Decision-makers report greater optimism and trust toward machine forecasters than human analysts, but a pervasive ambiguity-insensitivity—stronger among the financially literate—blunts the influence of beliefs on trust.
Trusting human versus machine predictions as a decision under ambiguity
In an incentivized lab experiment, participants are more optimistic and therefore more trusting of ML forecasters than human forecasters, but ambiguity-insensitivity—which increases with financial literacy—reduces the extent to which beliefs translate into trust.
Abstract We examine how decision-makers’ (DMs’) ambiguity attitudes shape trust for two different sources of financial forecasting: human or machine learning (ML). In an incentivized laboratory experiment, we measure participants’ ambiguity attitudes and optimism regarding forecast accuracy for both sources. Our results reveal that DMs are similarly ambiguity-seeking and ambiguity-generated insensitive (“a-insensitive”; i.e., they insufficiently discriminate between changes in the likelihood of prediction accuracy), regardless of the analyst type. DMs hold more optimistic beliefs about the accuracy of ML analysts, which predicts higher trust in ML analysts over human analysts. However, DMs who are more a-insensitive are less likely to incorporate their beliefs into their trust. DMs’ a-insensitivity increases with financial literacy, suggesting that financially literate DMs perceive greater ambiguity in prediction accuracy. Our findings demonstrate that a-insensitivity acts as a cognitive barrier between beliefs and trust.
Summary
Paper: Diecidue, Guecioueur & Xia (2026). "Trusting human versus machine predictions as a decision under ambiguity." Journal of Risk and Uncertainty. https://doi.org/10.1007/s11166-026-09485-x
Main Finding
Decision-makers (DMs) form more optimistic beliefs about ML forecasters than human analysts and are therefore more willing to trust ML forecasts — but this effect is weakened by "a‑insensitivity" (insensitivity to changes in likelihoods). A‑insensitivity acts as a cognitive barrier: when DMs are more a‑insensitive they incorporate their beliefs about accuracy into trust decisions less. Financial literacy is associated with higher a‑insensitivity, consistent with a‑insensitivity reflecting perceived ambiguity rather than mere cognitive deficits.
Key Points
- Trust split into two concepts:
- Trust decisions: incentivized choices to rely on an analyst.
- Trust attitudes: self-reported evaluations of an analyst.
- Participants showed similar ambiguity attitudes for human and ML analysts (no major source-dependent difference).
- Participants held more optimistic beliefs about ML analyst accuracy; optimism predicts greater choice of ML over human forecasts.
- A‑insensitivity (low discrimination across likelihood levels) moderates belief-to-trust mapping: higher a‑insensitivity reduces the extent to which optimistic beliefs translate into trusting choices.
- Financially literate participants exhibited greater a‑insensitivity toward analyst accuracy; interpreted as literate DMs perceiving more ambiguity in forecasting accuracy rather than having worse discrimination ability.
- The authors pre-registered hypotheses and tests (link in paper).
Data & Methods
- Design: Incentivized laboratory experiment where participants choose between one-month-ahead price/return forecasts produced by (a) real human equity analysts (I/B/E/S forecasts) and (b) a specially designed state‑of‑the‑art ML analyst. No prior performance information was provided to participants.
- Measurement approach:
- Elicited “matching probabilities” (per Baillon et al. 2018b) for six events partitioning the analyst accuracy rate: three single events (low [0,40), moderate [40,80], high (80,100]) and three composite unions.
- From matching probabilities constructed two orthogonal ambiguity indices:
- Ambiguity aversion index bk = 1 − ms k − mc k (ms = average single-event match prob; mc = average composite-event match prob). bk ∈ [−1,1]; positive indicates aversion.
- A‑insensitivity index ak = 3 × (1/3 − (mc k − ms k)). Larger ak means greater insensitivity to likelihood changes.
- Belief optimism sk derived following Li et al. (2019) — an ambiguity‑neutralized relative optimism measure built from the matching probabilities (paper used weak monotonicity because set‑monotonicity violations were frequent).
- Statistical tests: Examined relationships between optimism, bk and ak, and both trust attitudes and incentivized trust decisions; tested moderation of belief effect by a‑insensitivity; analyzed covariates including financial literacy and familiarity with analysts.
- Methodological notes/limitations reported by authors:
- High violation rate (≈37.9%) of set‑monotonicity required using weak monotonicity for belief extraction; robustness checks reported in appendices.
- Laboratory (artificial) setting and one‑month horizon; external validity caveats noted.
Implications for AI Economics
- Demand/adoption friction for ML in finance can come from cognitive attitudes toward ambiguity, not only from objective performance. Even when users are optimistic about ML accuracy, a‑insensitivity may block adoption.
- Financial literacy has a nuanced role: more literate users may perceive more ambiguity in forecasters’ accuracy (raising a‑insensitivity), which can slow adoption of algorithmic forecasts despite greater ability to evaluate models. This complicates assumptions that literacy uniformly increases technology uptake.
- Product and market design:
- Transparency and calibrated performance disclosures that reduce perceived ambiguity (not only increase raw accuracy) may increase translation of positive beliefs into actual use.
- Explainability tools or easily interpretable performance metrics that reduce a‑insensitivity could increase trust decisions even when beliefs are positive.
- Reputation-building and repeated-feedback mechanisms (field testing, track records) can lower perceived ambiguity and help beliefs drive behavior.
- Policy and regulation:
- Regulators/market designers aiming to foster responsible AI adoption in finance should consider interventions targeting perceived ambiguity (standardized reporting, auditability, model validation disclosures) rather than focusing solely on average performance.
- Directions for further research:
- Field experiments and longitudinal studies to test external validity (real investors, actual trading).
- Heterogeneity analyses: how professional status, domain expertise, or prior experience with algorithms moderates the a‑insensitivity effect.
- Testing interventions (explainability, calibrated forecast histories, interactive model summaries) to reduce a‑insensitivity and improve belief-to-action translation.
If you want, I can (a) extract the main equations and indexes in a one‑page technical note, (b) sketch experiment flow and incentives, or (c) propose a field experiment to test interventions that reduce a‑insensitivity in professional investors. Which would be most useful?
Assessment
Paper Typequasi_experimental
Evidence Strengthmedium — Behavioral, incentivized measures in a controlled lab setting provide credible evidence on beliefs and trust, and the direct comparison between human and ML forecasts supports internal validity; however, many key mediators (ambiguity attitudes, financial literacy) are measured not randomized, sample details and size are not provided in the abstract, and external validity is limited to the lab forecasting task and participant pool.
Methods Rigormedium — Use of incentivized belief elicitation and a controlled experimental environment are solid methodological choices; but the study appears to rely on measured (not manipulated) ambiguity attitudes, potentially limiting causal interpretation of those effects, and the abstract does not report robustness checks, sample representativeness, or pre-registration.
SampleIncentivized laboratory participants whose demographics/sampling frame and sample size are not specified in the abstract; participants completed forecasting tasks, had ambiguity attitudes and optimism measured, reported trust in human vs ML analysts, and answered financial literacy items.
Themeshuman_ai_collab adoption
IdentificationIncentivized laboratory experiment presenting participants with forecasts from two analyst types (human vs ML) and eliciting beliefs, ambiguity attitudes, optimism, and trust; identification of analyst-type effects appears to rely on experimental presentation (randomization or controlled comparison of analyst condition), while relationships between measured ambiguity attitudes/financial literacy and trust are correlational since attitudes were elicited rather than exogenously manipulated.
GeneralizabilityLab experimental setting may not reflect real-world high-stakes financial decision-making, Unclear participant population (e.g., students vs general adult population) limits external validity, Single domain (financial forecasting) — results may not generalize to other tasks or industries, Short-term, one-off decisions in the lab may not capture longitudinal trust/adoption dynamics, The ML analyst is a laboratory representation and may not reflect deployed, production-grade AI systems
Claims (5)
| Claim | Direction | Confidence | Outcome | Details |
|---|---|---|---|---|
| Decision-makers (DMs) are similarly ambiguity-seeking and ambiguity-generated insensitive (a-insensitive) regardless of whether the analyst is human or a machine learning (ML) model. Decision Quality | null_result | high | ambiguity attitude (ambiguity-seeking and a-insensitivity) |
0.48
|
| Decision-makers hold more optimistic beliefs about the accuracy of ML analysts than about human analysts, and this greater optimism predicts higher trust in ML analysts relative to human analysts. Decision Quality | positive | high | optimism about forecast accuracy and trust in analyst |
0.48
|
| Decision-makers who are more a-insensitive are less likely to incorporate their beliefs about forecast accuracy into their trust judgments. Decision Quality | negative | high | degree to which beliefs predict trust (belief–trust linkage) |
0.48
|
| A-insensitivity increases with financial literacy, suggesting financially literate decision-makers perceive greater ambiguity in prediction accuracy. Skill Acquisition | positive | high | a-insensitivity (ambiguity-generated insensitivity) |
0.48
|
| A-insensitivity acts as a cognitive barrier between beliefs and trust (i.e., it reduces the extent to which beliefs about forecast accuracy are translated into trust). Decision Quality | negative | high | belief-to-trust translation (strength of relationship between beliefs and trust) |
0.48
|