Do people trust algorithms more than humans?
Lay people lean towards algorithmic advice, and one visible mistake reverses it. What the studies say about when people trust algorithms, and when they should.
Answer to the question
Do people trust algorithms more than humans?
Sometimes. Lay people in Logg and colleagues' 2019 experiments followed advice labelled as algorithmic more closely than advice labelled as human. After seeing an algorithm make a mistake, Dietvorst's participants preferred a human forecaster, and experienced forecasters relied on algorithmic advice less than lay people did. Task type and expertise change the answer.
Key takeaways
- Logg and colleagues (2019) found in six experiments that lay people moved further towards advice described as algorithmic than advice described as human: 0.45 against 0.30 in a first experiment with 202 participants.
- Dietvorst and colleagues (2015), in five studies, found that people who watched an algorithm forecast were less likely to choose it over a human forecaster, including when it had outperformed the human.
- A 2021 replication in the Journal of Management Control, with 1,840 participants, found 67 per cent chose the algorithm without performance feedback and 47 per cent after seeing it perform.
- A 2019 Journal of Marketing Research paper found people trust algorithms less for tasks they see as subjective, across more than 1,400 laboratory and 56,000 field participants.
- Kahneman and his co-authors report that across 136 studies the rule beat the professional in 63, tied in 65 and lost in 8.
Researchers who study judgment and decision making were asked to predict how people would treat advice described as coming from an algorithm, set against advice described as coming from a person. They predicted the person would win. In Jennifer Logg's experiments the algorithm won. Do people trust algorithms more than humans? Sometimes, and four conditions do most of the work in deciding when people trust algorithms: whether the person has watched the algorithm make a mistake, how subjective the task feels, how much expertise the person brings, and whether the rival is a stranger or the person's own judgement. The evidence is easiest to follow from the mistake outward, because that is where the research on aversion began.
What One Visible Mistake Does to Trust in an Algorithm
The study that named algorithm aversion is the place to start. Berkeley Dietvorst and his co-authors ran five studies in which participants saw an algorithm make forecasts, saw a human make forecasts, saw both, or saw neither. They then chose whose future predictions would determine their own payment.
Those who had watched the algorithm work were less confident in it and less likely to choose it over a human forecaster, including when they had seen it beat that human. The authors traced this to confidence: after seeing the same kind of error, people lost faith in the algorithm faster than in the person.
Six years after that 2015 paper, a 2021 experimental study in the Journal of Management Control reproduced the pattern with 1,840 participants. In its replication, 67 per cent chose the algorithm without performance feedback and 47 per cent after seeing it perform. Time pressure reduced the aversion, which the authors link to lower confidence in participants' own forecasts. A goal to do their best and control over the input data did not.
Where Lay People Trust Algorithms, and Who Stops
Logg and her co-authors ran six experiments, all paid, with a single measure: how far a person moves their own estimate towards the advice they are given. In the first, 202 participants estimated a stranger's weight from a photograph. When the advice was described as an algorithm's calculation, they moved 0.45 of the way towards it. When it was described as the average of past participants, they moved 0.30. The authors called the pattern algorithm appreciation. Comparable gaps appeared when people forecast the popularity of songs and the romantic attraction between strangers.
The result held whether the two sources of advice appeared side by side or separately. People in both groups still gave their own estimate more weight than the advice. The finding is therefore a smaller shortfall in taking advice, not the absence of one. With no performance history to go on, lay people trust algorithms a little more than they trust other people, and the first experiment is only one of several that show it.
Set side by side, the aversion and appreciation studies describe one pattern. People start out ready to trust algorithms a little more than other people, a modest lean. Evidence of its performance then lowers that lean, and the loss is larger than the same evidence costs a human adviser. This is our reading of the designs, and neither paper states it.
A 2019 study in the Journal of Marketing Research found that people trust algorithms less for tasks they regard as subjective than for tasks they regard as objective. Across four online laboratory studies with more than 1,400 participants and two online field studies with more than 56,000, raising a task's perceived objectivity, or giving the algorithm a more humanlike emotional character, increased reliance on it. The more objective a task looks, the more readily people trust algorithms with it.
What an AI Label Does to a Product shows the same split in shopping, where the label cost a dress far more than a plain T-shirt. How the advice is weighed once it arrives is the subject of How Much Advice People Actually Take.
The Logg experiments also found two places where the lean towards the algorithm faded. It weakened when people chose between an algorithm's estimate and their own, as opposed to an outside person's. It weakened further in a group with forecasting experience. The experienced professionals, who made forecasts on a regular basis, relied less on algorithmic advice than lay people did, and their accuracy suffered for it. The people with the most forecasting practice gave algorithms the least weight, so the lay finding and the expert finding point in opposite directions.
| Condition | Effect on trust in the algorithm | Where it was found |
|---|---|---|
| No performance history, lay judges | Advice labelled algorithmic is followed more closely | Logg and colleagues, six experiments, 2019 |
| Algorithm seen to err | Choice of the algorithm falls, even when it beat the human | Dietvorst and colleagues, five studies, 2015; replicated 2021 |
| Task seen as subjective | Reliance falls | Journal of Marketing Research study, 2019 |
| Forecasting expertise, or a choice against one's own estimate | The lean towards the algorithm weakens | Logg and colleagues, 2019 |
What the Accuracy Record Says About the Whole Question
Whether people should trust algorithms is a separate question from whether they do. The accuracy record is more one-sided than the trust record, and Noise, by Daniel Kahneman and his co-authors, addresses it directly. The authors review a body of work on clinical versus mechanical prediction, begun in the 1950s, that compares professionals with simple rules built on the same information. In a review of 136 studies they report, the rule did better in 63, the two tied in 65, and the professional won in 8. Their explanation is that a rule gives the same answer to the same case every time, while a person does not.
They report a further result from later work: a model built only from a judge's own past judgments predicted outcomes better than the judge did, and in one replication it did so for every one of 98 participants. The model had kept the judge's consistent preferences and lost the judge's arbitrary variation. For hiring decisions, they give correlations with later performance of .28 for human judges and .44 for mechanical prediction.
The authors also say why this has changed so little in practice. A judge works towards a feeling that the evidence fits together, and that feeling is mistaken for confidence in being right. A rule cannot supply it, and the pull towards intuition is strongest where the situation is most uncertain. That suggests aversion should bite hardest in the same places, though neither literature tests it. Two Readings of the Same Case covers the same ground, with two readers of one case reaching different conclusions.
The authors of Noise conclude that people will keep deciding wherever a formula falls short, as it does in most real settings, so the practical work lies in improving how people judge. We agree.
Common questions
Do people trust algorithms more than humans?
Sometimes. In Logg and colleagues' 2019 experiments, lay people gave more weight to advice labelled as algorithmic than to advice labelled as human. After seeing an algorithm err, though, people in Dietvorst's studies preferred a human forecaster, and experienced forecasters relied on algorithmic advice less than lay people did.
What is algorithm aversion?
Algorithm aversion is the tendency to stop relying on an algorithm after seeing it make a mistake, even when it is more accurate than a person. Dietvorst and colleagues named it in 2015, and a 2021 replication confirmed the direction of the finding.
What is algorithm appreciation?
Algorithm appreciation is the tendency of lay people to follow advice more closely when it is described as coming from an algorithm than from a person. Logg and colleagues reported it in 2019, and it weakened among experts and when people compared the algorithm with their own judgement.
Sources
Kahneman, D., Sibony, O., & Sunstein, C. R. (2021). Noise: A Flaw in Human Judgment. London: William Collins. Seed book; vault hub Books/Noise.
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. doi:10.1016/j.obhdp.2018.12.005 https://doi.org/10.1016/j.obhdp.2018.12.005
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. doi:10.1037/xge0000033 https://doi.org/10.1037/xge0000033
Castelo, N., Bos, M. W., & Lehmann, D. R. (2019). “Task-dependent algorithm aversion.” Journal of Marketing Research, 56(5), 809-825. doi:10.1177/0022243719851788 https://doi.org/10.1177/0022243719851788
Jung, M., & Seiter, M. (2021). “Towards a better understanding on mitigating algorithm aversion in forecasting: an experimental study.” Journal of Management Control, 32(4). doi:10.1007/s00187-021-00326-3 https://doi.org/10.1007/s00187-021-00326-3