Do people accept algorithmic decisions in government? What surveys and experiments show

Governments increasingly use algorithms to score risk, triage applications and decide entitlements, but the legitimacy of those decisions depends on whether the people affected accept them. This study reviews two bodies of evidence on that question: behavioural experiments on "algorithm aversion" and "algorithm appreciation", and survey and conjoint experiments on algorithmic decision-making in the public sector, together with large public-opinion surveys in the United States and the United Kingdom. The evidence does not support a simple claim that people reject algorithms. People often trust algorithmic advice, sometimes more than human advice, especially for tasks seen as mechanical and when they are not themselves experts. But support falls sharply when the algorithm decides rather than advises, when the decision concerns personal circumstances such as welfare or justice, and when there is no route to a human. Across studies, the features that most consistently increase acceptance are the ability to appeal, meaningful human involvement, explanations of individual decisions, demonstrated accuracy, and trust in the organisation that deploys the system. In the UK, concern about AI deciding welfare eligibility rose from 44% to 59% between 2022 and 2024, and the share saying that appeal procedures would make them more comfortable rose from 59% to 65%.

Key findings

• Aversion and appreciation are both real; context decides. People abandon algorithms more quickly than humans after seeing them err (Dietvorst, Simmons and Massey), yet lay people often follow algorithmic advice more than human advice (Logg, Minson and Moore). Experts and people judging their own estimates show less appreciation.
• Advising is accepted more than deciding. In a preregistered experiment on EU policy-making (n=572), AI as the sole decision-maker was seen as illegitimate, while hybrid human-AI decision-making was rated no differently from the status quo on process and outcome legitimacy (Starke and Lünich).
• Personal and moral decisions attract the most resistance. For tasks requiring "human skills", algorithmic decisions were seen as less fair and trustworthy (Lee). In the US, 56% found criminal risk scores for parole unacceptable and 67-68% rejected automated video-interview analysis and personal-finance scoring (Pew, 2018).
• Appeals and human involvement matter most. In UK conjoint experiments on visa and parking permits, accuracy, an appeals system, transparency, lower cost, no data sharing and no private-company involvement all raised both acceptance and perceived fairness (Horváth and colleagues). In the UK public's own ranking, laws and regulation (72%) and appeal procedures (65%) top the list of what would increase comfort with AI (Ada Lovelace Institute and Alan Turing Institute, 2024/25).
• Explanations help, but access to code does not. Explaining why a decision was made increases trust more than making the algorithm accessible, and the effect varies by context (Grimmelikhuijsen).
• Who deploys matters as much as what is deployed. In Germany, acceptance of predictive policing and cancer-prediction algorithms depended more on trust in the deploying organisation and the importance of the problem than on the algorithm's measured performance (Wenzelburger and colleagues).
• Concern is rising. In the US, the share more concerned than excited about AI rose from 38% in December 2022 to 52% in August 2023; 62% of US adults in 2024 had little or no confidence in government to regulate AI effectively (Pew).

Research question and scope

The study asks: under what conditions do people accept decisions made or supported by algorithms in government, and what increases the legitimacy of such decisions? "Acceptance" includes preferences between human and algorithmic decision-makers, judgements of fairness and trust, and perceived legitimacy. The review covers behavioural experiments (mostly forecasting tasks), survey experiments and conjoint designs on public-sector scenarios, and large representative surveys. It focuses on the United States and Europe, where almost all the evidence originates; a 2022 systematic review found the evidence on fairness perceptions comes "almost exclusively from Western-democratic contexts".

Background

Two failure cases frame the debate in Europe and Australia. In the Netherlands, a risk-scoring and fraud-detection regime in childcare benefits wrongly treated thousands of families as fraudsters and led to the cabinet's resignation in January 2021. In Australia, the automated "Robodebt" scheme raised unlawful debts against welfare recipients, as a Royal Commission found in 2023. Both are documented elsewhere in this atlas. In the UK, the 2020 algorithm used to moderate school-leaving grades during the pandemic was abandoned after public protest. These cases share a profile: high-stakes decisions about individuals, weak routes to challenge, and systems whose logic was not explained to those affected.

The academic debate has two roots. Psychologists since Paul Meehl in the 1950s have shown that simple statistical rules often predict better than expert judgement, yet people resist using them, a phenomenon later named "algorithm aversion". Public-administration and legitimacy scholars ask a different question: not whether people use a forecast, but whether a decision taken by or with an algorithm is rightful. The second question draws on procedural justice theory, which holds that people accept unfavourable outcomes when they see the process as fair, transparent and respectful.

Method and sources

This is a narrative review of the strongest available evidence, chosen for design quality and relevance to government. It draws on two systematic reviews of algorithm aversion (Burton, Stein and Jensen: 61 articles; Mahmud and colleagues: 80 studies) and one of fairness perceptions (Starke, Baleis, Keller and Marcinkowski: 58 studies); individual experiments on public-sector scenarios in Europe and the United States; Pew Research Center surveys (2018, 2023, 2024); and the two waves of the Ada Lovelace Institute and Alan Turing Institute survey of the British and UK public (fieldwork November 2022 and October-November 2024). The OECD Survey on Drivers of Trust in Public Institutions (nearly 60,000 responses in 30 countries, late 2023) provides context on trust in government generally. Sample sizes and dates are given in the text so that readers can judge weight.

Findings

Aversion and appreciation

Dietvorst, Simmons and Massey coined "algorithm aversion" in five studies in which participants chose whether to tie their payment to a human's or a statistical model's forecasts. Participants who had seen the algorithm perform were less confident in it and less likely to choose it over an inferior human forecaster, "even among those who saw the algorithm outperform the human". The mechanism is asymmetric forgiveness: people lose confidence in algorithms more quickly than in humans after seeing the same mistake.

Logg, Minson and Moore found the opposite in six experiments: lay people adhered more to advice when they thought it came from an algorithm than from a person, in tasks ranging from estimating weights to forecasting song popularity. Appreciation waned when people chose between an algorithm's estimate and their own, and among experienced professionals, who relied less on algorithmic advice than lay people did, at a cost to their accuracy.

Two systematic reviews reconcile these results. Burton, Stein and Jensen, reviewing 61 peer-reviewed articles from 1950 to 2018, grouped the causes of aversion under expectations and expertise, decision autonomy, incentives, cognitive compatibility and divergent rationalities. Mahmud, Islam, Ahmed and Smolander, reviewing 80 empirical studies, grouped influencing factors as algorithm, individual, task and high-level factors, and noted that task and high-level (social and institutional) factors are under-studied. The common lesson is that aversion is conditional: it depends on who is judging, what the task is, and whether people keep some control.

Survey evidence points the same way. Araujo, Helberger, Kruikemeier and de Vreese, in a scenario-based survey experiment with a national sample of 958 respondents across media, public-health and judicial decisions, found respondents broadly concerned about the risks of automated decision-making and mixed on its fairness and usefulness at the societal level, yet for specific decisions they often rated AI on a par with, or better than, human experts.

For public policy, Kennedy, Waggoner and Ward ran experiments on forecasting political events and criminal recidivism. "Contrary to previous literature on algorithm aversion", they found high trust in algorithms relative to other sources of advice, even with minimal information about the algorithm. A conjoint experiment showed that risk aversion, the size of the training data, a human in the loop and the developer's reputation shaped preferences.

Who decides: legitimacy of hybrid and fully automated decisions

Starke and Lünich compared three arrangements for EU policy-making in a preregistered online experiment (n=572): politicians alone, AI alone, and politicians working with AI. The existing arrangement was rated most participatory and accessible (input legitimacy). On the process (throughput) and on outcomes (output legitimacy) there was no difference between the status quo and the hybrid arrangement, but respondents saw AI as the sole decision-maker as illegitimate. The finding is important for institution design: people distinguish sharply between AI as a tool of accountable officials and AI as a substitute for them.

Ingrams, Kaufmann and Jacobs reached a related result in a survey experiment on tax auditing: an AI-led decision was seen as lower in red tape than a human one, but also as less trustworthy. Efficiency and trust can move in opposite directions.

Lee's experiment on managerial decisions showed why task type matters. For "mechanical" tasks such as scheduling, algorithmic and human decisions were seen as equally fair and trustworthy. For tasks requiring "human skills", such as hiring and evaluation, algorithmic decisions were seen as less fair and trustworthy and provoked more negative emotion, attributed to the algorithm's perceived lack of intuition and to "the dehumanizing experience of being evaluated by machines". Binns and colleagues found that participants' sense of justice towards algorithmic decisions engaged the same dimensions as for human decisions, with concerns about arbitrariness, generalisation and dignity, summed up in the phrase that gives their paper its title: "It's reducing a human being to a percentage."

Not everyone prefers a human. Miller and Keiser found in a US survey experiment that Black respondents rated automated decision-making higher than police officers on fairness and preference when told that the police agency lacked passive representation; white respondents did not show this pattern. Where human decision-makers are distrusted, automation can look like an improvement.

What increases acceptance

Horváth, James, Banducci and Beduschi ran a preregistered conjoint experiment in the UK on AI processing of immigration-visa and parking-permit applications. More human involvement tended to increase acceptance, though perceived fairness was less affected, and when substantial human discretion was introduced in parking cases respondents preferred more limited human input. Human involvement did little to offset the AI's unfavourable attributes. By contrast, "high accuracy, the presence of an appeals system, increased transparency, reduced cost, non-sharing of data, and the absence of private company involvement" all raised both acceptance and perceived procedural fairness.

Grimmelikhuijsen tested two elements of transparency, accessibility (being able to see the algorithm) and explainability (being told why a decision was made), in a denied visa application and a suspicion of welfare fraud. Explainability had a more pronounced effect on trust than accessibility; transparency affected trust in the human decision-maker as well as in the algorithm; and the effects were not robust across the two contexts. Transparency "as accessibility" was insufficient.

Wenzelburger, König, Felfeli and Achtziger surveyed 2,661 respondents from a representative German online panel about predictive policing and skin-cancer prediction. Acceptance was "strongly shaped" by the personal importance of the problem and by trust in the organisations deploying the algorithm; the algorithm's objective performance mattered much less.

What the public says: large surveys

United States. Pew's 2018 survey of 4,594 adults found that 58% thought computer programs would always reflect some human bias. Majorities found three of four scenarios unacceptable: an automated personal-finance score (68% unacceptable), automated analysis of video job interviews (67%), and criminal risk assessment for parole decisions (56% unacceptable, 42% acceptable); 41% found automated résumé screening acceptable. Pew's August 2023 survey of 11,201 adults found 52% more concerned than excited about AI in daily life, up from 38% in December 2022. In August 2024, 51% of 5,410 adults were more concerned than excited, 62% had not too much or no confidence in government to regulate AI effectively, and 55% wanted more control over how AI is used in their lives.

United Kingdom. The Ada Lovelace Institute and Alan Turing Institute survey is the most detailed source on public-sector uses.

In 2024/25, 63% said they were not comfortable with AI making decisions that affect their lives; 62% said an explanation should always accompany an AI decision (including 37% who want humans to make and explain decisions); 71% of Black and Black British respondents were concerned about welfare-eligibility AI, against 59% nationally; and 50% felt their views were not represented in decisions about AI that affect their lives. In 2022/23, an independent regulator was the most popular choice (41%) for who should ensure AI is used safely; in 2024/25, when respondents could choose up to three options, it was chosen by 58%.

The contrast between facial recognition in policing (seen as beneficial by 91%) and welfare eligibility (47%) is instructive. Respondents appear to accept AI where it is seen as targeting others or as a tool for officials, and resist it where it decides their own entitlements.

What the evidence does not show

• Behaviour after real decisions. Most studies use hypothetical scenarios. There is little evidence on how people who actually received an automated decision, such as a benefit refusal, judge its legitimacy, comply or appeal.
• Long-run effects. Surveys show rising concern since 2022, but cannot separate reactions to AI in government from reactions to generative AI in general or to news events.
• Causal effects of safeguards in the field. Conjoint experiments show that appeals and explanations raise stated acceptance; whether introducing them in a real service raises trust has not been tested with field experiments.
• Global generalisation. The evidence base is overwhelmingly from the United States and Western Europe.
• The quality of acceptance. People may accept systems they should not, as the appreciation literature and the facial-recognition figures suggest. High acceptance is not evidence that a system is fair.

Implications for institution design

1. Keep a human decision-maker for consequential individual decisions and present the algorithm as advice. The evidence on hybrid legitimacy and on "human skills" tasks supports this more than any other design choice.
2. Make appeal to a human a right, not a courtesy. It is the feature most consistently associated with acceptance and is also the main safeguard against error.
3. Explain decisions, not just systems. Algorithm registers such as the UK's transparency records help oversight, but individual explanations matter more for trust.
4. Avoid private-sector data sharing and opaque supplier roles in high-stakes public uses, or justify them openly.
5. Invest in institutional trustworthiness. Acceptance depends on trust in the deploying agency; an untrusted agency will not be rescued by a better model.
6. Do not mistake acceptance for legitimacy. Systems should be evaluated for accuracy and fairness independently of how popular they are.

Open questions

• Do appeal rights increase trust when people know that few appeals succeed?
• How do people judge decisions in which AI and a human official disagree, and who should prevail?
• Does the rise in concern since 2022 reflect experience of AI in public services, or wider anxiety about generative AI?
• How do attitudes differ outside Western democracies, where trust in human officials may be lower?
• Can participatory processes, such as citizens' assemblies on algorithmic systems, raise legitimacy for non-participants?

Sources

1. Dietvorst, B. J., Simmons, J. P. and Massey, C. (2015). Algorithm aversion: People erroneously avoid algorithms after seeing them err. Journal of Experimental Psychology: General 144(1), 114-126.
2. Logg, J. M., Minson, J. A. and Moore, D. A. (2019). Algorithm appreciation: People prefer algorithmic to human judgment. Organizational Behavior and Human Decision Processes 151, 90-103.
3. Burton, J. W., Stein, M.-K. and Jensen, T. B. (2020). A systematic review of algorithm aversion in augmented decision making. Journal of Behavioral Decision Making 33(2), 220-239.
4. Mahmud, H., Islam, A. K. M. N., Ahmed, S. I. and Smolander, K. (2022). What influences algorithmic decision-making? A systematic literature review on algorithm aversion. Technological Forecasting and Social Change 175, 121390.
5. Starke, C., Baleis, J., Keller, B. and Marcinkowski, F. (2022). Fairness perceptions of algorithmic decision-making: A systematic review of the empirical literature. Big Data & Society 9(2).
6. Starke, C. and Lünich, M. (2020). Artificial intelligence for political decision-making in the European Union: Effects on citizens' perceptions of input, throughput, and output legitimacy. Data & Policy 2, e16.
7. Kennedy, R. P., Waggoner, P. D. and Ward, M. M. (2022). Trust in Public Policy Algorithms. The Journal of Politics 84(2), 1132-1148.
8. Lee, M. K. (2018). Understanding perception of algorithmic decisions: Fairness, trust, and emotion in response to algorithmic management. Big Data & Society 5(1).
9. Binns, R., Van Kleek, M., Veale, M., Lyngs, U., Zhao, J. and Shadbolt, N. (2018). 'It's Reducing a Human Being to a Percentage': Perceptions of Justice in Algorithmic Decisions. In Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems.
10. Miller, S. M. and Keiser, L. R. (2021). Representative Bureaucracy and Attitudes Toward Automated Decision Making. Journal of Public Administration Research and Theory 31(1), 150-165.
11. Ingrams, A., Kaufmann, W. and Jacobs, D. (2022). In AI we trust? Citizen perceptions of AI in government decision making. Policy & Internet 14(2), 390-409.
12. Grimmelikhuijsen, S. (2023). Explaining Why the Computer Says No: Algorithmic Transparency Affects the Perceived Trustworthiness of Automated Decision-Making. Public Administration Review 83(2), 241-262.
13. Horváth, L., James, O., Banducci, S. and Beduschi, A. (2023). Citizens' acceptance of artificial intelligence in public services: Evidence from a conjoint experiment about processing permit applications. Government Information Quarterly 40(4), 101876.
14. Wenzelburger, G., König, P. D., Felfeli, J. and Achtziger, A. (2024). Algorithms in the public sector. Why context matters. Public Administration 102(1), 40-60.
15. Araujo, T., Helberger, N., Kruikemeier, S. and de Vreese, C. H. (2020). In AI we trust? Perceptions about automated decision-making by artificial intelligence. AI & Society 35(3), 611-623.
16. Pew Research Center (2018). Public Attitudes Toward Computer Algorithms, 16 November 2018.
17. Pew Research Center (2023). Growing public concern about the role of artificial intelligence in daily life, 28 August 2023.
18. Pew Research Center (2025). How the U.S. Public and AI Experts View Artificial Intelligence, 3 April 2025.
19. Ada Lovelace Institute and The Alan Turing Institute (2023). How do people feel about AI? (2022/23 survey findings).
20. Ada Lovelace Institute and The Alan Turing Institute (2025). How do people feel about AI? (2024/25 survey findings).
21. OECD (2024). OECD Survey on Drivers of Trust in Public Institutions - 2024 Results. OECD Publishing.