In Algorithms We Trust? The Behavioural Economics of Trusting AI over Humans
From Algorithm Aversion to Algorithm Appreciation — and Back
5 min read · 932 words
Abstract
[Draft] When a forecast, a diagnosis, or a bail decision can come from either a person or an algorithm, whom do we trust — and should we? The experimental literature returns two contradictory answers: people abandon algorithms after seeing them err, even when the algorithm demonstrably outperforms the human (algorithm aversion), yet in other settings they weight identical advice more heavily when told it comes from an algorithm (algorithm appreciation). This article reconciles the two findings by treating trust in AI not as a fixed attitude but as the output of the same behavioural machinery documented for interpersonal trust: error salience, perceived control, task framing, expertise, and biased belief updating. Field evidence from bail decisions and radiology shows the stakes: judges respond to noise as if it were signal, and radiologists underweight AI predictions, so human–AI combinations can underperform either alone. The practical question is not whether to trust algorithms but how to design the choice architecture of trust.
1. From Trusting People to Trusting Machines
Plan: open with the trust-game definition carried through this series — trust as willingly accepted vulnerability on the expectation of competence and benevolence (Article 1). Ask what survives the substitution of an algorithm for the human trustee: competence expectations transfer, but benevolence, reciprocity and second-order beliefs do not. Introduce Glikson and Woolley’s (2020) review as the organising framework: cognitive versus emotional trust in AI, with tangibility, transparency, reliability and task characteristics as its empirical drivers.
2. Algorithm Aversion: Abandoning the Better Forecaster
Plan: present Dietvorst, Simmons and Massey (2015) — participants who saw an algorithm err abandoned it in favour of a human they had also seen err more. The driver is error salience, not error rates: an algorithm’s mistake is read as diagnostic (“it is broken”), a human’s as forgivable noise. Connect to Article 2’s asymmetric-updating evidence (Möbius et al., 2011). Close with the boundary condition: granting users even slight ability to modify the algorithm’s output restores use (Dietvorst, Simmons and Massey, 2018) — control, not accuracy, is the lever.
3. Algorithm Appreciation: The Opposite Result
Plan: present Logg, Minson and Moore (2019) — identical advice is weighted more when labelled algorithmic, at least by laypeople who have not yet seen the algorithm fail; experts discount algorithmic advice. Reconcile: aversion appears after observed errors and on subjectively framed tasks; appreciation appears before errors on objectively framed tasks. Since task framing is manipulable, it is both a design lever and a rhetorical one.
4. Field Evidence: What Is Actually at Stake
Plan: two field results anchor the welfare argument. Kleinberg et al. (2017; QJE 2018) show machine predictions of pre-trial flight risk could reduce crime by roughly a quarter at unchanged jailing rates — and that judges respond to noise as if it were signal. Agarwal et al. (2023) show that giving radiologists AI predictions does not improve average performance: they underweight the AI and treat its signal as independent of their own information, so the human–AI combination can underperform either alone. The bottleneck is the belief-updating machinery of Article 2 — conservatism and correlation neglect (Enke and Zimmermann, 2019; Benjamin, 2019).
5. Synthesis: Trust in AI as Choice Architecture
Plan: pull the threads together. Trust in algorithms is governed by the same behavioural forces as trust in people — there is no separate psychology of machines. Design implications: show track records rather than single errors; grant bounded control; frame tasks deliberately; and where updating errors are incorrigible, delegate cases wholly to human or machine rather than blending (Agarwal et al.’s delegation result). Normative coda: distinguish costly aversion (bail, diagnosis) from warranted scepticism (opaque or unvalidated systems). Close by linking back to Article 3: trusting an algorithm is the short-horizon cousin of trusting an institution with the future.
References (9)
- 1. Agarwal, N., Moehring, A., Rajpurkar, P., & Salz, T. (2023, rev. 2025). Combining human expertise with artificial intelligence: Experimental evidence from radiology. NBER Working Paper 31422. [I05]
- 2. Benjamin, D. J. (2019). Errors in probabilistic reasoning and judgment biases. In Handbook of Behavioral Economics, Vol. 2. NBER Working Paper 25200. [E01]
- 3. Dietvorst, B. J., Simmons, J. P., & Massey, C. (2015). Algorithm aversion: People erroneously avoid algorithms after seeing them err. Journal of Experimental Psychology: General, 144(1), 114–126. [I01]
- 4. Dietvorst, B. J., Simmons, J. P., & Massey, C. (2018). Overcoming algorithm aversion: People will use imperfect algorithms if they can (even slightly) modify them. Management Science, 64(3), 1155–1170. [to add to kit if used]
- 5. Enke, B., & Zimmermann, F. (2019). Correlation neglect in belief formation. Review of Economic Studies, 86(1), 313–332. [E02]
- 6. Glikson, E., & Woolley, A. W. (2020). Human trust in artificial intelligence: Review of empirical research. Academy of Management Annals, 14(2), 627–660. [I03]
- 7. Kleinberg, J., Lakkaraju, H., Leskovec, J., Ludwig, J., & Mullainathan, S. (2017). Human decisions and machine predictions. NBER Working Paper 23180; published in Quarterly Journal of Economics, 133(1), 237–293 (2018). [I04]
- 8. Logg, J. M., Minson, J. A., & Moore, D. A. (2019). Algorithm appreciation: People prefer algorithmic to human judgment. Organizational Behavior and Human Decision Processes, 151, 90–103. [I02]
- 9. Möbius, M. M., Niederle, M., Niehaus, P., & Rosenblat, T. S. (2011). Managing self-confidence: Theory and experimental evidence. NBER Working Paper 17014. [E03]