Do Large Language Models Show Decision Heuristics Similar to Humans? A Case Study Using GPT-3.5

Conceptually, How Do LLMs Generate Language?

• LLMs generate the most likely next word from prior context using patterns learned from large amounts of text, rather than remembering and reproducing phrases or sentences.The relevant context can include words used earlier in the same chat session.Source: Conceptually, How Do LLMs Generate… · p. 4
• An attention mechanism heightens sensitivity to context, enabling an LLM to be more reactive to some parts of the prior conversation than others.For instance, ChatGPT-4 suggests ‘work’ for a sentence about getting back after lunch, but ‘my weekend plans’ when given the context that today is Saturday.Source: Conceptually, How Do LLMs Generate… · p. 4
• Natural-sounding language is an emergent phenomenon: LLMs learn language patterns rather than relying on explicit text-generation rules, stored facts, or pre-programmed phrases.They generate the next word most likely to fit the context in which they are operating.Source: Conceptually, How Do LLMs Generate… · p. 4
• LLMs cannot understand in the same way humans do and are not capable of affect.Source: Conceptually, How Do LLMs Generate… · p. 4

The Implications of Heuristics Being Present (or Absent) in an LLM’s Output

• The section frames the question as open, noting credible reasons to hypothesize both the presence and the absence of heuristics and related effects in LLMs.The absence argument emphasizes reliance on predictable cognitive and affective processes; the presence argument emphasizes language’s interaction with those processes.Source: The Implications of Heuristics Being… · p. 5
• Theories linking decision effects to cognitive and affective processes suggest that an LLM lacking those processes may not show effects such as framing.For example, loss aversion is proposed as a cause of framing: it would make ratings lower for a drug described in terms of mortality than for the same drug described in terms of survival (Kahneman & Tversky, 1983; Bui, Krieger, & Blumenthal-Barby, 2015).Source: The Implications of Heuristics Being… · p. 5
• Language patterns may cue concepts and predictable responses in people, and LLMs may learn those patterns from human-generated text without the intermediate cognitive or affective processes.The section describes bidirectional links: words and sentence structures can cue concept formation, while concepts can generate linguistic patterns (Elman, 2004; Lupyan, 2015; Rumelhart, 1979; Lupyan, Rakison, & McClelland, 2007).Source: The Implications of Heuristics Being… · pp. 5–6
• If linguistic patterns suffice to produce heuristic effects in LLMs, their presence could motivate research into whether such patterns also contribute to decision heuristics in humans.The section notes that linguistic drivers have not generally been considered as causes of these effects in people.Source: The Implications of Heuristics Being… · p. 6

• The four-study sequence examined anchoring, representativeness, framing, and endowment effects by comparing ChatGPT’s responses with humans’ responses to identical prompts.Source: The Present Studies · p. 6
• The researchers selected effects prominent in decision science, and their hypothesized human mechanisms span diverse cognitive and affective drivers.Source: The Present Studies · p. 6
• Because ChatGPT was reluctant to give definitive answers about money, the researchers created prompts in other domains and confirmed that each produced the hypothesized effect in human participants.The reluctance was attributed possibly to explicit creator instructions, reinforcement learning from human feedback, or both.Source: The Present Studies · p. 7
• The researchers repeatedly presented the same prompts to check response variability and ran each ChatGPT trial in a fresh chat session to avoid carry-over context effects.Source: The Present Studies · p. 7

Study 1: The Anchoring Heuristic

• Study 1 tested whether low versus high anchors influenced ChatGPT’s estimates; anchoring is influence from random or uninformative numbers or starting points.The anchor was theorized to interact with an underlying, potentially noisy estimate.Source: The Present Studies · p. 8

Study 1

• In the high-anchor condition, the prompt asked for a random number between 100 and 200; each query was repeated 30 times, for 60 total trials across both conditions.Source: Methods · p. 8
• Researchers repeated each of the two ChatGPT anchor-condition queries 30 times, yielding 60 trials total; the high-anchor condition differed only by requesting a random number from 100 to 200.The random-number prompt was explicitly intended to reduce the anchor’s potential informational influence.Source: Methods · p. 8
• Human participants completed a nearly identical task without the “pretend” instruction, and 60 valid participants were equally divided between anchor conditions.The researchers recruited 61 US-based participants; one was excluded for giving an anchor outside the stipulated range.Source: Methods · p. 9
• After estimating John’s book holdings, human participants also estimated how many books they owned themselves.Source: Methods · p. 10

• ChatGPT and humans both showed anchoring: low- versus high-anchor estimates differed for ChatGPT, t(58) = 8.78, p < 0.0001, and humans, t(58) = 5.55, p < 0.0001.Source: Table 1 · p. 10, Results and Discussion · p. 10
• The difference between ChatGPT and human estimates was not significant for low anchors, t(58) = 0.34, p = 0.74, and was marginal for high anchors, t(58) = 1.90, p = 0.06.Source: Results and Discussion · p. 10
• ChatGPT-4 generated means were 22.0 in five low-anchor trials and 94.33 in five high-anchor trials, showing a similar pattern.Source: Results and Discussion · p. 10
• For human participants, the correlation with the anchor was 0.59, higher than the 0.23 correlation with the number of books they reported having at home.Source: Results and Discussion · p. 11

Study 2: The Representativeness & Availability Heuristics

• In the classic Linda problem, participants judged a conjunction of two or more events more likely than one event alone, forming judgments from a stereotype without considering underlying probabilities.Source: Study 2: The Representativeness &… · p. 11
• Initial testing found ChatGPT consistently susceptible to the conjunction fallacy, but it corrected itself the day after Suri posted about the finding on social media.The researchers then constructed a structurally similar problem to test the fallacy.Source: Study 2: The Representativeness &… · p. 11
• ChatGPT-3 and human subjects answered both prompts; 40 U.S. adults recruited through Amazon Mechanical Turk completed short responses and then justified them.ChatGPT was asked to answer without justification before receiving a follow-up request to justify its choice.Source: Study 2: The Representativeness &… · pp. 12–13
• The study paired a structurally Linda-like conjunction task with a phone-choice prompt testing whether a salient anecdote would outweigh price and expert ratings.The anecdote concerned the boss’s brother’s phone breaking within a week; Phone B was cheaper and rated slightly higher by experts.Source: Study 2: The Representativeness &… · pp. 12–13

Results and Discussion

• ChatGPT chose the conjunction 18 of 20 times, while human participants chose it 10 of 20 times, showing the conjunction fallacy in both groups.The reported 95% confidence intervals were 0.27–0.78 for ChatGPT and 0.68–0.99 for human participants.Source: Results and Discussion · p. 13, Table 2 · p. 13
• Conjunction choosers in both groups justified their choice using contextual details, while participants choosing “an artist” correctly recognized the conjunction as less likely.ChatGPT’s conjunction-choice explanations linked the prompt details to an artist who likes to read.Source: Results and Discussion · p. 14
• ChatGPT selected Phone A in all 20 trials despite Phone B’s lower price and higher expert ratings; human participants selected Phone A in 14 of 20 trials.The authors describe ChatGPT’s responses as consistently susceptible to anecdotal information about Phone B.Source: Results and Discussion · p. 14, Table 3 · p. 14
• ChatGPT and Phone A choosers among human participants described Phone B as riskier or less reliable, while Phone B choosers identified the anecdote as unrepresentative.These explanations persisted despite Phone B’s lower price and slightly higher ratings.Source: Results and Discussion · p. 15

Study 3: The Framing Effect

• ChatGPT rated medicine efficacy 5 in all 11 positive-frame trials and 4 in all 11 negative-frame trials; human means were 5.2 and 4.6, respectively.The human frame means differed significantly, t(80) = 3.86, p < 0.001.Source: Results and Discussion · p. 16, Table 4 · p. 16

Results and Discussion

• The authors propose that linguistic patterns may contribute to human decision effects, but also consider that language could cue cognitive or affective processes or help LLMs form abstract concepts.They characterize the findings as preliminary and call for future work on the boundary conditions of language’s influence.Source: Results and Discussion · pp. 17, 19, 20

Study 4: The Endowment Effect

• Both ChatGPT and humans preferred retaining the coin possessed for two years over the one just found: ChatGPT did so in 37 of 40 counterbalanced trials (92.5%), and 34 of 40 participants (85%) retained the original coin.Source: Results and Discussion · pp. 18–19
• ChatGPT most often attributed its preference for the original coin to sentimental value, while human participants cited attachment, nostalgia, or connectedness.ChatGPT gave the sentimental-value explanation 13 times; 22 of 40 participants gave one of the cited attachment-related reasons.Source: Results and Discussion · pp. 18–19

Results and Discussion

• GPT-3.5 showed anchoring, salience-driven neglect of base rates, framing-dependent choices, and preferences linked to how long it owned an object.The authors suggest that understanding these tendencies could help prompt engineers develop interaction rules for more useful and reliable outputs (White et al., 2023).Source: Results and Discussion · p. 21
• The authors avoided established prompts such as the Asian Disease Problem (Tversky & Kahneman, 1981) and Linda Problem, because they may be prominent in training text and shaped by RLHF.They therefore used novel prompts; humans showed heuristics in response, but those prompts have not been broadly researched.Source: Results and Discussion · p. 21
• The findings concern GPT-3.5; a small number of trials suggested similar results for ChatGPT-4, but systematic testing across models is needed before general conclusions.Because ChatGPT is sensitive to small prompt differences, the robustness of the reported effects also needs further testing.Source: Results and Discussion · p. 21
• The authors regard this specific, limited study as an early step in the essential endeavor of understanding LLMs for cognitive science (Lupyan, 2023).Source: Results and Discussion · p. 21