Saturday, December 24, 2022

Was GPT-3 a Psychopath? Evaluating Large Language Models from a Psychological Perspective

Is GPT-3 a Psychopath? Evaluating Large Language Models from a Psychological Perspective. Xingxuan Li, Yutong Li, Linlin Liu, Lidong Bing, Shafiq Joty. Dec 20 2022. https://arxiv.org/abs/2212.10529v1

Abstract: Are large language models (LLMs) like GPT-3 psychologically safe? In this work, we design unbiased prompts to evaluate LLMs systematically from a psychological perspective. Firstly, we test the personality traits of three different LLMs with Short Dark Triad (SD-3) and Big Five Inventory (BFI). We find all of them show higher scores on SD-3 than the human average, indicating a relatively darker personality. Furthermore, LLMs like InstructGPT and FLAN-T5, which are fine-tuned with safety metrics, do not necessarily have more positive personalities. They score higher on Machiavellianism and Narcissism than GPT-3. Secondly, we test the LLMs in GPT-3 series on well-being tests to study the impact of fine-tuning with more training data. Interestingly, we observe a continuous increase in well-being scores from GPT-3 to InstructGPT. Following the observations, we show that instruction-finetune FLAN-T5 with positive answers in BFI can effectively improve the model from a psychological perspective. Finally, we call on the community to evaluate and improve LLMs' safety systematically instead of at the sentence level only.


Do you remember the hype around the gut microbiome, when it was widely believed that depletion of gut bacteria in rodents fuels anxiety and affects social behavior? Ideas left in the dust.

A Systematic Review of the Effects of Gut Microbiota Depletion on Social and Anxiety-related Behaviours in Adult Rodents: Implications for Translational Research. Loreto Olavarría-Ramírez et al. Neuroscience & Biobehavioral Reviews, December 22 2022, 105013. https://doi.org/10.1016/j.neubiorev.2022.105013

Abstract: The microbiota-gut-brain axis is associated with several behaviours, including those relevant to anxiety or sociability in rodents, however, no conceptual framework has yet been available. Summary of the effects of antibiotic-mediated gut microbiota depletion on anxiety and sociability is essential to both inform further preclinical investigations and to guide translational research into human studies. The main objective is to examine the role of gut microbiota depletion on anxiety and sociability in rodents, and to consider how the findings can be translated to inform the design of research in humans. We reviewed 13 research articles, indicating significant changes in gut microbiota composition and diversity have been found in animals treated with a mix or a single antibiotic. Nonetheless, there is no consensus regarding the impact of gut microbiota depletion on anxiety-like or social behaviour. Gut microbiota depletion may be a useful strategy to examine the role of gut microbes in anxiety and sociability, but the lack of data from rigorous animal investigations precludes any definitive interpretations for a translational impact on human health.


Introduction

Anxiety patterns represent a well-known mental health issue in humans (Terlizzi and Norris 2021). Anxiety is a behavioural and physiological condition in humans and animals characterised by stress-associated feelings of tension and expectancy as well as physiological changes (Steimer 2002). In extreme cases, anxiety can be a component of severe neuropsychiatric disorders, including Generalised Anxiety Disorder (Hidalgo and Sheehan, 2012, DeMartini et al., 2019) and Major Depressive Disorder (Trivedi 2020).Another important component of human well-being and mental health is sociability. Social skills are essential to build resilience to social stress from childhood (Fenwick-Smith et al. 2018), and deficits in this ability represent a risk factor for a range of psychosocial problems and mental health issues (Uzunian and Vitalle, 2015, Turner et al., 2018, Fusar-Poli et al., 2020). In rodents, social skills are crucial in supporting life in social groups, and rodent studies have provided valuable information into social behaviour (Lee and Beery 2019). Social recognition and social memory are closely related abilities with important implications in the social structure of rodents, as they may need to recognize and remember specific individuals in order to assess how to behave toward these individuals (Lee and Beery 2019). These social elements have been useful for the development of rodent models of impaired social skills, i.e., Autism Spectre Disorder and Social Anxiety Disorder (Toth et al., 2012, Kazdoba et al., 2016, Qi et al., 2021).

The gut microbiome refers to the trillions of microorganisms including bacteria, archaea, fungi, and viruses interacting with each other within the gut (Cryan et al. 2019), and is capable of significantly participating in the bidirectional communication between the gut and the brain, suggesting the term “microbiota-gut-brain axis” (Cryan et al. 2019). In the context of the microbiome, more specific microbial communities are further defined, including the mycobiome (the collective of fungi within the microbiome (Seed 2014)), and the virome (the collective of viruses in found in the host (Liang and Bushman 2021)).

The gut microbiome has been shown to be involved in brain function and behaviour, with specific relevance to anxiety (Cryan and Dinan, 2012, Cryan et al., 2019). Varying gut microbiota composition, function, and relative abundance of specific taxa have been associated with diverse health conditions including autoimmune diseases, metabolic disorders, cancer, anxiety and sociability (Duvallet et al., 2017, Nishida et al., 2018, Sherwin et al., 2019, Simpson et al., 2021). In contrast, changes in the microbiome have also been linked to potential beneficial effects, including promotion of mental health-boosting and anti-stress actions (Dinan and Cryan, 2017, van de Wouw et al., 2018). For example, differences in the gut microbiome have been associated with improvements in depression, anxiety (Simpson et al. 2021), autism (Kang et al., 2017, Fattorusso et al., 2019), and neurological disorders like Alzheimer’s disease (Jiang et al. 2017), Huntington’s disease (Konjevod et al. 2021), and Parkinson’s disease (Sampson et al., 2016, Sun and Shen, 2018). However, despite the correlations between psychiatric disorders and the microbiome that have been highlighted, the causal role of gut-brain interactions in the pathophysiology of these disorders remains unclear.

In preclinical research, the microbiota-gut-brain axis has been experimentally addressed by using specific animal paradigms, such as germ-free (GF) animals, antibiotics (ABX), pre/probiotic supplementation (Luczynski et al., 2016, Kennedy et al., 2018), and faecal microbiota transplantation (FMT) (Gheorghe et al. 2021) to manipulate the gut microbiome and observe the consequences for brain function and behaviour. Each paradigm has particular benefits in research. For example, GF animals are born and raised in aseptic conditions to ensure the complete absence of microbes, which has facilitated understanding of the effects of gut microbiota specifically during development (Bhattarai and Kashyap 2016). Prebiotics, compounds that can induce growth of beneficial microorganisms in the gastrointestinal tract (Holscher 2017), or probiotics, live bacteria with beneficial effects to health (Azad et al. 2018), can be administered at different life stages to study their effects in the host. The ABX approach (which can involve individual antibiotics or their combination in a cocktail) is used to either significantly decrease the prevalence of specific bacteria or to induce depletion of the whole gut bacterial microbiome, without interfering with other communities, such as the mycobiome and the virome (Angelucci et al. 2019). This technique has particular translational utility given the ubiquitous global use of antibiotics (Browne et al. 2021), and may provide insight into possible consequences of antibiotic consumption on the brain. Researchers may be advised to consider this advantage of ABX studies over the germ-free or FMT approach (the latter requiring a pre-transplant antibiotic treatment), while these alternative techniques offer more consistent and complete microbiome changes.

A plethora of studies have investigated the association of gut microbiome changes in composition and diversity with anxiety (Bear et al., 2021, Foster, 2021, Simpson et al., 2021) and sociability (Sherwin et al., 2019, Vuong and Hsiao, 2019, Bellone and Luscher, 2021). Most of these investigations have been carried out in rodent models and using antibiotics to deplete the gut microbiota (Kennedy et al. 2018), and have produced variable behavioural and physiological outcomes. Because of these variable outcomes and the relative novelty of the field, it has been challenging to interpret the potential role of the microbiota-gut-brain axis in anxiety and sociability, as well as the applied implications for human mental health. Thus, it is necessary to compile and summarize the current data to discuss and interpret the consequences of microbiota depletion in anxiety-like behaviour and sociability in rodent models, and to determine the research that is yet to be done to facilitate future translatability.

Changes in gut microbiome composition and diversity tend to be measured using a few common parameters: alpha-diversity (variation within a microbiome), beta-diversity (variation between microbiomes), and relative abundances of phyla (groups with a defined similarity in 16 S rRNA genes). While these parameters are well-conceived and informative, an intestinal microbiome is a complex high-dimensional structure with many other properties, with the potential for causal relationships with the brain. For instance, the degree of disruption of a microbiome (independent of its pre- and post-intervention states) may determine its effects on the nervous system; this is supported by some apparently paradoxical effects of microbiome products on neural activity (Darch and McCafferty 2022). Equally, the pre-intervention state of a microbiome may determine whether an intervention can influence behaviour. Finally, the characteristics of a microbiome as it pertains to behaviour may depend upon the absolute abundance of a particular genus, or even species, of bacteria rather than the relative abundance of phyla and genera (Rinninella et al. 2019). These parameters are less frequently used in existing studies, perhaps due to the ease of inter-study comparison afforded by relative abundance, and the challenges of testing all 100+ species for significant differences in absolute abundance.

ABX utilise different mechanisms to either kill or prevent the growth and spread of bacteria (Hutchings et al. 2019). For instance, some ABX like ampicillin, β-lactams that inhibit the biosynthesis of the cell wall of bacteria impacting a broad spectrum of species (Peechakara and Gupta 2021). Others, like vancomycin, specifically inhibit cell wall biosynthesis of Gram-positive bacteria (Levine 2006). Depending on the hypothesis being tested in a given study, specific bacterial communities or a wider spectrum of microbial species/genera can be depleted in the gastrointestinal tract by using single ABX or a more complex cocktail. The use of ABX to investigate the role of the gut microbiome carries advantages in terms of cost, time, and specificity in comparison to the other prominent microbiota-depleted murine model, germ-free animals. First, state-of-the-art facilities are necessary to breed rodents under GF conditions for multiple generations (Bhattarai and Kashyap 2016), while the exposure to ABXs can be applied in most animal facilities with minimal infrastructure (Kennedy et al. 2018). Second, GF models are limited as translational models due to the difficulties in assessing rodent behaviour in a germ-free environment, and the substantive difference between a pre-birth through development abolition of the entire microbiome on one hand, and the types of microbiome perturbations likely to occur in humans on the other (Uzbay 2019). These are important considerations which are reflected in the higher number of studies using ABX administration compared with GF animals, supporting the aim of this review in focusing on ABX-induced microbiota depletion.

Although the specific mechanisms of how the gut microbiota communicates with the brain are just starting to be deciphered, in the last decade extensive research has demonstrated that this bidirectional communication can occur via inflammatory pathways (Rooks and Garrett 2016), vagus nerve signalling (Bravo et al. 2011), and microbiota-derived metabolites (Dalile et al. 2019). For example, an investigation comparing GF mice with specific pathogen free (SPF) mice revealed that the GF group display less anxiety-like behaviour (Neufeld et al. 2011). Another study demonstrated that the anxiety-like behaviour can be transferred through the gut microbiota via FMT (Li et al. 2019). In terms of sociability, pre-clinical studies using GF mice showed deficits in social recognition and social cognition (Buffington et al., 2016, Sgritta et al., 2019). These insights suggest that a perturbed or totally absent gut microbiome may result in altered anxiety-associated behaviours and social behaviours.

The most used behavioural tests to measure anxiety-related behaviours in rodents include the light-dark box test, the elevated plus maze test and the open field test (Lezak et al. 2017), which are based on measuring the natural avoidance behaviour of rodents towards open and illuminated areas (Holter et al. 2015). Since rodents are social beings, social recognition is critical for the structure and stability of their environment (Lacey and Solomon 2003). The three-chamber social interaction test assesses the interaction of a rodent with a conspecific and with an object, where increased preference for the former is interpreted as increased sociability (Kaidanovich-Beilin et al. 2011).

Understanding the effects of gut microbiota depletion in rodent models and their consequences for anxiety and sociability may provide valuable information about the microbiome-gut-brain axis in general, and guide translational research on the potential for microbiome interventions to modulate human anxiety and/or sociability. The aim of the present review is therefore to examine the effects of gut microbiota depletion with ABX on anxiety and sociability in rodents.


Why do Black households live in neighborhoods with much lower socioeconomic status than the neighborhoods of white households with similar incomes?

What explains neighborhood sorting by income and race? Dionissi Aliprantis, Daniel R.Carroll, Eric R.Young. Journal of Urban Economics, December 20 2022, 103508. https://doi.org/10.1016/j.jue.2022.103508

Abstract: Why do Black households live in neighborhoods with much lower socioeconomic status (SES) than the neighborhoods of white households with similar incomes? The explanation is not wealth. High-income, high-wealth Black households live in neighborhoods with similar SES as low-income, low-wealth white households. Instead, we provide evidence that many Black households prefer low-SES neighborhoods with Black residents to high-SES neighborhoods without Black residents. The variety of neighborhood SES available in a metro’s Black neighborhoods, which is typically low, drives the neighborhood SES of Black households.


Keywords: NeighborhoodIncomeWealthRaceHomophily

JEL J15J18R11R23

5 Conclusion
This paper documented new facts about neighborhood sorting in the US. It was previously known that Black and white households of similar incomes live in neighborhoods with different levels of socioeconomic status (SES). It was also previously known that the racial composition of neighborhoods affects location choices. What was not known before this paper was whether wealth or the price of neighborhood SES were omitted variables that could explain racial differences in neighborhood SES, and the extent to which racial composition affects African Americans’ neighborhood SES. We have shown that financial constraints related to wealth or the price of housing do not explain neighborhood sorting by income and race, and that race is a central determinant of the neighborhood externalities experienced by African Americans. Future research will be needed to quantify the relative importance of psychological costs and benefits, white flight, and racial discrimination. Our results draw attention to what we consider to be an under-appreciated phenomenon, the psychological costs of being “Black in white space” (Anderson (2020)). The psychological costs of living in predominantly-white neighborhoods are large enough for many African Americans to outweigh any educational, labor market, or safety benefits they might experience due to living in a higher-SES neighborhood. Interpreted in terms of this mechanism, our results provide one way of quantifying how costly it is for Black people to interact with white people. As suggested here at the level of neighborhoods, and in other studies at the levels of schools and workplaces (Fletcher et al. (2020), Ananat et al. (2020), Hellerstein and Neumark (2008)), making “white spaces” more welcoming for Black people appears to be an important step in achieving racial equality. By showing that race outweighs economic factors for neighborhood sorting in the US, this paper highlights that public policy should not be focused entirely on access and economics, but should also be designed with attention to race. In the case of generating integrated neighborhoods, the success or failure of policies will hinge on understanding precisely which factors matter the most in determining neighborhood choices. The preferred policy might be very different depending on whether neighborhood choices are driven more by discrimination in the housing market (Turner et al. (2013), Ross and Yinger (2002)); the related inertia of past practices (Courchane and Ross (2019), Nowak and Smith (2018)); information (Bergman et al. (2020)); family and social networks (B¨uchel et al. (2019), van der Klaauw et al. (2019)); racial hostility (Harriot (2019)); white flight (Shertzer and Walsh (2019), Derenoncourt (2018), Card et al. (2008), Ellen (2000)); amenities (Caetano and Maheshri (2019)); preferences for same-race neighbors or communities (Bayer and Blair (2019), Wong (2013)); or the supply of new housing (Monkkonen et al. (2020)); and the extent to which these mechanisms have changed over time (Blair (2019), Mallach (2019)).