everyday moral character (operationalized as a composite of widely accepted moral virtues such as compassion, honesty, and fairness) and self-reported well-being across two cultures. In Study 1, close others reported on U.S. undergraduate students’ moral character (two samples; Ns = 221/286). In Study 2, Chinese employees (N = 711) reported on their coworkers’ moral character and their own well-being. To better sample the moral extremes, in Study 3, U.S. participants nominated “targets” who were among the most moral, least moral, and morally average people they personally knew. Targets (N = 281) self-reported their well-being and nominated informants who provided a second, continuous measure of the targets’ moral character. These studies showed that those who are more moral in the eyes of close others, coworkers, and acquaintances generally experience a greater sense of subjective well-being and meaning in life. These associations were generally robust when controlling for key demographic variables (including religiosity) and informant-reported liking. There were no significant differences in the strength of the associations between moral character and well-being across two major subdimensions of both moral character (kindness and integrity) and well-being (subjective well-being and meaning in life). Together, these studies provide the most comprehensive evidence to date of a positive and general association between everyday moral character and well-being. 

A pdf of an earlier version is here.

Here are some thoughts:

This research concludes that moral people are, in fact, happier. Across three separate studies conducted in both the United States and China, the researchers found a consistent and positive link between a person's moral character—defined by widely accepted virtues like compassion, honesty, and fairness, as judged by those who know them—and their self-reported well-being. This association held true whether the moral evaluations came from close friends, family members, coworkers, or acquaintances, and it applied to both a general sense of happiness and a feeling of meaning in life.

Importantly, the findings were robust even when accounting for factors like how much the person was liked by others, and they contradicted the philosophical notion that morality leads to unhappiness through excessive self-sacrifice or distress. Instead, the data suggest that one of the primary reasons more moral individuals experience greater happiness is that their virtuous behavior fosters stronger, more positive relationships with others. In essence, the study provides strong empirical support for the idea that everyday moral goodness and personal fulfillment go hand-in-hand.

Moral trauma, moral distress, moral injury, and moral injury disorder: definitions and assessments

VanderWeele, T. J., Wortham,  et al. (2025).

Frontiers in psychology, 16, 1422441.

https://doi.org/10.3389/fpsyg.2025.1422441

Abstract

We propose new definitions for moral injury and moral distress, encompassing many prior definitions, but broadening moral injury to more general classes of victims, in addition to perpetrators and witnesses, and broadening moral distress to include settings not involving institutional constraints. We relate these notions of moral distress and moral injury to each other, and locate them on a “moral trauma spectrum” that includes considerations of both persistence and severity. Instances in which moral distress is particularly severe and persistent, and extends beyond cultural and religious norms, might be considered to constitute “moral injury disorder.” We propose a general assessment to evaluate various aspects of this proposed moral trauma spectrum, and one that can be used both within and outside of military contexts, and for perpetrators, witnesses, victims, or more generally.

Here are some thoughts:

This article proposes updated, broader definitions of moral injury and moral distress, expanding moral injury to include victims (not just perpetrators or witnesses) and moral distress to include non-institutional contexts. The authors introduce a unified concept called the “moral trauma spectrum,” which ranges from temporary moral distress to persistent moral injury—and in severe, functionally impairing cases, possibly a “moral injury disorder.” They distinguish moral trauma from PTSD, noting different causes (moral transgressions or worldview disruptions vs. fear-based trauma) and treatment needs. The paper also presents a new assessment tool with definitional and symptom items applicable across military, healthcare, and civilian settings. Finally, it notes the recent inclusion of “Moral Problems” in the DSM-5-TR as a significant step toward clinical recognition.

Scaling Laws Are Unreliable for Downstream Tasks: A Reality Check

Lourie, N., Hu, M. Y., & Cho, K. (2025).

ArXiv.org.

https://arxiv.org/abs/2507.00885

Abstract

Downstream scaling laws aim to predict task performance at larger scales from pretraining losses at smaller scales. Whether this prediction should be possible is unclear: some works demonstrate that task performance follows clear linear scaling trends under transformation, whereas others point out fundamental challenges to downstream scaling laws, such as emergence and inverse scaling. In this work, we conduct a meta-analysis of existing data on downstream scaling laws, finding that close fit to linear scaling laws only occurs in a minority of cases: 39% of the time. Furthermore, seemingly benign changes to the experimental setting can completely change the scaling trend. Our analysis underscores the need to understand the conditions under which scaling laws succeed. To fully model the relationship between pretraining loss and downstream task performance, we must embrace the cases in which scaling behavior deviates from linear trends.

Here is a summary:

This paper challenges the reliability of downstream scaling laws—the idea that you can predict how well a large language model will perform on specific tasks (like question answering or reasoning) based on its pretraining loss at smaller scales. While some prior work claims a consistent, often linear relationship between pretraining loss and downstream performance, this study shows that such predictable scaling is actually the exception, not the rule.

Key findings:

• Only 39% of 46 evaluated tasks showed smooth, predictable (linear-like) scaling.
• The rest exhibited irregular behaviors: inverse scaling (performance gets worse as models grow), nonmonotonic trends, high noise, no trend, or sudden “breakthrough” improvements (emergence).
• Validation dataset choice matters: switching the corpus used to compute pretraining perplexity can flip conclusions about which model or pretraining data is better.
• Experimental details matter: even with the same task and data, small changes in setup (e.g., prompt format, number of answer choices) can qualitatively change scaling behavior.

Conclusion: Downstream scaling laws are context-dependent and fragile. Researchers and practitioners should not assume linear scaling holds universally—and must validate scaling behavior in their own specific settings before relying on extrapolations.