What if artificial intelligence does not merely support investment decisions, but quietly reshapes the biases behind them?
Generative AI is rapidly becoming part of how investors search for information, test ideas and form judgements. Yet its most important influence may not lie in whether an answer is simply right or wrong. More fundamentally, it can alter what users notice, what they trust and how confidently they interpret uncertain information.
This work examines that emerging behavioural dimension of investment decision-making. Drawing on behavioural finance and research on large language models, it focuses on overconfidence, confirmation bias and anchoring. The central question is whether the use of systems such as ChatGPT can weaken existing distortions, reproduce them in a more persuasive form, or change the mechanisms through which they operate.
The analysis combines a systematic literature review with a qualitative prompt experiment based on a real investment case. ChatGPT was examined across seven controlled conditions and 21 independent trials, comparing neutral, bias-inducing and debiasing formulations. The results show that prompt design does not necessarily overturn the analytical structure of a response. Instead, it can substantially alter which arguments, risks and perspectives receive attention, depth and emphasis.
The implications extend beyond simple amplification or reduction. Confidence can shift from personal judgement to trust in AI-supported reasoning. Confirmation bias can evolve from selective information search into a reinforcing dialogue with the system. Anchors no longer need to originate solely from market prices or prior expectations. They can emerge from values, scenarios and reference points generated within the interaction itself.
All 21 model responses are reproduced verbatim in the appendix, allowing readers to trace the analysis directly back to its source material and assess the findings for themselves.
The result is a broader view of decision-making in the age of generative AI. Human judgement and algorithmic output no longer operate as separate forces. They increasingly form a shared cognitive environment in which prompts, interpretations and generated responses influence one another.
The crucial question is therefore no longer simply whether investors are biased. It is what becomes of those biases once artificial intelligence becomes part of the way decisions are formed.
Table of Contents
1 INTRODUCTION
1.1 PROBLEM STATEMENT AND RELEVANCE
1.2 STUDY OBJECTIVES
1.3 RESEARCH QUESTION
1.4 METHODOLOGICAL APPROACH
1.4.1 Systematic literature search
1.4.2 Inclusion and exclusion criteria
1.4.3 Identified records
1.4.4 Use of ChatGPT in prompt experiments
1.5 THESIS STRUCTURE
2 THEORETICAL FOUNDATIONS
2.1 DECISION-MAKING UNDER UNCERTAINTY
2.1.1 Bounded rationality
2.1.2 Information processing and heuristics
2.1.3 Investment decisions as strategic decisions
2.2 CAPITAL MARKETS AS A DECISION-MAKING ENVIRONMENT UNDER UNCERTAINTY
2.2.1 Forecast uncertainty and market volatility
2.2.2 Information complexity and asymmetries
2.2.3 Expectation formation and market narratives
2.2.4 Rationality assumption and behavioural critique
2.3 COGNITIVE BIASES IN INVESTMENT DECISIONS
2.3.1 Conceptual positioning within behavioural finance
2.3.2 Overconfidence bias
2.3.3 Confirmation bias
2.3.4 Anchoring bias
2.4 LLMS AS DECISION SUPPORT SYSTEMS
2.4.1 Fundamental principles
2.4.2 Human-AI interaction
2.4.3 Algorithmic biases
3 PRACTICAL FRAMEWORK CONDITIONS AND CHALLENGES USING LLMS IN INVESTMENT DECISIONS
3.1 LLMS IN CAPITAL MARKET-RELATED DECISION CONTEXTS
3.2 PERCEPTION OF AI AS AN OBJECTIVE AUTHORITY
3.3 PRACTICAL PROBLEM AREAS
3.3.1 Hallucinations and communication of uncertainty
3.3.2 Prompt-induced biases
3.3.3 Adoption of existing market narratives
3.4 REGULATORY AND LIABILITY IMPLICATIONS
4 SYSTEMATIC ANALYSIS OF BIAS DYNAMICS IN AI-ASSISTED INVESTMENT DECISIONS
4.1 OVERCONFIDENCE BIAS
4.1.1 Manifestations of overconfidence in capital market decisions
4.1.2 Potential attenuation through LLMs
4.1.3 Prompt experiment: mechanisms of reproduction and amplification
4.1.3.1 Prompt delivery
4.1.3.2 Qualitative assessment criteria
4.1.3.3 Expected outcomes
4.1.3.4 Comparison of overconfidence bias across conditions
4.1.3.5 Assessment of output consistency across prompt trials
4.1.3.6 Convergence of results and expected outcomes
4.1.4 Transformation dynamics
4.1.5 Strategic implications
4.2 CONFIRMATION BIAS
4.2.1 Confirmation-oriented decision logic in the investment process
4.2.2 Debiasing potential through structured information retrieval
4.2.3 Prompt experiment: reproduction of selective information search
4.2.3.1 Prompt delivery
4.2.3.2 Qualitative assessment criteria
4.2.3.3 Expected outcomes
4.2.3.4 Comparison of confirmation bias across conditions
4.2.3.5 Assessment of output consistency across prompt trials
4.2.3.6 Convergence of results and expected outcomes
4.2.4 Transformation through interactive dialogue structures
4.2.5 Strategic implications
4.3 ANCHORING BIAS
4.3.1 Reference points in capital market decisions
4.3.2 Potential neutralisation through alternative scenarios
4.3.3 Prompt experiment: context-dependent amplification through initial prompt information
4.3.3.1 Prompt delivery
4.3.3.2 Qualitative assessment criteria
4.3.3.3 Expected outcomes
4.3.3.4 Comparison of anchoring bias across conditions
4.3.3.5 Assessment of output consistency across prompt trials
4.3.3.6 Convergence of results and expected outcomes
4.3.4 Transformation dynamics
4.3.5 Strategic implications
4.4 COMPARATIVE SYNTHESIS AND STRUCTURAL CONSOLIDATION
4.4.1 Classification of bias dynamics
4.4.2 Emergence of autonomous algorithmic bias structures
4.4.3 Overarching strategic implications for investment decisions
5 CONCLUSION AND OUTLOOK
5.1 ANSWER TO THE RESEARCH QUESTION
5.2 KEY FINDINGS
5.3 LIMITATIONS
5.4 DIRECTIONS FOR FUTURE RESEARCH
Objectives & Focus Topics
The primary objective of this master's thesis is to examine systematically how Large Language Models (LLMs), exemplified by ChatGPT, influence cognitive biases in capital market investment decisions made under conditions of uncertainty. Moving beyond a simplistic dichotomy of whether artificial intelligence merely improves or worsens decision-making, the study investigates the underlying interaction mechanisms between human cognition and algorithmic systems to determine whether LLMs attenuate, reproduce, or transform overconfidence bias, confirmation bias, and anchoring bias.
- Behavioural finance foundations, bounded rationality, and decision-making under uncertainty in financial markets.
- The operational characteristics of LLMs as decision support systems, linguistic fluency, and perceived algorithmic authority.
- Empirical qualitative prompt experiments using ChatGPT across neutral, bias-inducing, and debiased conditions based on a Siemens Energy case study.
- Mechanisms of bias attenuation, reproduction, and the emergence of hybrid transformation dynamics.
- Autonomous algorithmic bias structures, regulatory governance, and strategic implications for individual and institutional investors.
Excerpt from the Book
4.4.1 Classification of bias dynamics
The analysis shows three recurring bias dynamics: attenuation, reproduction, and transformation. These dynamics are not mutually exclusive, since they can appear in the same decision environment, depending on how the investor uses the LLM and how critically the output is evaluated.
Attenuation describes cases in which LLMs reduce biased judgement. This can happen when the model helps the user structure information, compare alternatives, identify missing assumptions, or generate counterarguments. In relation to overconfidence, this may mean that the model makes downside risks and alternative scenarios more visible. Regarding confirmation bias, the user may receive contradictory evidence rather than only supportive arguments. For anchoring, it may mean that a reference point is questioned instead of accepted as the centre of the analysis. However, attenuation depends on a deliberate form of interaction. The user has to ask for uncertainty, weaknesses, and alternative perspectives. If the prompt is neutral and critical, the LLM can support reflection. If the prompt is one-sided, the same system may move in the opposite direction.
Reproduction describes cases in which LLMs continue or strengthen existing biases. This happens when the model adopts the assumptions already contained in the prompt. A user who asks for reasons why an investment thesis is strong may receive a coherent answer that supports this view. The model may not create the original bias, but it can translate it into a structured and persuasive explanation. This is relevant for all three biases. Overconfidence can be reinforced through confident AI-generated arguments, confirmation bias through selective support, and anchoring through repeated attention to a reference point.
Transformation goes one step further. LLMs do not simply weaken or strengthen traditional biases, but change their structure. Overconfidence may shift from confidence in one’s own judgement to confidence in AI-supported reasoning. Confirmation bias may shift from selective information search to selective dialogue with the model. Anchoring may shift from external market reference points to reference points generated or stabilised during the conversation.
These three dynamics show that LLM-supported investment decisions should be understood as hybrid decision processes. The relevant question is not only whether the investor is biased or whether the model is biased, but how human assumptions and model outputs interact. Bias can emerge from this interaction even when neither side alone fully explains the result.
Summary of Chapters
1 INTRODUCTION: Introduces the research problem, outlines how generative AI intersects with financial decision-making under uncertainty, establishes the core research questions, and details the deductive qualitative methodology combining a PRISMA-guided systematic literature review with controlled prompt experiments.
2 THEORETICAL FOUNDATIONS: Explores bounded rationality, heuristics, and reflexivity in capital markets, conceptualises overconfidence, confirmation, and anchoring biases within behavioural finance, and examines the fundamental architecture and cognitive offloading dynamics of LLMs as decision support systems.
3 PRACTICAL FRAMEWORK CONDITIONS AND CHALLENGES USING LLMS IN INVESTMENT DECISIONS: Analyses real-world applications in market analysis, the risks of perceived AI objectivity and linguistic fluency, practical challenges such as hallucinations and prompt-induced distortions, and the ensuing regulatory, oversight, and liability implications.
4 SYSTEMATIC ANALYSIS OF BIAS DYNAMICS IN AI-ASSISTED INVESTMENT DECISIONS: Delivers detailed qualitative prompt-based experiments using ChatGPT across overconfidence, confirmation, and anchoring biases in a Siemens Energy scenario, synthesising how algorithmic responses attenuate, reproduce, or transform human biases into hybrid cognitive structures.
5 CONCLUSION AND OUTLOOK: Concludes by answering the primary research question, outlining core findings regarding the conditional nature of LLM outputs and prompt literacy, addressing conceptual and methodological limitations, and proposing directions for future empirical and governance research.
Keywords
Artificial Intelligence, Large Language Models, ChatGPT, Behavioural Finance, Cognitive Biases, Overconfidence Bias, Confirmation Bias, Anchoring Bias, Decision Support Systems, Algorithmic Bias, Prompt Engineering, Capital Markets, Investment Decisions, Bounded Rationality, Hybrid Cognition
Frequently Asked Questions
What is the core focus of this thesis?
The thesis investigates how Large Language Models (LLMs), such as ChatGPT, influence cognitive biases when used as decision-support tools for capital market investment decisions under conditions of uncertainty.
Which specific cognitive biases are investigated?
The work concentrates on three major biases from behavioural finance: overconfidence bias, confirmation bias, and anchoring bias, chosen because they influence different stages of the investment decision process.
What is the primary research question?
The study examines to what extent the use of LLMs as decision support under uncertainty alters the dynamics of cognitive biases in capital market investment decisions by generating mechanisms of attenuation, reproduction, or transformation.
Which scientific methodology was applied?
A deductive and qualitative research design was utilized, consisting of a PRISMA-compliant systematic literature review across academic databases coupled with structured qualitative prompt experiments in ChatGPT Version 5.5.
What does the empirical analysis in the main body entail?
The analytical core evaluates ChatGPT's responses across neutral, bias-inducing, and debiased prompt conditions using a realistic corporate case study of Siemens Energy, assessing uncertainty communication, risk depth, evidence balance, and scenario generation.
What are the primary keywords that characterize the work?
Key terms include Artificial Intelligence, Large Language Models, ChatGPT, Behavioural Finance, Cognitive Biases, Overconfidence, Confirmation Bias, Anchoring Bias, Decision Support Systems, and Prompt Engineering.
How does the author define the "transformation" of cognitive biases by AI?
Transformation occurs when human-AI interaction alters the nature of classical biases—such as shifting overconfidence from personal self-assessment to trust in algorithmic authority, turning confirmation bias into an interactive feedback loop, or replacing market reference points with synthetic, AI-generated anchors.
Do LLMs automatically improve or worsen investment decision-making?
No. The research shows that LLM effects are non-deterministic and strictly conditional; an LLM can attenuate biases when prompted to critically test assumptions and seek counterarguments, but will reproduce and reinforce biases when presented with leading, one-sided prompts.
What practical recommendations are offered for investors and institutions?
Investors must preserve analytical friction by designing prompts that demand downside risks and failure conditions, while financial institutions need clear governance frameworks that enforce human oversight and prevent treating fluent AI narratives as verified financial advice.
- Quote paper
- Nino Fabrizio Hein (Author), 2026, Capital Market Investment Decisions under Artificial Intelligence Assistance, Munich, GRIN Verlag, https://www.grin.com/document/1771795
