Ethics of psychotherapy rationing: A review of ethical and regulatory documents in Canadian professional psychology

Gower, H. K., & Gaine, G. S. (2023).

Canadian Psychology / Psychologie canadienne. 

Advance online publication.

https://doi.org/10.1037/cap0000372

Abstract

Ethical and regulatory documents in Canadian professional psychology were reviewed for principles and standards related to the rationing of psychotherapy. Despite Canada’s high per capita health care expenses, mental health in Canada receives relatively low funding. Further, surveys indicated that Canadians have unmet needs for psychotherapy. Effective and ethical rationing of psychological treatment is a necessity, yet the topic of rationing in psychology has received scant attention. The present study involved a qualitative review of codes of ethics, codes of conduct, and standards of practice documents for their inclusion of rationing principles and standards. Findings highlight the strengths and shortcomings of these documents related to guiding psychotherapy rationing. The discussion offers recommendations for revising these ethical and regulatory documents to promote more equitable and cost-effective use of limited psychotherapy resources in Canada.

Impact Statement

Canadian professional psychology regulatory documents contain limited reference to rationing imperatives, despite scarce psychotherapy resources. While the foundation of distributive justice is in place, rationing-specific principles, standards, and practices are required to foster the fair and equitable distribution of psychotherapy by Canadian psychologists.

From the recommendations:

Recommendations for Canadian Psychology Regulatory Documents

1. Explicitly widen psychologists’ scope of concern to include not only current clients but also waiting clients and those who need treatment but face access barriers.
2. Acknowledge the scarcity of health care resources (in public and private settings) and the high demand for psychology services (e.g., psychotherapy) and admonish inefficient and cost-ineffective use.
3. Draw an explicit connection between the general principle of distributive justice and the specific practices related to rationing of psychology resources, including, especially, mitigation of biases likely to weaken ethical decision making.
4. Encourage the use of outcome monitoring measures to aid relative utility calculations for triage and termination decisions and to ensure efficiency and distributive justice.
5. Recommend advocacy by psychologists to address barriers to accessing needed services (e.g., psychotherapy), including promoting the cost effectiveness of psychotherapy as well as highlighting systemic barriers related to presenting problem, disability, ethnicity, race, gender, sexuality, or income.

Toward Parsimony in Bias Research: A Proposed Common Framework of Belief-Consistent Information Processing for a Set of Biases

Oeberst, A., & Imhoff, R. (2023).

Perspectives on Psychological Science, 0(0).

https://doi.org/10.1177/17456916221148147

Abstract

One of the essential insights from psychological research is that people’s information processing is often biased. By now, a number of different biases have been identified and empirically demonstrated. Unfortunately, however, these biases have often been examined in separate lines of research, thereby precluding the recognition of shared principles. Here we argue that several—so far mostly unrelated—biases (e.g., bias blind spot, hostile media bias, egocentric/ethnocentric bias, outcome bias) can be traced back to the combination of a fundamental prior belief and humans’ tendency toward belief-consistent information processing. What varies between different biases is essentially the specific belief that guides information processing. More importantly, we propose that different biases even share the same underlying belief and differ only in the specific outcome of information processing that is assessed (i.e., the dependent variable), thus tapping into different manifestations of the same latent information processing. In other words, we propose for discussion a model that suffices to explain several different biases. We thereby suggest a more parsimonious approach compared with current theoretical explanations of these biases. We also generate novel hypotheses that follow directly from the integrative nature of our perspective.

Here is my summary:

The authors argue that many different biases, such as the bias blind spot, hostile media bias, egocentric/ethnocentric bias, and outcome bias, can be traced back to the combination of a fundamental prior belief and humans' tendency toward belief-consistent information processing.

Belief-consistent information processing is the process of attending to, interpreting, and remembering information in a way that is consistent with one's existing beliefs. This process can lead to biases when it results in people ignoring or downplaying information that is inconsistent with their beliefs, and giving undue weight to information that is consistent with their beliefs.

The authors propose that different biases can be distinguished by the specific belief that guides information processing. For example, the bias blind spot is characterized by the belief that one is less biased than others, while hostile media bias is characterized by the belief that the media is biased against one's own group. However, the authors also argue that different biases may share the same underlying belief, and differ only in the specific outcome of information processing that is assessed. For example, both the bias blind spot and hostile media bias may involve the belief that one is more objective than others, but the bias blind spot is assessed in the context of self-evaluations, while hostile media bias is assessed in the context of evaluations of others.

The authors' framework has several advantages over existing theoretical explanations of biases. First, it provides a more parsimonious explanation for a wide range of biases. Second, it generates novel hypotheses that can be tested empirically. For example, the authors hypothesize that people who are more likely to believe in one bias will also be more likely to believe in other biases. Third, the framework has implications for interventions to reduce biases. For example, the authors suggest that interventions to reduce biases could focus on helping people to become more aware of their own biases and to develop strategies for resisting the tendency toward belief-consistent information processing.

AI Will—and Should—Change Medical School, Says Harvard’s Dean for Medical Education

Hswen Y, Abbasi J.

JAMA. Published online October 25, 2023.

doi:10.1001/jama.2023.19295

Here is an excerpt:

Dr Bibbins-Domingo: When these types of generative AI tools first came into prominence or awareness, educators, whatever level of education they were involved with, had to scramble because their students were using them. They were figuring out how to put up the right types of guardrails, set the right types of rules. Are there rules or danger zones right now that you’re thinking about?

Dr Chang: Absolutely, and I think there’s quite a number of these. This is a focus that we’re embarking on right now because as exciting as the future is and as much potential as these generative AI tools have, there are also dangers and there are also concerns that we have to address.

One of them is helping our students, who like all of us are still new to this within the past year, understand the limitations of these tools. Now these tools are going to get better year after year after year, but right now they are still prone to hallucinations, or basically making up facts that aren’t really true and yet saying them with confidence. Our students need to recognize why it is that these tools might come up with those hallucinations to try to learn how to recognize them and to basically be on guard for the fact that just because ChatGPT is giving you a very confident answer, it doesn’t mean it’s the right answer. And in medicine of course, that’s very, very important. And so that’s one—just the accuracy and the validity of the content that comes out.

As I wrote about in my Viewpoint, the way that these tools work is basically a very fancy form of autocomplete, right? It is essentially using a probabilistic prediction of what the next word is going to be. And so there’s no separate validity or confirmation of the factual material, and that’s something that we need to make sure that our students understand.

The other thing is to address the fact that these tools may inherently be structurally biased. Now, why would that be? Well, as we know, ChatGPT and these other large language models [LLMs] are trained on the world’s internet, so to speak, right? They’re trained on the noncopyrighted corpus of material that’s out there on the web. And to the extent that that corpus of material was generated by human beings who in their postings and their writings exhibit bias in one way or the other, whether intentionally or not, that’s the corpus on which these LLMs are trained. So it only makes sense that when we use these tools, these tools are going to potentially exhibit evidence of bias. And so we need our students to be very aware of that. As we have worked to reduce the effects of systematic bias in our curriculum and in our clinical sphere, we need to recognize that as we introduce this new tool, this will be another potential source of bias.

The info is here.

Here is my summary:

Bernard Chang, the Dean for Medical Education at Harvard Medical School, argues that artificial intelligence (AI) is poised to transform medical education. AI has the potential to improve the way medical students learn and train, and that medical schools should not only embrace AI, but also take an active role in shaping its development and use.

Chang identifies several areas where AI could have a significant impact on medical education. First, AI could be used to personalize learning and provide students with more targeted feedback. For example, AI-powered tutors could help students learn complex medical concepts at their own pace, and AI-powered diagnostic tools could help students practice their clinical skills.

Second, AI could be used to automate tasks that are currently performed by human instructors, such as grading exams and providing feedback on student assignments. This would free up instructors to focus on more high-value activities, such as mentoring students and leading discussions.

Third, AI could be used to create new educational experiences that are not possible with traditional methods. For example, AI could be used to create virtual patients that students can interact with to practice their clinical skills. AI could also be used to develop simulations of complex medical procedures that students can practice in a safe environment.

Chang argues that medical schools have a responsibility to prepare students for the future of medicine, which will be increasingly reliant on AI. He writes that medical schools should teach students how to use AI effectively, and how to critically evaluate AI-generated information. Medical schools should also develop new curricula that take into account the potential impact of AI on medical practice.

One strike and you’re a lout: Cherished values increase the stringency of moral character attributions

Rottman, J., Foster-Hanson, E., & Bellersen, S.

(2023). Cognition, 239, 105570.

https://doi.org/10.1016/j.cognition.2023.105570 

Abstract

Moral dilemmas are inescapable in daily life, and people must often choose between two desirable character traits, like being a diligent employee or being a devoted parent. These moral dilemmas arise because people hold competing moral values that sometimes conflict. Furthermore, people differ in which values they prioritize, so we do not always approve of how others resolve moral dilemmas. How are we to think of people who sacrifice one of our most cherished moral values for a value that we consider less important? The “Good True Self Hypothesis” predicts that we will reliably project our most strongly held moral values onto others, even after these people lapse. In other words, people who highly value generosity should consistently expect others to be generous, even after they act frugally in a particular instance. However, reasoning from an error-management perspective instead suggests the “Moral Stringency Hypothesis,” which predicts that we should be especially prone to discredit the moral character of people who deviate from our most deeply cherished moral ideals, given the potential costs of affiliating with people who do not reliably adhere to our core moral values. In other words, people who most highly value generosity should be quickest to stop considering others to be generous if they act frugally in a particular instance. Across two studies conducted on Prolific (N = 966), we found consistent evidence that people weight moral lapses more heavily when rating others’ membership in highly cherished moral categories, supporting the Moral Stringency Hypothesis. In Study 2, we examined a possible mechanism underlying this phenomenon. Although perceptions of hypocrisy played a role in moral updating, personal moral values and subsequent judgments of a person’s potential as a good cooperative partner provided the clearest explanation for changes in moral character attributions. Overall, the robust tendency toward moral stringency carries significant practical and theoretical implications.

The research is paywalled here.

My take aways: 

The results showed that participants were more likely to rate the person as having poor moral character when the transgression violated a cherished value. This suggests that when we see someone violate a value that we hold dear, it can lead us to question their entire moral compass.

The authors argue that this finding has important implications for how we think about moral judgment. They suggest that our own values play a significant role in how we judge others' moral character. This is something to keep in mind the next time we're tempted to judge someone harshly.

Here are some additional points that are made in the article:

• The effect of cherished values on moral judgment is stronger for people who are more strongly identified with their values.
• The effect is also stronger for transgressions that are seen as more serious.
• The effect is not limited to personal values. It can also occur for group-based values, such as patriotism or religious beliefs.

Patients need doctors who look like them. Can medicine diversify without affirmative action?

Kat Stafford

apnews.com

Originally posted 11 September 23

Here are two excerpts:

But more than two months after the Supreme Court struck down affirmative action in college admissions, concerns have arisen that a path into medicine may become much harder for students of color. Heightening the alarm: the medical field’s reckoning with longstanding health inequities.

Black Americans represent 13% of the U.S. population, yet just 6% of U.S. physicians are Black. Increasing representation among doctors is one solution experts believe could help disrupt health inequities.

The disparities stretch from birth to death, often beginning before Black babies take their first breath, a recent Associated Press series showed. Over and over, patients said their concerns were brushed aside or ignored, in part because of unchecked bias and racism within the medical system and a lack of representative care.

A UCLA study found the percentage of Black doctors had increased just 4% from 1900 to 2018.

But the affirmative action ruling dealt a “serious blow” to the medical field’s goals of improving that figure, the American Medical Association said, by prohibiting medical schools from considering race among many factors in admissions. The ruling, the AMA said, “will reverse gains made in the battle against health inequities.”

The consequences could affect Black health for generations to come, said Dr. Uché Blackstock, a New York emergency room physician and author of “LEGACY: A Black Physician Reckons with Racism in Medicine.”

(cut)

“As medical professionals, any time we see disparities in care or outcomes of any kind, we have to look at the systems in which we are delivering care and we have to look at ways that we are falling short,” Wysong said.

Without affirmative action as a tool, career programs focused on engaging people of color could grow in importance.

For instance, the Pathways initiative engages students from Black, Latino and Indigenous communities from high school through medical school.

The program starts with building interest in dermatology as a career and continues to scholarships, workshops and mentorship programs. The goal: Increase the number of underrepresented dermatology residents from about 100 in 2022 to 250 by 2027, and grow the share of dermatology faculty who are members of color by 2%.

Tolliver credits her success in becoming a dermatologist in part to a scholarship she received through Ohio State University’s Young Scholars Program, which helps talented, first-generation Ohio students with financial need. The scholarship helped pave the way for medical school, but her involvement in the Pathways residency program also was central.

The info is here.

The Best-Case Heuristic: 4 Studies of Relative Optimism, Best-Case, Worst-Case, & Realistic Predictions in Relationships, Politics, & a Pandemic

Sjåstad, H., & Van Bavel, J. (2023).

Personality and Social Psychology Bulletin, 0(0).

https://doi.org/10.1177/01461672231191360

Abstract

In four experiments covering three different life domains, participants made future predictions in what they considered the most realistic scenario, an optimistic best-case scenario, or a pessimistic worst-case scenario (N = 2,900 Americans). Consistent with a best-case heuristic, participants made “realistic” predictions that were much closer to their best-case scenario than to their worst-case scenario. We found the same best-case asymmetry in health-related predictions during the COVID-19 pandemic, for romantic relationships, and a future presidential election. In a fully between-subject design (Experiment 4), realistic and best-case predictions were practically identical, and they were naturally made faster than the worst-case predictions. At least in the current study domains, the findings suggest that people generate “realistic” predictions by leaning toward their best-case scenario and largely ignoring their worst-case scenario. Although political conservatism was correlated with lower covid-related risk perception and lower support of early public-health interventions, the best-case prediction heuristic was ideologically symmetric.

The article is paywalled here.

Here is my summary:

This research examined how people make predictions about the future in different life domains, such as health, relationships, and politics. The researchers found that people tend to make predictions that are closer to their best-case scenario than to their worst-case scenario, even when asked to make a "realistic" prediction. This is known as the best-case heuristic.

The researchers conducted four experiments to test the best-case heuristic. In the first experiment, participants were asked to make predictions about their risk of getting COVID-19, their satisfaction with their romantic relationship in one year, and the outcome of the next presidential election. Participants were asked to make three predictions for each event: a best-case scenario, a worst-case scenario, and a realistic scenario. The results showed that participants' "realistic" predictions were much closer to their best-case predictions than to their worst-case predictions.

The researchers found the same best-case asymmetry in the other three experiments, which covered a variety of life domains, including health, relationships, and politics. The findings suggest that people use a best-case heuristic when making predictions about the future, even in serious and important matters.

The best-case heuristic has several implications for individuals and society. On the one hand, it can help people to maintain a positive outlook on life and to cope with difficult challenges. On the other hand, it can also lead to unrealistic expectations and to a failure to plan for potential problems.

Overall, the research on the best-case heuristic suggests that people's predictions about the future are often biased towards optimism. This is something to be aware of when making important decisions and when planning for the future.

Humans’ Bias Blind Spot and Its Societal Significance

Pronin, E., & Hazel, L. (2023).

Current Directions in Psychological Science, 0(0).

https://doi.org/10.1177/09637214231178745

Abstract

Human beings have a bias blind spot. We see bias all around us but sometimes not in ourselves. This asymmetry hinders self-knowledge and fuels interpersonal misunderstanding and conflict. It is rooted in cognitive mechanics differentiating self- and social perception as well as in self-esteem motives. It generalizes across social, cognitive, and behavioral biases; begins in childhood; and appears across cultures. People show a bias blind spot in high-stakes contexts, including investing, medicine, human resources, and law. Strategies for addressing the problem are described.

(cut)

Bias-limiting procedures

When it comes to eliminating bias, attempts to overcome it via conscious effort and educational training are not ideal. A different strategy is worth considering, when possible: preventing people’s biases from having a chance to operate in the first place, by limiting their access to biasing information. Examples include conducting auditions behind a screen (discussed earlier) and blind review of journal submissions. If fully blocking access to potentially biasing information is not possible or carries more costs than benefits, another less stringent option is worth considering, that is, controlling when the information is presented so that potentially biasing information comes late, ideally after a tentative judgment is made (e.g., “sequential unmasking”; Dror, 2018; “temporary cloaking”; Kang, 2021).

Because of the BBS, people can be resistant to procedures like this that limit their access to biasing information (see Fig. 3). For example, forensics experts prefer consciously trying to avoid bias over being shielded from even irrelevant biasing information (Kukucka et al., 2017). When high school teachers and ensemble singers were asked to assess blinding procedures (in auditioning and grading), they opposed them more for their own group than for the other group and even more for themselves personally (Pronin et al., 2022). This opposition is consistent with experiments showing that people are unconcerned about the effects of biasing decision processes when it comes to their own decisions (Hansen et al., 2014). In those experiments, participants made judgments using a biasing decision procedure (e.g., judging the quality of paintings only after looking to see if someone famous painted them). They readily acknowledged that the procedure was biased, nonetheless made decisions that were biased by that procedure, and then insisted that their conclusions were objective. This unwarranted confidence is a barrier to the self-imposition of bias-reducing procedures. It suggests the need for adopting procedures like this at the policy level rather than counting on individuals or their organizations to do so.

A different bias-limiting procedure that may induce resistance for these same reasons, and that therefore may also benefit from institutional or policy-level implementation, involves precommitting to decision criteria (e.g., Norton et al., 2004; Uhlmann & Cohen, 2005). For example, the human resources officer who precommits to judging job applicants more on the basis of industry experience versus educational background cannot then change that emphasis after seeing that their favorite candidate has unusually impressive academic credentials. This logic is incorporated, for example, into the system of allocating donor organs in the United States, which has explicit and predetermined criteria for making those allocations in order to avoid the possibility of bias in this high-stakes arena. When decision makers are instructed to provide objective criteria for their decision not before making that decision but rather when providing it—that is, the more typical request made of them—this not only makes bias more likely but also, because of the BBS, may even leave decision makers more confident in their objectivity than if they had not been asked to provide those criteria at all.

Here's my brief summary:

The article discusses the concept of the bias blind spot, which refers to people's tendency to recognize bias in others more readily than in themselves. Studies have consistently shown that people rate themselves as less susceptible to various biases than the average person. The bias blind spot occurs even for well-known biases that people readily accept exist. This blind spot has important societal implications, as it impedes recognition of one's own biases. It also leads to assuming others are more biased than oneself, resulting in decreased trust. Overcoming the bias blind spot is challenging but important for issues from prejudice to politics. It requires actively considering one's own potential biases when making evaluations about oneself or others.

I Have a Question for the Famous People Who Have Tried to Apologize

Elizabeth Spiers

The New York Times - Guest Opinion

Originally posted 22 September 23

Here is an excerpt:

As a talk show host, Ms. Barrymore has been lauded in part for her empathy. She is vulnerable, and that makes her guests feel like they can be, too. But even nice people can be self-centered when they’re on the defensive. That’s what happened when people objected to the news that her show would return to production despite the writers’ strike. In a teary, rambling video on Instagram, which was later deleted, she spoke about how hard the situation had been — for her. “I didn’t want to hide behind people. So I won’t. I won’t polish this with bells and whistles and publicists and corporate rhetoric. I’ll just stand out there and accept and be responsible.” (Ms. Barrymore’s awkward, jumbled sentences unwittingly demonstrated how dearly she needs those writers.) Finally, she included a staple of the public figure apology genre: “My intentions have never been in a place to upset or hurt anyone,” she said. “It’s not who I am.”

“This is not who I am” is a frequent refrain from people who are worried that they’re going to be defined by their worst moments. It’s an understandable concern, given the human tendency to pay more attention to negative events. People are always more than the worst thing they’ve done. But it’s also true that the worst things they’ve done are part of who they are.

Somehow, Mila Kunis’s scripted apology was even worse. She and Mr. Kutcher had weathered criticism for writing letters in support of their former “That ’70s Show” co-star Danny Masterson after he was convicted of rape. Facing her public, she spoke in the awkward cadence people have when they haven’t memorized their lines and don’t know where the emphasis should fall. “The letters were not written to question the legitimacy” — pause — “of the judicial system,” she said, “or the validity” — pause — “of the jury’s ruling.” For an actress, it was not a very convincing performance. Mr. Kutcher, who is her husband, was less awkward in his delivery, but his defense was no more convincing. The letters, he explained, were only “intended for the judge to read,” as if the fact that the couple operated behind the scenes made it OK.

The article is paywalled here.

Here are my observations about the main theme of this article:

Miller argues that many celebrity apologies fall short because they are not sincere. She says that they often lack the essential elements of a good apology: acknowledging the offense, providing an explanation, expressing remorse, and making amends. Instead, many celebrity apologies are self-serving and aimed at salvaging their public image.

Miller concludes by saying that if celebrities want their apologies to be meaningful, they need to be honest, take responsibility for their actions, and show that they are truly sorry for the harm they have caused.

I would also add that celebrity apologies can be difficult to believe because they often follow a predictable pattern. The celebrity typically issues a statement expressing their regret and apologizing to the people they have hurt. They may also offer a brief explanation for their behavior, but they often avoid taking full responsibility for their actions. And while some celebrities may make amends in some way, such as donating to charity or volunteering their time, many do not.

As a result, many people are skeptical of celebrity apologies. They see them as nothing more than a way for celebrities to save face and get back to their normal lives. This is why it is so important for celebrities to be sincere and genuine when they apologize.

The Young Conservatives Trying to Make Eugenics Respectable Again

Adam Serwer

The Atlantic

Originally posted 15 September 23

Here are two excerpts:

One explanation for the resurgence of scientific racism—what the psychologist Andrew S. Winston defines as the use of data to promote the idea of an “enduring racial hierarchy”—is that some very rich people are underwriting it. Mathias notes that “rich benefactors, some of whose identities are unknown, have funneled hundreds of thousands of dollars into a think tank run by Hanania.” As the biological anthropologist Jonathan Marks tells the science reporter Angela Saini in her book Superior, “There are powerful forces on the right that fund research into studying human differences with the goal of establishing those differences as a basis of inequalities.”

There is no great mystery as to why eugenics has exerted such a magnetic attraction on the wealthy. From god emperors, through the divine right of kings, to social Darwinism, the rich have always sought an uncontestable explanation for why they have so much more money and power than everyone else. In a modern, relatively secular nation whose inequalities of race and class have been shaped by slavery and its legacies, the justifications tend toward the pseudoscience of an unalterable genetic aristocracy with white people at the top and Black people at the bottom.

“The lay concept of race does not correspond to the variation that exists in nature,” the geneticist Joseph L. Graves wrote in The Emperor’s New Clothes: Biological Theories of Race at the Millennium. “Instead, the American concept of race is a social construction, resulting from the unique political and cultural history of the United States.”

Because race is a social reality, genuine disparities among ethnic groups persist in measures such as education and wealth. Contemporary believers in racial pseudoscience insist these disparities must necessarily have a genetic explanation, one that happens to correspond to shifting folk categories of race solidified in the 18th century to justify colonialism and enslavement. They point to the external effects of things like war, poverty, public policy, and discrimination and present them as caused by genetics. For people who have internalized the logic of race, the argument may seem intuitive. But it is just astrology for racists.

(cut)

Race is a sociopolitical category, not a biological one. There is no genetic support for the idea that humans are divided into distinct races with immutable traits shared by others who have the same skin color. Although qualified geneticists have debunked the shoddy arguments of race scientists over and over, the latter maintain their relevance in part by casting substantive objections to their assumptions, methods, and conclusions as liberal censorship. There are few more foolproof ways to get Trump-era conservatives to believe falsehoods than to insist that liberals are suppressing them. Race scientists also understand that most people can evaluate neither the pseudoscience they offer as proof of racial differences nor the actual science that refutes it, and will default to their political sympathies.

Three political developments helped renew this pseudoscience’s appeal. The first was the election of Barack Obama, an emotional blow to those adhering to the concept of racial hierarchy from which they have yet to recover. Then came the rise of Bernie Sanders, whose left-wing populism blamed the greed of the ultra-wealthy for the economic struggles of both the American working class and everyone in between. Both men—one a symbol of racial equality, the other of economic justice—drew broad support within the increasingly liberal white-collar workforce from which the phrenologist billionaires of Big Tech draw their employees. The third was the election of Donald Trump, itself a reaction to Obama and an inspiration to those dreaming of a world where overt bigotry does not carry social consequences.

The info is here.

Here is my brief synopsis:

Young conservatives are often influenced by far-right ideologues who believe in the superiority of the white race and the need to improve the human gene pool.  Serwer argues that the resurgence of interest in eugenics is part of a broader trend on the right towards embracing racist and white supremacist ideas. He also notes that the pseudoscience of race is being used to justify hierarchies and provide an enemy to rail against.

It is important to note that eugenics is a dangerous and discredited ideology. It has been used to justify forced sterilization, genocide, and other atrocities. The resurgence of interest in eugenics is a threat to all people, especially those who are already marginalized and disadvantaged.

The Myth of the Secret Genius

Brian Klaas

The Garden of Forking Path

Originally posted 30 Nov 22

Here are two excepts: 

A recent research study, involving a collaboration between physicists who model complex systems and an economist, however, has revealed why billionaires are so often mediocre people masquerading as geniuses. Using computer modelling, they developed a fake society in which there is a realistic distribution of talent among competing agents in the simulation. They then applied some pretty simple rules for their model: talent helps, but luck also plays a role.

Then, they tried to see what would happen if they ran and re-ran the simulation over and over.

What did they find? The most talented people in society almost never became extremely rich. As they put it, “the most successful individuals are not the most talented ones and, on the other hand, the most talented individuals are not the most successful ones.”

Why? The answer is simple. If you’ve got a society of, say, 8 billion people, there are literally billions of humans who are in the middle distribution of talent, the largest area of the Bell curve. That means that in a world that is partly defined by random chance, or luck, the odds that someone from the middle levels of talent will end up as the richest person in the society are extremely high.

Look at this first plot, in which the researchers show capital/success (being rich) on the vertical/Y-axis, and talent on the horizontal/X-axis. What’s clear is that society’s richest person is only marginally more talented than average, and there are a lot of people who are extremely talented that are not rich.

Then, they tried to figure out why this was happening. In their simulated world, lucky and unlucky events would affect agents every so often, in a largely random pattern. When they measured the frequency of luck or misfortune for any individual in the simulation, and then plotted it against becoming rich or poor, they found a strong relationship.

(cut)

The authors conclude by stating “Our results highlight the risks of the paradigm that we call “naive meritocracy", which fails to give honors and rewards to the most competent people, because it underestimates the role of randomness among the determinants of success.”

Indeed.

The article is here.

Here is my summary:

The myth of the secret genius: The belief that some people are just born with natural talent and that there is nothing we can do to achieve the same level of success.

The importance of hard work: The vast majority of successful people are not geniuses. They are simply people who have worked hard and persevered in the face of setbacks.

The power of luck: Luck plays a role in everyone's success. Some people are luckier than others, and most people do not factor in luck, as well as other external variables, into their assessment.  This bias is another form of the Fundamental Attribution Error.

The importance of networks: Our networks play a big role in our success. We need to be proactive in building relationships with people who can help us achieve our goals.

Ontario court rules against Jordan Peterson, upholds social media training order

Canadian Broadcasting Company

Originally posted 23 August 23

An Ontario court ruled against psychologist and media personality Jordan Peterson Wednesday, and upheld a regulatory body's order that he take social media training in the wake of complaints about his controversial online posts and statements.

Last November, Peterson, a professor emeritus with the University of Toronto psychology department who is also an author and media commentator, was ordered by the College of Psychologists of Ontario to undergo a coaching program on professionalism in public statements.

That followed numerous complaints to the governing body of Ontario psychologists, of which Peterson is a member, regarding his online commentary directed at politicians, a plus-sized model, and transgender actor Elliot Page, among other issues. You can read more about those social media posts here.

The college's complaints committee concluded his controversial public statements could amount to professional misconduct and ordered Peterson to pay for a media coaching program — noting failure to comply could mean the loss of his licence to practice psychology in the province.

Peterson filed for a judicial review, arguing his political commentary is not under the college's purview.

Three Ontario Divisional Court judges unanimously dismissed Peterson's application, ruling that the college's decision falls within its mandate to regulate the profession in the public interest and does not affect his freedom of expression.

"The order is not disciplinary and does not prevent Dr. Peterson from expressing himself on controversial topics; it has a minimal impact on his right to freedom of expression," the decision written by Justice Paul Schabas reads, in part.

The info is here.

The court decision is here.

My take:

Peterson has argued that the order violates his right to free speech. He has also said that the complaints against him were politically motivated. However, the court ruled that the college's order was justified in order to protect the public from harm.

The case of Jordan Peterson is a reminder that psychologists, like other human beings, are not infallible. They are capable of making mistakes and of expressing harmful views. It is important to hold psychologists accountable for their actions, and to ensure that they are held to the highest ethical standards.

In addition to the steps outlined above, there are a number of other things that can be done to mitigate bias in psychology. These include:

• Increasing diversity in the field of psychology
• Promoting critical thinking and self-reflection among psychologists
• Developing more specific ethical guidelines for psychologists' use of social media
• Holding psychologists accountable for their online behavior

Can Confirmation Bias Improve Group Learning?

Gabriel, N. and O'Connor, C. (2022)

[Preprint]

Abstract

Confirmation bias has been widely studied for its role in failures of reasoning. Individuals exhibiting confirmation bias fail to engage with information that contradicts their current beliefs, and, as a result, can fail to abandon inaccurate beliefs. But although most investigations of confirmation bias focus on individual learning, human knowledge is typically developed within a social structure. How does the presence of confirmation bias influence learning and the development of consensus within a group? In this paper, we use network models to study this question. We find, perhaps surprisingly, that moderate confirmation bias often improves group learning. This is because confirmation bias leads the group to entertain a wider variety of theories for a longer time, and prevents them from prematurely settling on a suboptimal theory. There is a downside, however, which is that a stronger form of confirmation bias can cause persistent polarization, and hurt the knowledge producing capacity of the community. We discuss implications of these results for epistemic communities, including scientific ones.

Conclusion

We find that confirmation bias, in a more moderate form, improves the epistemic performance of agents in a networked community. This is perhaps surprising given that previous work mostly emphasizes the epistemic harms of confirmation bias. By decreasing the chances that a group pre-emptively settles on a

promising theory or option, confirmation bias can improve the likelihood that the group chooses optimal options in the long run. In this, it can play a similar role to decreased network connectivity or stubbornness (Zollman, 2007, 2010; Wu, 2021). The downside is that more robust confirmation bias, where agents entirely ignore data that is too disconsonant with their current beliefs, can lead to polarization, and harm the epistemic success of a community. Our modeling results thus provide potential support for the arguments of Mercier & Sperber (2017) regarding the benefits of confirmation bias to a group, but also a caution.  Too much confirmation bias does not provide such benefits.

There are several ongoing discussions in philosophy and the social sciences where these results are relevant. Mayo-Wilson et al. (2011) use network models to argue for the independence thesis—that rationality of individual agents and rationality of the groups they form sometimes come apart. I.e., individually rational agents may form groups which are not ideally rational, and rational groups may sometimes consist in individually irrational agents. Our results lend support to this claim. While there is a great deal of evidence suggesting that confirmation bias is not ideal for individual reasoners, our results suggest that it can nonetheless improve group reasoning under the right conditions.

The research is here.

The authors conclude that confirmation bias can have both positive and negative effects on group learning. The key is to find a moderate level of confirmation bias that allows the group to explore a variety of theories without becoming too polarized.

Here are some of the key findings of the paper:

• Moderate confirmation bias can improve group learning by preventing the group from prematurely settling on a suboptimal theory.
• Too much confirmation bias can lead to polarization and a decrease in the group's ability to learn.
• The key to effective group learning is to find a moderate level of confirmation bias.

Evidence for Anchoring Bias During Physician Decision-Making

Ly, D. P., Shekelle, P. G., & Song, Z. (2023).

JAMA Internal Medicine, 183(8), 818.

https://doi.org/10.1001/jamainternmed.2023.2366

Abstract

Introduction

Cognitive biases are hypothesized to influence physician decision-making, but large-scale evidence consistent with their influence is limited. One such bias is anchoring bias, or the focus on a single—often initial—piece of information when making clinical decisions without sufficiently adjusting to later information.

Objective

To examine whether physicians were less likely to test patients with congestive heart failure (CHF) presenting to the emergency department (ED) with shortness of breath (SOB) for pulmonary embolism (PE) when the patient visit reason section, documented in triage before physicians see the patient, mentioned CHF.

Design, Setting, and Participants

In this cross-sectional study of 2011 to 2018 national Veterans Affairs data, patients with CHF presenting with SOB in Veterans Affairs EDs were included in the analysis. Analyses were performed from July 2019 to January 2023.

Conclusions and Relevance

In this cross-sectional study among patients with CHF presenting with SOB, physicians were less likely to test for PE when the patient visit reason that was documented before they saw the patient mentioned CHF. Physicians may anchor on such initial information in decision-making, which in this case was associated with delayed workup and diagnosis of PE.

Here is the conclusion of the paper:

In conclusion, among patients with CHF presenting to the ED with SOB, we find that ED physicians were less likely to test for PE when the initial reason for visit, documented before the physician's evaluation, specifically mentioned CHF. These results are consistent with physicians anchoring on initial information. Presenting physicians with the patient’s general signs and symptoms, rather than specific diagnoses, may mitigate this anchoring. Other interventions include refining knowledge of findings that distinguish between alternative diagnoses for a particular clinical presentation.

Quick snapshot:

Anchoring bias is a cognitive bias that causes us to rely too heavily on the first piece of information we receive when making a decision. This can lead us to make inaccurate or suboptimal decisions, especially when the initial information is not accurate or relevant.

The findings of this study suggest that anchoring bias may be a significant factor in physician decision-making. This could lead to delayed or missed diagnoses, which could have serious consequences for patients.

A Federal Judge Asks: Does the Supreme Court Realize How Bad It Smells?

Michael Ponsor

The New York Times: Opinion

Originally posted 14 July 23

What has gone wrong with the Supreme Court’s sense of smell?

I joined the federal bench in 1984, some years before any of the justices currently on the Supreme Court. Throughout my career, I have been bound and guided by a written code of conduct, backed by a committee of colleagues I can call on for advice. In fact, I checked with a member of that committee before writing this essay.

A few times in my nearly 40 years on the bench, complaints have been filed against me. This is not uncommon for a federal judge. So far, none have been found to have merit, but all of these complaints have been processed with respect, and I have paid close attention to them.

The Supreme Court has avoided imposing a formal ethical apparatus on itself like the one that applies to all other federal judges. I understand the general concern, in part. A complaint mechanism could become a political tool to paralyze the court or a playground for gadflies. However, a skillfully drafted code could overcome this problem. Even a nonenforceable code that the justices formally pledged to respect would be an improvement on the current void.

Reasonable people may disagree on this. The more important, uncontroversial point is that if there will not be formal ethical constraints on our Supreme Court — or even if there will be — its justices must have functioning noses. They must keep themselves far from any conduct with a dubious aroma, even if it may not breach a formal rule.

The fact is, when you become a judge, stuff happens. Many years ago, as a fairly new federal magistrate judge, I was chatting about our kids with a local attorney I knew only slightly. As our conversation unfolded, he mentioned that he’d been planning to take his 10-year-old to a Red Sox game that weekend but their plan had fallen through. Would I like to use his tickets?

The information is here.

Cheap promises: Evidence from loan repayment pledges in an online experiment

Bhanot, S. P. (2017).

Journal of Economic Behavior & 

Organization, 142, 250-261.

https://doi.org/10.1016/j.jebo.2017.03.003

Abstract

Across domains, people struggle to follow through on their commitments. This can happen for many reasons, including dishonesty, forgetfulness, or insufficient intrinsic motivation. Social scientists have explored the reasons for persistent failures to follow through, suggesting that eliciting explicit promises can be an effective way to motivate action. This paper presents a field experiment that tests the effect of explicit promises, in the form of “honor pledges,” on loan repayment rates. The experiment was conducted with LendUp, an online lender, and targeted 4,883 first-time borrowers with the firm. Individuals were randomized into four groups, with the following experimental treatments: (1) having no honor pledge to complete (control); (2) signing a given honor pledge; (3) re-typing the same honor pledge as in (2) before signing; and (4) coming up with a personal honor pledge to type and sign. I also randomized whether or not borrowers were reminded of the honor pledge they signed prior to the repayment deadline. The results suggest that the honor pledge treatments had minimal impacts on repayment, and that reminders of the pledges were similarly ineffective. This suggests that borrowers who fail to repay loans do so not because of dishonesty or behavioral biases, but because they suffer from true financial hardship and are simply unable to repay.

Discussion

Literature in experimental economics and psychology often finds impacts of promises and explicit honor pledges on behavior, and in particular on reducing dishonest behavior. However, the results of this field experiment suggest no meaningful effects from an explicit promise (and indeed, a salient promise) on loan repayment behavior in a real-world setting, with money at stake. Furthermore, a self-written honor pledge was no more efficacious than any other, and altering the salience of the honor pledge, both at loan initiation and in reminder emails, had negligible impacts on outcomes. In other words, I find no evidence for the hypotheses that salience, reminders, or personalization strengthen the impact of a promise on behavior.  Indeed, the results of the study suggest that online loan repayment is a domain where such behavioral tools do not have an impact on decisions. This is a significant result, because it provides insights into why borrowers might fail to repay loans; most notably, it suggests that the failure to repay short-term loans may not be a question of dishonest behavior or behavioral biases, but rather an indication of true financial hardship. Simply put, when repayment is not financially possible, framing, reminders, or other interventions utilizing behavioral science are of limited use.

A pdf of the research is here.

The persistence of cognitive biases in financial decisions across economic groups

Ruggeri, K., Ashcroft-Jones, S. et al.

Sci Rep 13, 10329 (2023).

https://doi.org/10.1038/s41598-023-36339-2

Abstract

While economic inequality continues to rise within countries, efforts to address it have been largely ineffective, particularly those involving behavioral approaches. It is often implied but not tested that choice patterns among low-income individuals may be a factor impeding behavioral interventions aimed at improving upward economic mobility. To test this, we assessed rates of ten cognitive biases across nearly 5000 participants from 27 countries. Our analyses were primarily focused on 1458 individuals that were either low-income adults or individuals who grew up in disadvantaged households but had above-average financial well-being as adults, known as positive deviants. Using discrete and complex models, we find evidence of no differences within or between groups or countries. We therefore conclude that choices impeded by cognitive biases alone cannot explain why some individuals do not experience upward economic mobility. Policies must combine both behavioral and structural interventions to improve financial well-being across populations.

From the Discussion section

This study aimed to determine if rates of cognitive biases were different between positive deviants and low-income adults in a way that might explain some elements of what impedes or facilitates upward economic mobility. We anticipated finding small-to-moderate effects between groups indicating positive deviants were less prone to biases involving risk and uncertainty in financial choices. However, across a sample of nearly 5000 participants from 27 countries, of which 1458 were low-income or positive deviants, we find no evidence of any difference in the rates of cognitive biases—minor or otherwise—and no systematic variability to indicate patterns vary globally.

In sum, we find clear evidence that resistance to cognitive biases is not a factor contributing to or impeding upward economic mobility in our sample. Taken along with related work showing that temporal choice anomalies are tied more to economic environment rather than individual financial circumstances, our findings are (unintentionally) a major validation of arguments (especially that of Bertrand, Mullainathan, and Shafir) stating that poorer individuals are not uniquely prone to cognitive biases that alone explain protracted poverty. It also supports arguments that scarcity is a greater driver of decisions, as individuals of different income groups are equally influenced by biases and context-driven cues.

What makes these findings particularly reliable is that multiple possible approaches to analyses had to be considered while working with the data, some of which were considered into extreme detail before selecting the optimal approach. As our measures were effective at eliciting biases on a scale to be expected based on existing research, and as there were relatively low correlations between individual biases (e.g., observing loss aversion in one participant is not necessarily a strong predictor of also observing any other specific bias), we conclude that there is no evidence from our sample to support that biases are directly associated with potentially harming optimal choices uniquely amongst low-income individuals.

Conclusion

We sought to determine if individuals that had overcome low-income childhoods showed significantly different rates of cognitive biases from individuals that remained low-income as adults. We comprehensively reject our initial hypotheses and conclude that outcomes are not tied—at least not exclusively or potentially even meaningfully—to resistance to cognitive biases. Our research does not reject the notion that individual behavior and decision-making may directly relate to upward economic mobility. Instead, we narrowly conclude that biased decision-making does not alone explain a significant proportion of population-level economic inequality. Thus, any attempts to reduce economic inequality must involve both behavioral and structural aspects. Otherwise, similar decisions between disadvantaged individuals may not lead to similar outcomes. Where combined effectively, it will be possible to assess if genuine impact has been made on the financial well-being of individuals and populations.

How to Use AI Ethically for Ethical Decision-Making

Demaree-Cotton, J., Earp, B. D., & Savulescu, J.

(2022). American Journal of Bioethics, 22(7), 1–3.

https://doi.org/10.1080/15265161.2022.2075968

Here is an excerpt:

The  kind  of AI  proposed  by  Meier  and  colleagues  (2022) has  the  fascinating  potential to improve the transparency of ethical decision-making, at least if it is used as a decision aid rather than a decision replacement (Savulescu & Maslen 2015). While artificial intelligence cannot itself engage in the human communicative process of justifying its decisions to patients, the AI they describe (unlike “black-box” AI) makes explicit which values and principles are involved and how much weight they are given.

By contrast, the moral principles or values underlying human moral intuition are not always consciously, introspectively accessible (Cushman, Young, and  Hauser  2006).  While humans sometimes have a fuzzy, intuitive sense of some of the factors that are relevant to their moral judgment, we often have strong moral intuitions without being sure of their source,  or  with- out being clear on precisely how strongly different  factors  played  a  role in  generating  the intuitions.  But  if  clinicians  make use  of  the AI  as a  decision  aid, this  could help  them  to transparently and precisely communicate the actual reasons behind their decision.

This is so even if the AI’s recommendation is ultimately rejected. Suppose, for example, that the AI recommends a course of action, with a certain amount of confidence, and it specifies the exact values or  weights it has assigned  to  autonomy versus  beneficence  in coming  to  this conclusion. Evaluating the recommendation made by the AI could help a committee make more explicit the “black box” aspects  of their own reasoning.  For example, the committee might decide that beneficence should actually be  weighted more heavily in this case than the AI suggests. Being able to understand the reason that their decision diverges from that of the AI gives them the opportunity to offer a further justifying reason as to why they think beneficence should be given more weight;  and  this,  in  turn, could improve the  transparency of their recommendation. 

However, the potential for the kind of AI described in the target article to improve the accuracy of moral decision-making may be more limited. This is so for two reasons. Firstly, whether AI can be expected to outperform human decision-making depends in part on the metrics used to train it. In non-ethical domains, superior accuracy can be achieved because the “verdicts” given to the AI in the training phase are not solely the human judgments that the AI is intended to replace or inform. Consider how AI can learn to detect lung cancer from scans at a superior rate to human radiologists after being trained on large datasets and being “told” which scans show cancer and which ones are cancer-free. Importantly, this training includes cases where radiologists  did  not  recognize  cancer  in  the  early  scans  themselves,  but  where  further information verified the correct diagnosis later on (Ardila et al. 2019). Consequently, these AIs are  able  to  detect  patterns  even in  early  scans  that  are  not  apparent  or  easily detectable  by human radiologists, leading to superior accuracy compared to human performance.  

A pdf of the paper is here.

Generative AI companies must publish transparency reports

A. Narayanan and S. Kapoor

Knight First Amendment Institute

Originally published 26 June 23

Here is an excerpt:

Transparency reports must cover all three types of harms from AI-generated content

There are three main types of harms that may result from model outputs.

First, generative AI tools could be used to harm others, such as by creating non-consensual deepfakes or child sexual exploitation materials. Developers do have policies that prohibit such uses. For example, OpenAI's policies prohibit a long list of uses, including the use of its models to generate unauthorized legal, financial, or medical advice for others. But these policies cannot have real-world impact if they are not enforced, and due to platforms' lack of transparency about enforcement, we have no idea if they are effective. Similar challenges in ensuring platform accountability have also plagued social media in the past; for instance, ProPublica reporters repeatedly found that Facebook failed to fully remove discriminatory ads from its platform despite claiming to have done so.

Sophisticated bad actors might use open-source tools to generate content that harms others, so enforcing use policies can never be a comprehensive solution. In a recent essay, we argued that disinformation is best addressed by focusing on its distribution (e.g., on social media) rather than its generation. Still, some actors will use tools hosted in the cloud either due to convenience or because the most capable models don’t tend to be open-source. For these reasons, transparency is important for cloud-based generative AI.

Second, users may over-rely on AI for factual information, such as legal, financial, or medical advice. Sometimes they are simply unaware of the tendency of current chatbots to frequently generate incorrect information. For example, a user might ask "what are the divorce laws in my state?" and not know that the answer is unreliable. Alternatively, the user might be harmed because they weren’t careful enough to verify the generated information, despite knowing that it might be inaccurate. Research on automation bias shows that people tend to over-rely on automated tools in many scenarios, sometimes making more errors than when not using the tool.

ChatGPT includes a disclaimer that it sometimes generates inaccurate information. But OpenAI has often touted its performance on medical and legal exams. And importantly, the tool is often genuinely useful at medical diagnosis or legal guidance. So, regardless of whether it’s a good idea to do so, people are in fact using it for these purposes. That makes harm reduction important, and transparency is an important first step.

Third, generated content could be intrinsically undesirable. Unlike the previous types, here the harms arise not because of users' malice, carelessness, or lack of awareness of limitations. Rather, intrinsically problematic content is generated even though it wasn’t requested. For example, Lensa's avatar creation app generated sexualized images and nudes when women uploaded their selfies. Defamation is also intrinsically harmful rather than a matter of user responsibility. It is no comfort to the target of defamation to say that the problem would be solved if every user who might encounter a false claim about them were to exercise care to verify it.

The info is here.

Quick summary: 

The call for transparency reports aims to increase accountability and understanding of the inner workings of generative AI models. By disclosing information about the data used to train the models, the companies can address concerns regarding potential biases and ensure the ethical use of their technology.

Transparency reports could include details about the sources and types of data used, the demographics represented in the training data, any data augmentation techniques applied, and potential biases detected or addressed during model development. This information would enable users, policymakers, and researchers to evaluate the capabilities and limitations of the generative AI systems.

Gender-Affirming Care for Cisgender People

Theodore E. Schall and Jacob D. Moses

Hastings Center Report 53, no. 3 (2023): 15-24.

DOI: 10.1002/hast.1486 

Abstract

Gender-affirming care is almost exclusively discussed in connection with transgender medicine. However, this article argues that such care predominates among cisgender patients, people whose gender identity matches their sex assigned at birth. To advance this argument, we trace historical shifts in transgender medicine since the 1950s to identify central components of "gender-affirming care" that distinguish it from previous therapeutic models, such as "sex reassignment." Next, we sketch two historical cases-reconstructive mammoplasty and testicular implants-to show how cisgender patients offered justifications grounded in authenticity and gender affirmation that closely mirror rationales supporting gender-affirming care for transgender people. The comparison exposes significant disparities in contemporary health policy regarding care for cis and trans patients. We consider two possible objections to the analogy we draw, but ultimately argue that these disparities are rooted in "trans exceptionalism" that produces demonstrable harm.

The info is paywalled here.

Here is my summary:

The authors cite several examples of gender-affirming care for cisgender people, such as breast reconstruction following mastectomy, penile implants following testicular cancer, hormone replacement therapy, and hair removal. They argue that these interventions can be just as important for cisgender people's mental and physical health as they are for transgender people.

The authors also note that gender-affirming care for cisgender people is often less scrutinized and less stigmatized than such care for transgender people. Cisgender people do not need special letters of permission from mental health providers to access care whose primary purpose is to affirm their gender identity. And insurance companies are less likely to exclude gender-affirming care for cisgender people from their coverage.

The authors argue that the differences in the conceptualization and treatment of gender-affirming care for cisgender and transgender people reflect broad anti-trans bias in society and health care. They call for a more inclusive view of gender-affirming care that recognizes the needs of all people, regardless of their gender identity.

Final thoughts:

1. Gender-affirming care can be lifesaving. It can help reduce anxiety, depression, and suicidal thoughts.  Gender-affirming care can be framed as suicide prevention.
2. Gender-affirming care is not experimental. It has been studied extensively and is safe and effective. See other posts on this site for more comprehensive examples.
3. All people deserve access to gender-affirming care, regardless of their gender identity. This is basic equality and fairness in terms of access to medical care.

Inducing anxiety in large language models increases exploration and bias

Coda-Forno, J., Witte, K., et al. (2023).

arXiv preprint arXiv:2304.11111.

Abstract

Large language models are transforming research on machine learning while galvanizing public debates. Understanding not only when these models work well and succeed but also why they fail and misbehave is of great societal relevance. We propose to turn the lens of computational psychiatry, a framework used to computationally describe and modify aberrant behavior, to the outputs produced by these models. We focus on the Generative Pre-Trained Transformer 3.5 and subject it to tasks commonly studied in psychiatry. Our results show that GPT-3.5 responds robustly to a common anxiety questionnaire, producing higher anxiety scores than human subjects. Moreover, GPT-3.5's responses can be predictably changed by using emotion-inducing prompts. Emotion-induction not only influences GPT-3.5's behavior in a cognitive task measuring exploratory decision-making but also influences its behavior in a previously-established task measuring biases such as racism and ableism. Crucially, GPT-3.5 shows a strong increase in biases when prompted with anxiety-inducing text. Thus, it is likely that how prompts are communicated to large language models has a strong influence on their behavior in applied settings. These results progress our understanding of prompt engineering and demonstrate the usefulness of methods taken from computational psychiatry for studying the capable algorithms to which we increasingly delegate authority and autonomy.

From the Discussion section

What do we make of these results? It seems like GPT-3.5 generally performs best in the neutral condition, so a clear recommendation for prompt-engineering is to try and describe a problem as factually and neutrally as possible. However, if one does use emotive language, then our results show that anxiety-inducing scenarios lead to worse performance and substantially more biases. Of course, the neutral conditions asked GPT-3.5 to talk about something it knows, thereby possibly already contextualizing the prompts further in tasks that require knowledge and measure performance. However, that anxiety-inducing prompts can lead to more biased outputs could have huge consequences in applied scenarios. Large language models are, for example, already used in clinical settings and other high-stake contexts. If they produce higher biases in situations when a user speaks more anxiously, then their outputs could actually become dangerous. We have shown one method, which is to run psychiatric studies, that could capture and prevent such biases before they occur.

In the current work, we intended to show the utility of using computational psychiatry to understand foundation models. We observed that GPT-3.5 produced on average higher anxiety scores than human participants. One possible explanation for these results could be that GPT-3.5’s training data, which consists of a lot of text taken from the internet, could have inherently shown such a bias, i.e. containing more anxious than happy statements. Of course, large language models have just become good enough to perform psychological tasks, and whether or not they intelligently perform them is still a matter of ongoing debate.

A pdf is here.