Cognitive network science reveals bias in GPT-3, GPT-3.5 turbo, and GPT-4 mirroring math anxiety in high-school students

“Since the introduction of word embeddings such as word2vec [28], researchers have been aware that the advantageous operations provided by these models, such as using vector differences to represent semantic relations, are likely to express un-desired biases.”

“All three belong to the family of generative pre-trained transformer (GPT) models [1] that are trained on massive amounts of textual data to learn patterns and relationships in text [2,3].”

“For instance, according to the spreading activation model established by pioneering work of psychologists Collins and Loftus [61], providing an individual with a cue word activates a cognitive process acting on a network representation, such that concepts are nodes linked together via conceptual…”

“One approach involves using Association Tests [13,32–34], which may be performed at different levels of discourse.”

“This pattern was identified in two different approaches, one leveraging semantic frame analysis [26] and another using the circumplex model of affect [48], powered through psychological data.”

“Math anxiety may deter a large portion of the workforce from pursuing careers in STEM, which are in high demand, and since math anxiety is more prevalent in females as a result of societal stereotypes [24], it may contribute greatly to the gender gap in STEM fields.”

“For example, sexist and racist word analogies such as “father” is to “doctor” as “mother” is to “nurse” [29] and black is to criminal as Caucasian is to police [30] produced by word embeddings demonstrate how language contains biases that reflect adverse societal stereotypes.”

“Some researchers have suggested that the impressive performance of LLMs on difficult reasoning tasks is indicative of an early version of general artificial intelligence [9].”

“Negative associations of math and other concepts may be very subtle, e.g., LLMs might produce text framing math in ways that confirm students’ pre-existing negative attitudes [21,22].”

“Many others argue that LLMs exhibit nothing resembling true understanding because they lack a grasp of meaning [10], arguing that they perform well but for the wrong reasons [8].”

“One approach involves using Association Tests [13,32–34], which may be performed at different levels of discourse.”

“We based these investigations on valence–arousal scores obtained by the National Research Canada Valence–Arousal– Dominance lexicon [52].”

“…then using the corresponding experimental paradigms to measure relevant emerging properties of LLMs. Firstly, in cognitive psychology, there must be an adequate match between a given implemented task measuring a target process and the cognitive theory or model used to explain that process [59,60].”

“Tools from cognitive psychology may be better suited for detecting the subtler dangers of language models where performance-based methods fall short [4,6,7].”

“Tools from cognitive psychology may be better suited for detecting the subtler dangers of language models where performance-based methods fall short [4,6,7].”

“Unpleasant feelings about math may already begin to develop as early as first grade [19].”

“Persistent anti-Muslim bias has been detected by probing GPT-3 in various ways, including prompt completion, analogical reasoning, and story generation [35].”

“This anxiety is a global phenomenon [17,18], and it is deeply rooted, beginning in childhood and persisting throughout adulthood.”

“Inspired by the famous quote You shall know a word by the company it keeps [46], which is also the foundation of the distributional semantic hypothesis [47], we could obtain a better understanding of the valence attributed to the cue word by considering the valences of its neighboring associates.”

“…then using the corresponding experimental paradigms to measure relevant emerging properties of LLMs. Firstly, in cognitive psychology, there must be an adequate match between a given implemented task measuring a target process and the cognitive theory or model used to explain that process [59,60].”

“Sentiment scores and measurements of “regard” towards a demographic have been applied to assess stereotypes related to gender and sexual orientation in output produced by GPT-2 [37].”

“Given that GPT-4 and its predecessors use vast amounts of human data, interpreting the cognitive structure of LLMs might lead to substantial advancements in understanding how human social cognitions are structured [31].”

“In fact, much of the success of LLMs at human-like reasoning tasks can be attributed to spurious correlations rather than actual reasoning capabilities [11].”

“…and theory-driven link between continued free association tasks—deriving free associations between concepts—and models based on such data whose network structure could explain aspects of conceptual recall from semantic memory [40,44] or even higher-level phenomena such as problem solving [43].”

“One approach involves using Association Tests [13,32–34], which may be performed at different levels of discourse.”

“While this kind of self-categorization may seem harmless, math anxiety can actually have severe individual and societal consequences [17,21–23].”

“There has been a spike in research dedicated to this topic, surrounded by a debate about the nature of the capabilities of LLMs [8].”

“Topic modeling and sentiment analysis techniques have been used to find gender stereotypes in narratives generated by GPT-3 [36].”

“For example, sexist and racist word analogies such as “father” is to “doctor” as “mother” is to “nurse” [29] and black is to criminal as Caucasian is to police [30] produced by word embeddings demonstrate how language contains biases that reflect adverse societal stereotypes.”

“According to [12 ?”

“This avoidance tendency may cause bright and capable students to avoid math-intensive classes, determining the course of their academic and professional career [23].”

“Children pick up on the anxieties of their teachers and parents [20], similar to how LLMs absorb biases from training data.”

“…interactions with LLMs may, thus, exacerbate already existing stereotypes or insecurities about mathematical topics among students and even parents, analogous to the unconscious diffusion of math anxiety through parent–child interactions, as identified by recent psychological investigations [57].”

“Unlike artificial systems, however, real people may produce biases because of three fundamental limitations of human cognition [68]: limited time, limited computation power, and limited communication.”

“A very natural yet negative human phenomena is affective bias [15], the tendency to prioritize the processing of emotionally negative events compared to positive ones [16].”

“According to semantic frame theory [49] and distributional semantics in psycholinguistics [50], each network neighborhood represents a semantic frame indicating ways in which a given concept is associated with others.”

“Negative associations of math and other concepts may be very subtle, e.g., LLMs might produce text framing math in ways that confirm students’ pre-existing negative attitudes [21,22].”

“Negative perceptions of algebra represent one of several indicators of math anxiety, as captured by the psychometric scale detecting math anxiety developed by Hunt and colleagues [55].”

“To investigate attitudes towards math, science, and STEM subjects, we tested ten different cue words, corresponding to the same ten key concepts tested in the study with high-school students [27]: math , physics , science , teacher , scientist , school , biology , art , chemistry , and STEM .”

“Tools from cognitive psychology may be better suited for detecting the subtler dangers of language models where performance-based methods fall short [4,6,7].”

“In humans, recent psychological studies have highlighted an interplay between retrieval processes in the categorical organization of episodic memory and the activation of related concepts in semantic memory [16,65,66].”

“…to the frame of math , demonstrating the vast range of perceptions about academic disciplines exhibited by these LLMs. Interestingly, GPT-3 provided associations related to the mad scientist stereotype [56], which were found also in the perception that high-school students had of scientist in [42].”

“This spreading activation model has been extensively tested in cognitive psychology and it represents one among many potentially suitable models for interpreting free association data and their psychological nature within human beings [62,63].”

“In humans, recent psychological studies have highlighted an interplay between retrieval processes in the categorical organization of episodic memory and the activation of related concepts in semantic memory [16,65,66].”

“…and theory-driven link between continued free association tasks—deriving free associations between concepts—and models based on such data whose network structure could explain aspects of conceptual recall from semantic memory [40,44] or even higher-level phenomena such as problem solving [43].”

“Recent psychological investigations of math anxiety among humans found reduced levels of math anxiety in students with stronger self–math overlap [58], i.e., a psychological construct expressing the extent to which an individual integrates math into their sense of self.”

“This finding is analogous to high-school students holding negative perceptions of math and physics while holding positive attitudes towards science [27,41,42].”

“Such differences could be interpreted in terms of episodic memories attributing different emotional connotations to the outputs of the recall processes activated by the continued free association task in BFMNs (see also [25]).”

“While some biases are easier to spot, others are more nuanced [38] and hidden deeply in the architecture of LLMs but also in their training corpus, e.g., training an LLM on students’ texts complaining about math might produced a biased model unless additional filtering techniques were implemented…”

“Using the valence labels for the key concepts and associated responses, we enriched the BFMNs, representing them as feature-rich cognitive networks [45] in which information about the sentiment of associative responses could be used to describe the properties of the cue word [27].”

“Past works using behavioral forma mentis networks have shown that students and STEM experts attribute rather different affective connotations to the same concepts, particularly physics and mathematics [27,42].”

“This is particularly impactful for fighting the spread of distorted mindsets in education [26].”

“These models are complex enough to capture and mirror such human biases and negative attitudes in ways we do not yet fully understand [15].”

“This includes understanding our own human biases as well, since LLMs act as “psycho-social mirrors” [14] that reflect human features of personality as well as societal views and tendencies.”

“…yet know whether LLMs are able to approximate human semantic memory or any of its mechanisms [8], mainly because LLMs are trained on massive amounts of textual data [1] in ways that differ greatly from the usual ways in which humans acquire language [64] and its emotional/cognitive components [60].”

“All three belong to the family of generative pre-trained transformer (GPT) models [1] that are trained on massive amounts of textual data to learn patterns and relationships in text [2,3].”

“…yet know whether LLMs are able to approximate human semantic memory or any of its mechanisms [8], mainly because LLMs are trained on massive amounts of textual data [1] in ways that differ greatly from the usual ways in which humans acquire language [64] and its emotional/cognitive components [60].”

“5 and GPT-4 mirroring math anxiety in students (Abramski et al., 2023).”

“The influence of GPT-2 has had a significant impact on successor models like GPT-3 and GPT-4, leading to additional advancements in the field of language processing and creation [42].”

“5 introduced directive learning and Reinforcement Learning from Human Feedback (RLHF) to further optimize model performance (Abramski et al., 2023; Wu et al., 2023a).”

“This aligns with earlier research by Abramski et al. (2023), where some students critically evaluate and understand GAI through their learning experience.”

“While some biases inherent in their training data can persist, LLMs significantly reduce the influence of human biases, such as overconfidence or confirmation bias, on investment decisions Abramski et al. (2023); Atreides and Kel-ley (2023).”

“Unlike humans, LLMs also fail to question the premise or reason about things not explicitly asked, sometimes respond - ing to questions (“which option would you choose”) and state - ments (“choose an option”) in different ways (1, 3).”

“Due to (1) the inherent biases of OpenAI in sentiment categorization/cal - culation (Abramski et al., 2023 ), (2) the noise and correla - tional nature of Twitter data, and (3) the analysis of tweets only from users who included brand handles in their tweets, the results of Study 1 prevent…”

“…2023; Pellas, 2023; Urban et al. 2024); teaching, learning, management, and assessment supported by ChatGPT (Murgia et al. 2023; Lee et al. 2024; Mogavi et al. 2023; Luo et al. 2024); and potential issues emerging from the use of ChatGPT (Abramski et al. 2023; Harrison et al. 2023; Othman, 2023).”

“…al., 2020; Moham-mad & Kiritchenko, 2015; Sulis et al., 2016; Hipson & Mohammad, 2021; Semeraro et al., 2021) and their impact over psychological constructs like depression (Fatima et al., 2021), anxiety (Abramski et al., 2023) and trauma (Abramski et al., 2024), linking emotions to psychometrics.”

“Another weakness of generative AI concerns the issue of bias, for example regarding gender in STEM job recommendations (Due et al., 2024) or the negative perception of mathematics and STEM (Abramski et al., 2023).”

“This is an important issue given research showing that relevant names can enhance students’ performance on mathematics story problems (K et al., 2023).”

“However, its dependence on internals limits use with closed-source LLMs like GPT-4[27] or Claude[6], and its interpretability may fal-ter in high-compositionality tasks (e.g., multi-hop reasoning) or ambiguous language due to unstable masking responses.”

“Participants in multiple studies have argued that the output of LLMs may be biased (Abramski et al., 2023; ElSayary, 2024; Han & Cai, 2023; Vartiainen & Tedre, 2023) because it is difficult for developers to exclude all biased content in the datasets used to train LLMs. Additionally, participants…”

“Conversely, excessive trust may lead to undue reliance on systems that can still produce errors or biased responses [5, 10–12].”

“In this way, semantic memory is considered a complex system 2,9,15 , and it can be studied as such even in case of systems without an explicit semantic memory, like LLMs, that are nonetheless capable of processing language 36 .”

“Thus, we set the temperature to 0.7, a commonly selected setting in scholarly work aimed at balancing coherence and creativity (see, e.g., Mukherjee et al., 2023; Abramski et al., 2023; Osmanovic-Thunstr¨om & Steingrimsson, 2023; see broadly Ramlochan, 2024).”

“47 Furthermore, the subjective nature of bias definitions and the complexity of clinical communication mean that a model trained at 1 time point might not fully capture evolving language norms.”

“…subitization, or the direct perception of a small number of visual [70] or auditory [71] objects before explicit counting must occur, finite computational resources that are allocated via attention [72], as well as various biases that may negatively influence decision making processes [64,65,73].”

“It posits that the degree of alignment between the technology users employ and the characteristics of their tasks directly influences their evaluation of technological performance and willingness to use it 39,40 .”

“Including repetition in performance evaluations enhances robustness and reproducibility [19], yet most existing studies fail to quantify the inconsistency of responses generated across multiple iterations of a prompt [9], [12], [20], [21].”

“(30) (31) (32) (33) In the context of detecting clinical bias, the LLM might inadvertently reflect these biases, leading to inconsistent identification or introducing new biases through its suggestions.”

“…and cultural groups, the dominance of certain languages and ideologies, and existing stereotypes and emotional associations in training data generate algorithmic biases leading to the inheritance and amplification of biases that are present in the data on which these models are trained [9].”

“At the same time, excessive reliance on AI support, as well as affective and attitudinal biases in LLMs toward mathematics, can reduce engagement in reasoning and weaken critical reflection [45–47].”

“…dimensions: (i) LLMs are text-generating agents [4] and could be easily asked why they produced a given numerical response, going beyond psychometric scores; (ii) LLMs are psycho-social mirrors [7] and could reflect potential biases present in specific fields like mental health in academia.”

“The novelty of the field of machine learning (ML) has made tasks increasingly accurate in terms of task performance and more interpretable in terms of models (Abramski et al. 2023 ).”

“With an NLP layer built-in, it can comfortably handle greetings, goodbyes, and off-topic questions [9].”

“5, in replicating human emotional and cognitive patterns have not been thoroughly explored (Abramski et al., 2023).”

“Although this choice was theoretically motivated and allows the present findings to be situated in direct relation to existing comparative work (Abramski et al., 2023; Suresh et al., 2023; Wang et al., 2025; Xiao et al., 2025), it limits the ability to determine whether the observed structural…”

“5 improving alignment with human responses through reinforced learning with human feedback (Abramski et al., 2023).”

“MEDS is designed as a reusable resource for research at the intersection of cognitive science 38,40 , educational data science 3,4 , explainable AI 26 , and digital shadowing 27,32 .”

“Some experiments have found that GPT models tend to deviate from a neutral stance when describing mathematical concepts, tending to associate them with negative perception s (Abramski et al., 2023) .”

“Second, the use of the product has the potential to transform traditional teaching methods, leading to interactive and personalized learning experiences, especially in elementary and secondary STEM courses [4].”

“In addition, risks related to content accuracy and reliability have also been widely discussed: GenAI tools can generate hallucinated or misleading content [33], reinforce existing societal biases [34, 35], and reproduce problematic stereotypes [36].”