This paper introduces a novel method, referred to as "hashing", which involves masking potentially bias-inducing words in large language models (LLMs) with hash-like meaningless identifiers to reduce cognitive biases and reliance on external knowledge. The method was tested across three sets of experiments involving a total of 490 prompts. Statistical analysis using chi-square tests showed significant improvements in all tested scenarios, which covered LLama, ChatGPT, Copilot, Gemini and Mixtral models. In the first experiment, hashing decreased the fallacy rate in a modified version of the "Linda" problem aimed at evaluating susceptibility to cognitive biases. In the second experiment, it improved LLM results on the frequent itemset extraction task. In the third experiment, we found hashing is also effective when the Linda problem is presented in a tabular format rather than text, indicating that the technique works across various input representations. Overall, the method was shown to improve bias reduction and incorporation of external knowledge. Despite bias reduction, hallucination rates were inconsistently reduced across types of LLM models. These findings suggest that masking bias-inducing terms can improve LLM performance, although its effectiveness is model- and task-dependent.

Although large language models (LLMs) have demonstrated remarkable proficiency in "understanding" text and generating human-like text, they may exhibit biases acquired from training data in doing so. Specifically, LLMs may be susceptible to a common cognitive trap in human decision-making called the representativeness heuristic. This is a concept in psychology that refers to judging the likelihood of an event based on how closely it resembles a well-known prototype or typical example versus considering broader facts or statistical evidence. This work investigates the impact of the representativeness heuristic on LLM reasoning. We created ReHeAT (Representativeness Heuristic AI Testing), a dataset containing a series of problems spanning six common types of representativeness heuristics. Experiments reveal that four LLMs applied to REHEAT all exhibited representativeness heuristic biases. We further identify that the model's reasoning steps are often incorrectly based on a stereotype rather than the problem's description. Interestingly, the performance improves when adding a hint in the prompt to remind the model of using its knowledge. This suggests the uniqueness of the representativeness heuristic compared to traditional biases. It can occur even when LLMs possess the correct knowledge while failing in a cognitive trap. This highlights the importance of future research focusing on the representativeness heuristic in model reasoning and decision-making and on developing solutions to address it.

The field of Artificial Intelligence (AI) is focusing on creating automated decision-making (ADM) systems that operate as close as possible to human-like intelligence. This effort has pushed AI researchers into exploring cognitive fields like psychology. The work of Daniel Kahneman and the late Amos Tversky on biased human decision-making, including the study of the conjunction fallacy, has experienced a second revival because of this. Under the conjunction fallacy a human decision-maker will go against basic probability laws and rank as more likely a conjunction over one of its parts. It has been proven overtime through a set of experiments with the Linda Problem being the most famous one. Although this interdisciplinary effort is welcomed, we fear that AI researchers ignore the driving force behind the conjunction fallacy as captured by the Linda Problem: the fact that Linda must be stereotypically described as a woman. In this paper we revisit the Linda Problem and formulate it as a fairness problem. In doing so we introduce perception as a parameter of interest through the structural causal perception framework. Using an illustrative decision-making example, we showcase the proposed conceptual framework and its potential impact for developing fair ADM systems.

The EU AI Act Article 5 is designed to regulate AI manipulation to prevent potential harmful consequences. However, the practical implementation of this legislation is challenging due to the ambiguous terminologies and the unclear presentations of manipulative techniques. Moreover, the Article 5 also suffers criticize of inadequate protective efficacy. This paper attempts to clarify terminologies and to enhance the protective efficacy by integrating insights from psychology and behavioral economics. Firstly, this paper employs cognitive psychology research to elucidate the term subliminal techniques and its associated representation. Additionally, this paper extends the study of heuristics: a set of thinking shortcuts which can be aroused for behavior changing from behavior economics to the realm of manipulative techniques. The elucidation and expansion of terminologies not only provide a more accurate understanding of the legal provision but also enhance its protective efficacy. Secondly, this paper proposes five classical heuristics and their associated examples to illustrate how can AI arouse those heuristics to alter users behavior. The enumeration of heuristics serves as a practical guide for stakeholders such as AI developers, algorithm auditors, users, and legal practitioners, enabling them to identify manipulative techniques and implement countermeasures. Finally, this paper critically evaluates the protective efficacy of Article 5 for both the general public and vulnerable groups. This paper argues that the current protective efficacy of Article 5 is insufficient and thus proposes specific revision suggestions to terms a and b in Article 5 to enhance its protective efficacy. This work contributes to the ongoing discourse on AI ethics and legal regulations, providing a practical guide for interpreting and applying the EU AI Act Article 5.