Folk psychological attributions of consciousness to large language models

Colombatto, C., & Fleming, S. M. (2024).

Neuroscience of consciousness, 2024(1)

https://doi.org/10.1093/nc/niae013

Abstract

Technological advances raise new puzzles and challenges for cognitive science and the study of how humans think about and interact with artificial intelligence (AI). For example, the advent of large language models and their human-like linguistic abilities has raised substantial debate regarding whether or not AI could be conscious. Here, we consider the question of whether AI could have subjective experiences such as feelings and sensations ('phenomenal consciousness'). While experts from many fields have weighed in on this issue in academic and public discourse, it remains unknown whether and how the general population attributes phenomenal consciousness to AI. We surveyed a sample of US residents (n = 300) and found that a majority of participants were willing to attribute some possibility of phenomenal consciousness to large language models. These attributions were robust, as they predicted attributions of mental states typically associated with phenomenality-but also flexible, as they were sensitive to individual differences such as usage frequency. Overall, these results show how folk intuitions about AI consciousness can diverge from expert intuitions-with potential implications for the legal and ethical status of AI.

Conclusions

In summary, our investigation of folk psychological attributions of consciousness revealed that most  people are willing  to attribute some form of phenomenality to LLMs: only a third of our sample thought that ChatGPT definitely did not have subjective experience, while two-thirds of our sample thought that  ChatGPT had varying degrees of phenomenal consciousness. The relatively high rates of consciousness  attributions in this sample are somewhat surprising, given that experts in neuroscience and consciousness science currently estimate that LLMs are highly unlikely to be conscious (Butlin et al. 2023, LeDoux et al. 2023). These findings thus highlight a discrepancy between folk  intuitions and expert  opinions on artificial consciousness—with significant implications for the ethical, legal, and moral status of AI.

Here is a link to the pdf.

Summary:

The paper examines how people attribute consciousness and mental states to large language models (LLMs) like GPT-3 based on folk psychology intuitions.  Folk psychology refers to how laypeople reason about minds and mental states based on observable behavior.

Attributions of Consciousness

In two studies, participants interacted with GPT-3 and rated the extent to which they attributed various mental states to the LLM.  Participants attributed significant levels of consciousness, including the ability to experience emotions, have beliefs, and understand language.  However, attributions were lower for more advanced capacities like self-awareness and intentionality.

Factors Influencing Attributions

Attributions increased the more the LLM's responses appeared coherent, thoughtful and human-like.

Attributions decreased when participants were reminded that the LLM is an artificial system without subjective experiences. Individuals' beliefs about machine consciousness and their exposure to AI also impacted attributions.

Implications

The studies reveal a tendency for people to anthropomorphize and over-attribute mental capacities to sophisticated language models based on their surface behavior.  This has implications for managing public expectations around AI capabilities and potential risks of deception. The findings highlight the need for transparency about the actual cognitive architectures underlying LLMs to mitigate misunderstandings.

In summary, the research demonstrates how people's folk psychology leads them to project human-like mental states onto LLMs in ways that may not accurately reflect the systems' true capabilities and nature.

What does my group consider moral?: How social influence shapes moral expressions

del Rosario, K., Van Bavel, J. J., & West, T.

PsyArXiv (2024, May 8).

https://doi.org/10.31234/osf.io/dwzq2

Abstract

Although morality is often characterized as a set of stable values that are deeply held, we argue that moral expressions are highly malleable and sensitive to social norms. For instance, norms can either lead people to exaggerate their expressions of morality (such as on social media) or restrain them (such as in professional settings). In this paper, we discuss why moral expressions are subject to social influence by considering two goals that govern social influence: affiliation goals (the desire to affiliate with one’s group) and accuracy goals (the desire to be accurate in ambiguous situations). Different from other domains of social influence, we argue that moral expressions often satisfy both affiliation goals (“I want to fit in with the group”) and accuracy goals (“I want to do the right thing”). As such, the fundamental question governing moral expressions is: “what does my group consider moral?” We argue that this central consideration achieves both goals underlying social influence and drives moral expressions. We outline the ways in which social influence shapes moral expressions, from unconsciously copying others’ behavior to expressing outrage to gain status within the group. Finally, we describe when the same goals can result in different behaviors, highlighting how context-specific norms can encourage (or discourage) moral expressions. We explain how this framework will be helpful in understanding how identity, norms, and social contexts shape moral expressions.

Conclusion

Our review examines moral expressions through the lens of social influence, illustrating the critical role of the social environment in shaping moral expressions. Moral expressions serve a social purpose, such as affiliating with a group, and are influenced by various goals, including understanding the appropriate emotional response to moral issues and conforming to others' expressions to fit in. These influences become evident in different contexts, where norms either encourage exaggerated expressions, like on social media, or restraint, such as in professional settings. For this reason, different forms of influence can have vastly different implications. As such, the fundamental social question governing moral expressions for people in moral contexts is: “What does my group consider moral?” However, much of the morality literature does not account for the role of social influence in moral expressions. Thus, a social norms framework will be helpful in understanding how social contexts shape moral expression.

Here is a summary:

The research argues that moral expressions (outward displays of emotions related to right and wrong) are highly malleable and shaped by social norms and contexts, contrary to the view that morality reflects stable convictions. It draws from research on normative influence (conforming to gain social affiliation) and informational influence (seeking accuracy in ambiguous situations) to explain how moral expressions aim to satisfy both affiliation goals ("fitting in with the group") and accuracy goals ("doing the right thing").

The key points are:

1. Moral expressions vary across contexts because people look to their social groups to determine what is considered moral behavior.
2. Affiliation goals (fitting in) and accuracy goals (being correct) are intertwined for moral expressions, unlike in other domains where they are distinct.
3. Social influence shapes moral expressions in various ways, from unconscious mimicry to outrage expressions for gaining group status.
4. Context-specific norms can encourage or discourage moral expressions by prioritizing affiliation over accuracy goals, or vice versa.
5. The motivation to be seen as moral contributes to the malleability of moral expressions across social contexts.

Examining Potential Psychological Protective and Risk Factors for Stress and Burnout in Social Workers

Maddock, A.

Clin Soc Work J (2024).

https://doi.org/10.1007/s10615-024-00924-3

Abstract

Social work professionals experience high levels of stress and burnout. Stress and burnout can have a negative impact on the individual social worker, the organisations they work for, and perhaps most importantly, the quality of care that marginalised groups that are supported by social workers receive. Several work-related predictors of stress and burnout have been identified; however, no studies have examined the underlying psychological protective and risk factors which might help to explain changes in social worker stress and burnout. Using the clinically modified Buddhist psychological model (CBPM) as a theoretical framework, this cross-sectional study attempted to identify psychological protective and risk factors for stress and burnout in 121 social workers in Northern Ireland, using structural equation modelling, and conditional process analyses. This study provided promising preliminary evidence for a mediated effect CBPM as being a potentially useful explanatory framework of variation in social worker stress, emotional exhaustion, and depersonalisation. This study also provided evidence that several CBPM domains could have a direct effect on personal accomplishment. This study provides preliminary evidence that support programmes, which have the capacity to improve each CBPM domain (mindfulness, acceptance, attention regulation/decentering, self-compassion, non-attachment, and non-aversion) and reduce experiences of worry and rumination, are likely to support social workers to experience reduced stress, emotional exhaustion, depersonalisation of service users, and improvements in personal accomplishment.

From the Discussion

The aims of this paper were to provide more theoretical transparency on what some of the most important protective and risk factors for social worker stress and burnout are, using the data attained from social workers in Northern Ireland. To support our analysis, the CBPM (Maddock, 2023), which is a multi-faceted stress coping, cognitive and emotional regulation theory was used. Using structural equation modelling, though the direct and mediated effects CBPM was found to be an acceptable fit to the data on perceived stress, emotional exhaustion, and depersonalisation, our results indicate that the mediated effects CBPM model was a better fit to the data on each of these outcomes. Most of the significant conditional effects found using Process, between the CBPM domains and perceived stress, emotional exhaustion, depersonalisation were also mediated by either worry or rumination and sometimes both (e.g., stress), highlighting that negative thinking styles, such as worry and rumination, are likely to be a key risk factor for the development of stress and emotional exhaustion in social workers along with the depersonalisation of service users. This supports Kazdin (2009), who asserted that individual risk or protective factors (in our case, worry and rumination respectively) can impact multiple outcomes. This highlights how interventions e.g., MBPs or CBT, that aim to reduce feelings of stress, emotional exhaustion, and depersonalisation of service users in social work, could be more parsimonious, and effective, if they focussed on supporting social workers to regulate the extent to which they engage in worry or rumination in response to feelings of stress or burnout. This could be achieved, particularly by MBPs, through the development of each CBPM domain (i.e., mindfulness, attention regulation/decentering, acceptance, self-compassion, non-attachment and non-aversion), each of which have been identified as approach oriented coping strategies, which have been the capacity to support social workers to regulate the extent to which they worry or rumination (Maddock, 2023).

It is clear from this study that the effects of different potential psychological protective and risk factors for social worker stress and burnout, are likely to be complex. The limited literature available attempting to explain the patterns of relationships between mindfulness variables and mental health and well-being outcomes such as stress and burnout has usually identified either significant direct (e.g., Hölzel et al., 2011) or mediated (e.g., Gu et al., 2015) pathways, but not both at the same time. This study thus highlights the potentially complex direct and mediated interactions between mindfulness variables e.g., acceptance, attention regulation, stress, and different domains of burnout in social work. This is supported by the fact that most of the significant effects of each CBPM domain on stress, burnout-emotional exhaustion, burnout-depersonalisation, and burnout-personal accomplishment were found to be mediated by either worry or rumination. A number of CBPM domains e.g., acceptance and attention regulation/decentering also appeared to have a direct effect on stress and burnout-depersonalisation. These findings also support Kazdin (2009) who highlighted that outcomes, such as stress and depersonalisation, can be reduced through multiple pathways i.e., through both direct and mediated relationships.

The health care workforce crisis is already here

Caitlin Owen

axios.com

Originally posted 7 June 24

Demoralized doctors and nurses are leaving the field, hospitals are sounding the alarm about workforce shortages and employees are increasingly unionizing and even going on strike in high-profile disputes with their employers.

Why it matters: Dire forecasts of health care worker shortages often look to a decade or more from now, but the pandemic — and its ongoing fallout — has already ushered in a volatile era of dissatisfied workers and understaffed health care facilities.

• Some workers and experts say understaffing is, in some cases, the result of intentional cost cutting. Regardless, patients' access to care and the quality of that care are at risk.
• "There are 83 million Americans today who don't have access to primary care," said Jesse Ehrenfeld, president of the American Medical Association. "The problem is here. It's acute in rural parts of the country, it's acute in underserved communities."

The big picture: Complaints about understaffing, administrative burdens and inadequate wages aren't new, but they are getting much louder — and more health workers are leaving their jobs or cutting back their hours.

The info is here.

Here are some thoughts:

The news of the healthcare workforce crisis being "already here" is deeply concerning.  It's not just about future projections; it's about the impact on patient care, provider well-being, and the ethical obligations we all share.

Providers will likely walk an ethical tightrope, that will likely have negative consequences. Imagine a doctor facing a packed waiting room, knowing some patients won't receive the time and attention they deserve.  This is the reality for many providers stretched thin by staffing shortages. It creates an ethical tightrope: how to deliver quality care amidst overwhelming pressure.  Burnout, compassion fatigue, and even medical errors become more likely.  This is likely the starting point for the possibility of moral distress and/or moral injury.

The crisis isn't just a burden on healthcare providers or institutions. It's a societal challenge.  Policymakers, educators, and even patients themself can play a role.

This isn't about pointing fingers; it's about recognizing a shared responsibility.  By working together, we can ensure a healthcare system that is ethical, sustainable, and provides quality care for all.

Morals Versus Ethics: Building An Organizational Culture Of Trust And Transparency

Pamela Furr

Forbes.com

Originally posted 6 May 24

Here are two excerpts:

Prioritize Transparency And Integrity

Our team is a diverse mix of ages, cultures, races and backgrounds, and we all bring unique experiences and perspectives to the table. If a colleague says or does something that doesn’t sit right with you, take a moment to pause, process and then approach them. Share how you felt in the moment—this can be as simple as saying, “My feelings were hurt when you did that” or “I didn’t think the language you used earlier was appropriate.” Give them the opportunity to explain or apologize before gossiping with coworkers or silently holding onto resentments. Trust each other to have open, honest conversations, and you can often defuse conflicts before they escalate.

(cut)

Build A Sense Of Community

Set the tone for open dialogue and mutual respect in your organization. By modeling these values in your interactions with others, you can inspire your team to uphold the same standards. Foster a culture in which you advocate for yourself and others and try to learn from others as well. Approach things you don’t understand with a spirit of curiosity and compassion, assuming positive intent until proven otherwise. Ask questions, and truly seek to understand someone else’s point of view.

I believe that an essential part of being a leader is ensuring that our employees feel safe, protected and heard when they come to work. We can work to hold external governing boards accountable to the standards they set, but we can also do everything in our power to create a culture of trust, transparency and accountability within our own organizations.

The article is here.

Here is my summary:

The article discusses the difference between morals and ethics. Morals are personal beliefs and values that guide our actions, while ethics are a set of rules established by a community or governing body.

The author describes a situation where a trainee made a false sexual harassment claim against her mentor. The certifying board refused to take any action because they saw it as an employment contract issue. The author argues that governing boards should take a stronger stance in upholding ethics within their professions.

The article concludes with the author's thoughts on creating an ethical and transparent workplace culture. The author emphasizes the importance of open communication, understanding policies and procedures, and building a sense of community. By following these principles, organizations can create a safe and supportive environment for their employees.

Attributions toward artificial agents in a modified Moral Turing Test

Aharoni, E., Fernandes, S., Brady, D.J. et al.

Sci Rep 14, 8458 (2024).

https://doi.org/10.1038/s41598-024-58087-7

Abstract

Advances in artificial intelligence (AI) raise important questions about whether people view moral evaluations by AI systems similarly to human-generated moral evaluations. We conducted a modified Moral Turing Test (m-MTT), inspired by Allen et al. (Exp Theor Artif Intell 352:24–28, 2004) proposal, by asking people to distinguish real human moral evaluations from those made by a popular advanced AI language model: GPT-4. A representative sample of 299 U.S. adults first rated the quality of moral evaluations when blinded to their source. Remarkably, they rated the AI’s moral reasoning as superior in quality to humans’ along almost all dimensions, including virtuousness, intelligence, and trustworthiness, consistent with passing what Allen and colleagues call the comparative MTT. Next, when tasked with identifying the source of each evaluation (human or computer), people performed significantly above chance levels. Although the AI did not pass this test, this was not because of its inferior moral reasoning but, potentially, its perceived superiority, among other possible explanations. The emergence of language models capable of producing moral responses perceived as superior in quality to humans’ raises concerns that people may uncritically accept potentially harmful moral guidance from AI. This possibility highlights the need for safeguards around generative language models in matters of morality.

Here is my summary:

The researchers conducted a modified Moral Turing Test (m-MTT) to investigate if people view moral evaluations by advanced AI systems similarly to those by humans. They had participants rate the quality of moral reasoning from the AI language model GPT-4 and from humans, while initially blinded to the source.

Key Findings

• Remarkably, participants rated GPT-4's moral reasoning as superior in quality to humans' across dimensions like virtuousness, intelligence, and trustworthiness. This is consistent with passing the "comparative MTT" proposed previously.
• When later asked to identify if the moral evaluations came from a human or computer, participants performed above chance levels.
• However, GPT-4 did not definitively "pass" this test, potentially because its perceived superiority made it identifiable as AI.

Artificial Intelligence Feedback on Physician Notes Improves Patient Care

NYU Langone Health

Research, Innovation

Originally posted 17 APR 24

Artificial intelligence (AI) feedback improved the quality of physician notes written during patient visits, with better documentation improving the ability of care teams to make diagnoses and plan for patients’ future needs, a new study finds.

Since 2021, NYU Langone Health has been using pattern-recognizing, machine-learning AI systems to grade the quality of doctors’ clinical notes. At the same time, NYU Langone created data informatics dashboards that monitor hundreds of measures of safety and the effectiveness of care. The informatics team over time trained the AI models to track in dashboards how well doctors’ notes achieved the “5 Cs”: completeness, conciseness, contingency planning, correctness, and clinical assessment.

Now, a new case study, published online April 17 in NEJM Catalyst Innovations in Care Delivery, shows how notes improved by AI, in combination with dashboard innovations and other safety initiatives, resulted in an improvement in care quality across four major medical specialties: internal medicine, pediatrics, general surgery, and the intensive care unit.

This includes improvements across the specialties of up to 45 percent in note-based clinical assessments (that is, determining diagnoses) and reasoning (making predictions when diagnoses are unknown). In addition, contingency planning to address patients’ future needs saw improvements of up to 34 percent.

Last year, NYU Langone added to this long-standing effort a newer form of AI that develops likely options for the next word in any sentence based on how billions of people used language on the internet over time. A result of this next-word prediction is that generative AI chatbots like GPT-4 can read physician notes and make suggestions. In a pilot within the case study, the research team supercharged their machine-learning AI model, which can only give physicians a grade on their notes, by integrating a chatbot that added an accurate written narrative of issues with any note.

The info is here.

The article is linked above.  Here is the abstract:

Abstract

Electronic health records have become an integral part of modern health care, but their implementation has led to unintended consequences, such as poor note quality. This case study explores how NYU Langone Health leveraged artificial intelligence (AI) to address the challenge to improve the content and quality of medical documentation. By quickly and accurately analyzing large volumes of clinical documentation and providing feedback to organizational leadership and individually to providers, AI can help support a culture of continuous note quality improvement, allowing organizations to enhance a critical component of patient care.

A Doctor at Cigna Said Her Bosses Pressured Her to Review Patients’ Cases Too Quickly

P. Rucker and D. Armstrong

Propublica.org

Originally posted 29 APR 24

Here is an excerpt:

As ProPublica and The Capitol Forum reported last year, Cigna built a computer program that allowed its medical directors to deny certain claims in bulk. The insurer’s doctors spent an average of just 1.2 seconds on each of those cases. Cigna at the time said the review system was created to speed up approval of claims for certain routine screenings; the company later posted a rebuttal to the story. A congressional committee and the Department of Labor launched inquiries into this Cigna program. A spokesperson for Rep. Cathy McMorris Rodgers, the chair of the congressional committee, said Rodgers continues to monitor the situation after Cigna shared some details about its process. The Labor Department is still examining such practices.

One figure on Cigna’s January and February 2022 dashboards was like a productivity score; the news organizations found that this number reflects the pace at which a medical director clears cases.

Cigna said it was incorrect to call that figure on its dashboard a productivity score and said its “view on productivity is defined by a range of factors beyond elements included in a single spreadsheet.” In addition, the company told the news organizations, “The copy of the dashboard that you have is inaccurate and secondary calculations made using its contents may also be inaccurate.” The news organizations asked what was inaccurate, but the company wouldn’t elaborate.

Nevertheless, Cigna said that because the dashboard created “inadvertent confusion” the company was “reassessing its use.”

The info is here.

Here is my summary:

The article reports on Dr. Debby Day, who alleges that Cigna, her employer, pressured her to prioritize speed over thoroughness when reviewing patients' requests for healthcare coverage.

According to Day, managers emphasized meeting quotas and processing claims quickly, even if it meant superficially reviewing cases. Dr. Day said Cigna expected medical directors to review cases in as little as 4 minutes, which she felt was too rushed to properly evaluate them.  The pressure to deny claims quickly was nicknamed "click and close" by some employees.

Day felt this practice compromised patient care and refused to expedite reviews at the expense of quality. The article suggests this may have led to threats of termination from Cigna.

Large Language Models as Moral Experts? GPT-4o Outperforms Expert Ethicist in Providing Moral Guidance

Dillion, D., Mondal, D., Tandon, N.,

& Gray, K. (2024, May 29).

https://doi.org/10.31234/osf.io/w7236

Abstract

AI has demonstrated expertise across various fields, but its potential as a moral expert remains unclear. Recent work suggests that Large Language Models (LLMs) can reflect moral judgments with high accuracy. But as LLMs are increasingly used in complex decision-making roles, true moral expertise requires not just aligned judgments but also clear and trustworthy moral reasoning. Here, we advance work on the Moral Turing Test and find that advice from GPT-4o is rated as more moral, trustworthy, thoughtful, and correct than that of the popular The New York Times advice column, The Ethicist. GPT models outperformed both a representative sample of Americans and a renowned ethicist in providing moral explanations and advice, suggesting that LLMs have, in some respects, achieved a level of moral expertise. The present work highlights the importance of carefully programming ethical guidelines in LLMs, considering their potential to sway users' moral reasoning. More promisingly, it suggests that LLMs could complement human expertise in moral guidance and decision-making.

The research is here.

Here are my thoughts:

This research on GPT-4o's moral reasoning is fascinating, but caution is warranted. While exceeding human performance in explanations and perceived trustworthiness is impressive, true moral expertise goes beyond these initial results.

Here's why:

First, there are nuances to all moral dilemmas. Real-world dilemmas often lack clear-cut answers. Can GPT-4o navigate the gray areas and complexities of human experience?

Next, everyone has a rich experience, values, perspectives, and biases.  What ethical framework guides GPT-4o's decisions? Transparency in its programming is crucial.

Finally, the consequences of AI-driven moral advice can be far-reaching. Careful evaluation of potential biases and unintended outcomes is essential.  There is no objective algorithm.  There is no objective morality.  All moral decisions, no matter how well-reasoned, have pluses and minuses.  Therefore, AI can be used as a starting point for decision-making and planning.