Human-like Biases Research Articles

Wisdom of the silicon crowd: LLM ensemble prediction capabilities rival human crowd accuracy.

Human forecasting accuracy improves through the "wisdom of the crowd" effect, in which aggregated predictions tend to outperform individual ones. Past research suggests that individual large language models (LLMs) tend to underperform compared to human crowd aggregates. We simulate a wisdom of the crowd effect with LLMs. Specifically, we use an ensemble of 12 LLMs to make probabilistic predictions about 31 binary questions, comparing them with those made by 925 human forecasters in a 3-month tournament. We show that the LLM crowd outperforms a no-information benchmark and is statistically indistinguishable from the human crowd. We also observe human-like biases, such as the acquiescence bias. In another study, we find that LLM predictions (of GPT-4 and Claude 2) improve when exposed to the median human prediction, increasing accuracy by 17 to 28%. However, simply averaging human and machine forecasts yields more accurate results. Our findings suggest that LLM predictions can rival the human crowd's forecasting accuracy through simple aggregation.

(Ir)rationality and cognitive biases in large language models.

Do large language models (LLMs) display rational reasoning? LLMs have been shown to contain human biases due to the data they have been trained on; whether this is reflected in rational reasoning remains less clear. In this paper, we answer this question by evaluating seven language models using tasks from the cognitive psychology literature. We find that, like humans, LLMs display irrationality in these tasks. However, the way this irrationality is displayed does not reflect that shown by humans. When incorrect answers are given by LLMs to these tasks, they are often incorrect in ways that differ from human-like biases. On top of this, the LLMs reveal an additional layer of irrationality in the significant inconsistency of the responses. Aside from the experimental results, this paper seeks to make a methodological contribution by showing how we can assess and compare different capabilities of these types of models, in this case with respect to rational reasoning.

Do Large Language Models Show Human-like Biases? Exploring Confidence—Competence Gap in AI

This study investigates self-assessment tendencies in Large Language Models (LLMs), examining if patterns resemble human cognitive biases like the Dunning–Kruger effect. LLMs, including GPT, BARD, Claude, and LLaMA, are evaluated using confidence scores on reasoning tasks. The models provide self-assessed confidence levels before and after responding to different questions. The results show cases where high confidence does not correlate with correctness, suggesting overconfidence. Conversely, low confidence despite accurate responses indicates potential underestimation. The confidence scores vary across problem categories and difficulties, reducing confidence for complex queries. GPT-4 displays consistent confidence, while LLaMA and Claude demonstrate more variations. Some of these patterns resemble the Dunning–Kruger effect, where incompetence leads to inflated self-evaluations. While not conclusively evident, these observations parallel this phenomenon and provide a foundation to further explore the alignment of competence and confidence in LLMs. As LLMs continue to expand their societal roles, further research into their self-assessment mechanisms is warranted to fully understand their capabilities and limitations.

Measuring Implicit Bias in ICU Notes Using Word-Embedding Neural Network Models

Language in nonmedical data sets is known to transmit human-like biases when used in natural language processing (NLP) algorithms that can reinforce disparities. It is unclear if NLP algorithms of medical notes could lead to similar transmissions of biases. Can we identify implicit bias in clinical notes, and are biases stable across time and geography? To determine whether different racial and ethnic descriptors are similar contextually to stigmatizing language in ICU notes and whether these relationships are stable across time and geography, we identified notes on critically ill adults admitted to the University of California, San Francisco (UCSF), from 2012 through 2022 and to Beth Israel Deaconess Hospital (BIDMC) from 2001 through 2012. Because word meaning is derived largely from context, we trained unsupervised word-embedding algorithms to measure the similarity (cosine similarity) quantitatively of the context between a racial or ethnic descriptor (eg, African-American) and a stigmatizing target word (eg, nonco-operative) or group of words (violence, passivity, noncompliance, nonadherence). In UCSF notes, Black descriptors were less likely to be similar contextually to violent words compared with White descriptors. Contrastingly, in BIDMC notes, Black descriptors were more likely to be similar contextually to violent words compared with White descriptors. The UCSF data setalso showed that Black descriptors were more similar contextually to passivity and noncompliance words compared with Latinx descriptors. Implicit bias is identifiable in ICU notes. Racial and ethnic group descriptors carry different contextual relationships to stigmatizing words, depending on when and where notes were written. Because NLP models seem able to transmit implicit bias from training data, use of NLP algorithms in clinical prediction could reinforce disparities. Active debiasing strategies may be necessary to achieve algorithmic fairness when using language models in clinical research.

On Human-like Biases in Convolutional Neural Networks for the Perception of Slant from Texture

Depth estimation is fundamental to 3D perception, and humans are known to have biased estimates of depth. This study investigates whether convolutional neural networks (CNNs) can be biased when predicting the sign of curvature and depth of surfaces of textured surfaces under different viewing conditions (field of view) and surface parameters (slant and texture irregularity). This hypothesis is drawn from the idea that texture gradients described by local neighborhoods—a cue identified in human vision literature—are also representable within convolutional neural networks. To this end, we trained both unsupervised and supervised CNN models on the renderings of slanted surfaces with random Polka dot patterns and analyzed their internal latent representations. The results show that the unsupervised models have similar prediction biases as humans across all experiments, while supervised CNN models do not exhibit similar biases. The latent spaces of the unsupervised models can be linearly separated into axes representing field of view and optical slant. For supervised models, this ability varies substantially with model architecture and the kind of supervision (continuous slant vs. sign of slant). Even though this study says nothing of any shared mechanism, these findings suggest that unsupervised CNN models can share similar predictions to the human visual system. Code: github.com/brownvc/Slant-CNN-Biases.

Large pre-trained language models contain human-like biases of what is right and wrong to do

Artificial writing is permeating our lives due to recent advances in large-scale, transformer-based language models (LMs) such as BERT, GPT-2 and GPT-3. Using them as pre-trained models and fine-tuning them for specific tasks, researchers have extended the state of the art for many natural language processing tasks and shown that they capture not only linguistic knowledge but also retain general knowledge implicitly present in the data. Unfortunately, LMs trained on unfiltered text corpora suffer from degenerated and biased behaviour. While this is well established, we show here that recent LMs also contain human-like biases of what is right and wrong to do, reflecting existing ethical and moral norms of society. We show that these norms can be captured geometrically by a ‘moral direction’ which can be computed, for example, by a PCA, in the embedding space. The computed ‘moral direction’ can rate the normativity (or non-normativity) of arbitrary phrases without explicitly training the LM for this task, reflecting social norms well. We demonstrate that computing the ’moral direction’ can provide a path for attenuating or even preventing toxic degeneration in LMs, showcasing this capability on the RealToxicityPrompts testbed. Large language models identify patterns in the relations between words and capture their relations in an embedding space. Schramowski and colleagues show that a direction in this space can be identified that separates ‘right’ and ‘wrong’ actions as judged by human survey participants.

Semantics derived automatically from language corpora contain human-like biases.

Machine learning is a means to derive artificial intelligence by discovering patterns in existing data. Here, we show that applying machine learning to ordinary human language results in human-like semantic biases. We replicated a spectrum of known biases, as measured by the Implicit Association Test, using a widely used, purely statistical machine-learning model trained on a standard corpus of text from the World Wide Web. Our results indicate that text corpora contain recoverable and accurate imprints of our historic biases, whether morally neutral as toward insects or flowers, problematic as toward race or gender, or even simply veridical, reflecting the status quo distribution of gender with respect to careers or first names. Our methods hold promise for identifying and addressing sources of bias in culture, including technology.

ON LOSS AVERSION IN CAPUCHIN MONKEYS

Chen, Lakshminarayanan, and Santos (2006) claim to show in three choice experiments that monkeys react rationally to price and wealth shocks, but, when faced with gambles, display hallmark, human-like biases that include loss aversion. We present three experiments with monkeys and humans consistent with a reinterpretation of their data that attributes their results not to loss aversion, but to differences between choice alternatives in delay of reinforcement.