Cognitive Bias in Project Decision-Making in the Era of Conversational Generative AI Appendices

APPENDICES Appendix A. Cognitive Bias Taxonomy Used in the Review TIER BIAS DEFINITION KEY AUTHORS Tier 1 Optimism Bias Systematic tendency to overestimate favorable outcomes while underestimating potential difficulties, risks, or costs. • Kahneman & Tversky (1979) • Flyvbjerg (2006, 2017, 2021) Tier 1 Planning Fallacy Tendency to underestimate the time, effort, and resources required to complete tasks or projects, even when experience suggests otherwise. • Kahneman & Tversky (1979) • Flyvbjerg (2006, 2017, 2021) Tier 1 Anchoring Overreliance on an initial value or piece of information, which subsequently shapes and constrains judgement. • Kahneman & Tversky (1979) • Flyvbjerg (2006, 2021) Tier 1 Availability Bias Judgement influenced by information that is most salient or easily recalled rather than by objective or representative evidence. • Tversky & Kahneman (1974) Tier 1 Overconfidence Inflated belief in one’s own knowledge, accuracy, or predictive abilities, leading to miscalibrated decisionmaking. • Kahneman (2011) • Moore & Healy (2008) Tier 1 Hindsight Bias Inclination to view past events as more predictable or inevitable after their outcomes are known. • Fischhoff (1975 Tier 1 Uniqueness Bias Assumption that one’s project or situation is fundamentally different from comparable cases, thereby limiting learning from prior evidence. • Flyvbjerg (2017, 2021) Tier 1 Base-Rate Fallacy Neglect or underweighting of statistical base-rate information in favor of case-specific or anecdotal details. • Kahneman & Tversky (1973) Tier 1 Escalation of Commitment Continued investment in a failing course of action due to prior resource commitments, despite evidence of diminishing returns. • Staw (1976) • Brockner (1992) Tier 2 Automation Bias Tendency to over-rely on AI-generated outputs, accepting automated recommendations even when they are incomplete, low-quality, or incorrect. • Parasuraman & Riley (1997). Tier 2 Algorithm Aversion Tendency to reject or avoid algorithmic advice after observing errors, showing lower tolerance for machine mistakes than for human ones. • Dietvorst, Simmons & Massey (2015). Tier 2 Trust Bias Disposition to grant unwarranted trust or distrust to an AI system based on perceived reliability, confidence cues, or presentation style rather than evidence of actual performance. • Lee & See (2004); • Hoff & Bashir (2015). Tier 2 Anthropomorphism Bias Attribution of human-like intentions, understanding, or expertise to an AI system, leading users to overestimate its cognitive or interpretive capabilities. • Epley, Waytz & Cacioppo (2007). Tier 2 Expertise Bias * Tendency to misjudge the competence of an AI system, attributing expertise or authority beyond its actual capabilities and deferring to its recommendations inappropriately. • Endsley (1995, 1998); • Bansal et al. (2021) • Klein & Reddy (2024). Tier 2 Illusion of Understanding * Overestimation of the AI’s understanding due to the fluency, coherence, or confidence of the model’s responses. • Rozenblit & Keil (2002) • Binz & Schulz, (2023) • Powell & Riccardi (2025). Tier 2 Framing Effects * Sensitivity of decisions to how options or information are presented, with AI-generated wording, emphasis, or narrative structure influencing user preferences and evaluations. • Tversky & Kahneman (1981); • Lu et al. (2024). Tier 2 Loss Aversion * Tendency to weigh potential losses more heavily than equivalent gains, particularly the perceived loss of status, autonomy, or professional relevance when interacting with or relying on conversational generative AI. • Kahneman & Tversky (1979); • Leoni et al. (2024). Appendix B. Mapping of Explicit Bias Terms to the Cognitive Bias Taxonomy This appendix lists all explicit cognitive biases identified in the included studies and illustrates how each term aligns with the review’s two-tier cognitive bias taxonomy. ART. ID VERBATIM (WITH LOCATION) EXPLICIT BIAS NORMALIZED TAXONOMY TERM 3 “Participants may have enjoyed their first interaction with the LLM-based chatbot due to its novelty…” (Dortheimer et al., 2024, p. 11) Novelty effect • Automation Bias (Tier 2) 4 Automation Bias occurs when users blindly trust AI results…” (Frater & Mushininga, 2025, p. 59) Automation bias • Automation Bias (Tier 2) 5 “Confirmation bias…” (Handler et al., 2024, p. 5) Confirmation bias • Confirmation bias (Tier 1) 8 “Organizations must address the risks of overreliance on AI…” (Hettrich et al., 2025, p. 2) Overreliance • Automation Bias (Tier 2) 8 “Overreliance on GenAI could limit employees’ learning…” (Hettrich et al., 2025, p. 11) Overreliance • Automation Bias (Tier 2) 8 “There is a risk of over-reliance on potentially incorrect results…” (Hettrich et al., 2025, p. 10) Overreliance • Automation Bias (Tier 2) 13 “This human reflection process is susceptible to recency bias…” (Martin et al., 2025, p. 12) Recency bias • Availability Bias (Tier 1) 13 “Loss aversion…” (Martin et al., 2025, p. 2) Loss aversion • Loss aversion (Ties 1) 13 “This corresponds to the optimism bias concept…” (Martin et al., 2025, p. 14) Optimism bias • Optimism Bias (Tier 1) 13 “Group 2’s decreased residual risk ratings… may suggest optimism bias…” (Martin et al., 2025, p. 14) Optimism bias • Optimism Bias (Tier 1) 14 “Users may habitually accept AI-generated answers without rationalizing or critiquing them.” (Mbizo et al., 2024, p. 365) Overreliance • Automation Bias (Tier 2) 14 “Over-reliance on generative AI assistance by developers.” (Mbizo et al., 2024, p. 361) Overreliance • Automation Bias (Tier 2) 16 Mechanisms to avoid automation bias are largely needed…” (NguyenDuc et al., 2023, p. 59) Automation bias • Automation Bias (Tier 2) 17 “Becoming overconfident in GenAI tools may result in a lack of criticism…” (Nyqvist et al., 2024, p. 43) Overconfidence • Automation Bias (Tier 2) 20 “Users tend to overestimate the accuracy of LLM responses…” (Steyvers et al., 2025, p. 221) Overconfidence • Automation Bias (Tier 2) 20 “Users consistently overestimated how accurate LLM outputs were…” (Steyvers et al., 2025, p. Overestimate • Automation Bias (Tier 2) 20 “Human miscalibration is primarily due to overconfidence…” (Steyvers et al., 2025, p. 224) Overconfidence • Automation Bias (Tier 2) 20 “Overconfidence in LLM capabilities is an important concern…” (Steyvers et al., 2025, p. 226) Overconfidence • Automation Bias (Tier 2) 20 “Overconfidence in LLM capabilities is an important concern…” (Steyvers et al., 2025, p. 226) Overestimate • Availability Bias (Tier 1) 21 “Potential for over-reliance on automated systems…” (Vergara et al., 2025, p. 2) Overreliance • Automation Bias (Tier 2) 21 “Over-reliance on automated decision making can reduce human oversight…” (Vergara et al., 2025, p. 14) Overreliance • Automation Bias (Tier 2) Appendix C. Inferred Cognitive Biases: Textual Triggers, and Coding Justifications This appendix provides illustrative examples of the textual or contextual triggers used to infer cognitive biases in studies where the authors did not explicitly name a bias. Each entry demonstrates how inferences were grounded in observable indicators and justified using the review’s interpretive framework. ART. ID VERBATIM EXCERPT WITH TRIGGER HIGHLIGHTED + LOCATION INFERRED BIAS JUSTIFICATION 2 “As a steering committee member, GenAI understands its role and provides presentations upon request…” (Bahi et al., 2024, p. 59) Anthropomorphism Attributing understanding, role awareness, and advice-giving capacity to AI 2 “It seems that GenAI understands the prompt pattern... able to create an action plan...” (Bahi et al., 2024, p. 59) Illusion of Understanding Fluent output misread as deep reasoning 3 “However, a part of the prompt influences how the human engages with the chatbot...” (Dortheimer et al., 2024, p. 4) Framing Effect Prompt wording shapes interpretation 4 “Concerns about AI replacing human roles were also prominent, with 53% believing that AI will significantly replace IT jobs...” (Frater & Mushininga, 2025, p. 66) Loss Aversion The mention of AI “replacing human roles” signals a perceived threat of job loss, prompting lossoriented reactions. 5 Within the context of decision support, this means that language models… may parrot plausible-sounding guidance… without actually offering well-informed or well-reasoned advice. This poses a danger to practitioners, who may act uncritically on generated suggestions.” (Handler et al., 2024, p. 4) Automation Bias The AI’s fluent and coherent output creates an illusion of correctness that prompts users to accept suggestions uncritically. Appendix D. Detailed Search Strings for ProQuest and Web of Science ProQuest Search String (noft(project manage* OR product owner OR scrum master OR project planning OR project control OR program management OR portfolio selection OR portfolio management) AND noft(generative AI OR large language model* OR ChatGPT OR GPT OR AI-powered OR cognitive agents OR Conversational AI) AND noft(cognitive bias* OR heuristic* OR judgment OR decision making OR behavioral bias* OR human AI interaction)) Advanced Search in “Search by command” A peer-reviewed filter was applied 01-01-2023 to 11-03-2025 Language: English Document type: Article Web of Science String TS=(("project manage*" OR "product owner" OR "scrum master" OR "project planning" OR "project control" OR "program management" OR "portfolio selection" OR "portfolio management") AND ("generative AI" OR "large language model*" OR "ChatGPT" OR "GPT" OR "AI-powered" OR "cognitive agents" OR "conversational AI") AND ("cognitive bias*" OR heuristic* OR judgment OR "decision making" OR "behavioral bias*"OR "human AI interaction")) Advanced Search Query Builder 01-01-2023 to 03-11-2025 Document type: No filter applied Language: English Under Database, the option Research Commons was disabled