AI recommendations nudge human judgment: LLM-generated numeric guidance anchors decisions and loss-framed outputs increase risk sensitivity, while higher trust in the AI predicts greater reliance on its suggestions.

Main Finding

LLM-generated recommendations materially shape human decisions: numerical anchors produced by LLMs shift choices (anchoring), gain-versus-loss framing from the LLM alters risk preferences, and people show stronger sensitivity to losses than to equivalent gains (consistent with Prospect Theory). Self-reported trust in the LLM significantly predicts greater influence and reliance on its recommendations.

Key Points

• Research question: Do LLM outputs create anchoring effects and interact with framing and trust to shape decision making?
• Sample & design: Between-subjects experiment with 251 participants.
• Manipulations: numeric anchor (AI recommendation) and outcome framing (gain vs loss).
• Main outcomes: investment / risk-choice responses, measures of AI influence/reliance, and AI trust.
• Results summary:
  • Exposure to LLM numerical recommendations produced anchoring on participants' decisions.
  • Gain/loss framing by the LLM significantly altered risk-taking (risk-averse in gains; more risk-seeking in losses).
  • Loss aversion observed—losses carried greater weight than equivalent gains.
  • Higher AI trust significantly predicted stronger influence and reliance on LLM recommendations.
• Analytical methods included descriptive stats, reliability analysis, independent and paired t-tests, chi-square tests, correlations and regressions, and ANOVA.
• The authors situate findings within Prospect Theory, arguing LLMs can set or shift reference points and thereby alter economic choices.

Data & Methods

• Sample: 251 participant responses (between-subjects).
• Experiment: Participants were assigned to conditions varying AI-provided numerical recommendations (anchors) and framing (gain vs loss). They then made investment or risk-related decisions; additional survey items captured AI trust and self-reported AI influence/reliance.
• Measures: decision choices (primary), AI trust scale, AI influence/reliance indicators (self-report); loss-aversion inferred from choices across framed conditions.
• Statistical approach: reliability checks for scales; inferential tests comparing groups (independent t-tests, paired t-tests), association tests (chi-square), relationships and predictions (correlation, regression), and variance analyses (ANOVA).
• Theoretical lens: Prospect Theory (reference dependence, loss aversion, framing effects).
• Reporting note: the paper reports significance of effects but does not provide full effect-size breakdowns in the excerpt (readers should check the full paper for magnitudes, CIs, and robustness checks).

Implications for AI Economics

• Behavioral influence of AI matters for markets and consumer welfare:
  • LLM-produced anchors can sway investment decisions and valuations, potentially amplifying herding or coordination effects when many users receive similar AI outputs.
  • Framing by LLMs can systematically shift risk-taking across consumer populations, affecting demand for risky assets, insurance purchase behavior, or responses to policy communications.
• Modeling and forecasting:
  • Economic models that assume independent information processing should incorporate reference-point shifts and automation-driven anchoring when AI advice is widespread.
  • Demand and adoption models should include endogenous trust dynamics: higher trust increases reliance, changing the propagation and persistence of AI-originated signals.
• Policy and regulation:
  • Disclosure standards and transparency (e.g., stating that a recommendation is AI-generated, showing reasoning or uncertainty) may be necessary to mitigate undesired anchoring and automation bias.
  • Consumer protection frameworks should consider framing effects from AI (e.g., in fintech, robo-advice) and require checks against manipulative framing.
• Design and market design interventions:
  • LLMs and interfaces can be engineered to reduce anchoring (e.g., present ranges, counterfactuals, uncertainty, debiased prompts) and to avoid manipulative framing.
  • Firms offering AI advice should monitor how default presentations of recommendations affect client choices and potential systemic risk.
• Research and measurement priorities:
  • Need for effect-size estimates, heterogeneous-treatment analyses (by expertise, stake size, demographic), and field studies with real financial stakes to quantify welfare impacts.
  • Investigate long-run dynamics: does repeated exposure to consistent AI anchors change reference points permanently, and how do competing AI sources interact?
• Broader economic concerns:
  • If LLMs converge on similar recommendations, anchoring could increase price/cohort volatility or slow information incorporation.
  • Loss-averse responses to AI framing could bias aggregate investment toward safer assets in some contexts, with implications for capital allocation and market liquidity.

Limitations to consider (from available information): moderate sample size, reliance on an experimental (likely hypothetical) investment task and self-reported influence/trust; effect magnitudes and robustness checks are not detailed in the excerpt. For policy or market conclusions, follow-up replications, larger/stakeful field experiments, and disclosure of effect sizes are recommended.