Decision support

Decision support is the set of tools and practices that help an individual make a better decision themselves, rather than deciding for them: structured decision aids, coaching, techniques to counter known biases, ways of clarifying what one values, ways of weighing advice, and, more recently, software that advises or questions the decision-maker. Its central disputes are what counts as a “better” decision when the right answer is unknown, whether the techniques transfer from laboratory and clinic to everyday choices, and whether AI assistance improves decisions at all, since combinations of people and AI often do no better than the stronger party alone. The biases and decision theories that motivate the field are covered in Judgment and decision-making.

History

To do: the origins of decision aids, of debiasing programmes and of judge-advisor research.

Core concepts

What counts as a better decision

In health decision-aid research, decision quality is defined as a choice that is congruent with informed and considered values (International Patient Decision Aids Standards Collaboration, 2024). It is measured as a composite: a knowledge score, and whether the treatment received matches the patient’s stated goals, the latter inferred by regressing treatment on goals (Massachusetts General Hospital Health Decision Sciences Center, n.d.). The two halves have unequal support. In a validation for breast cancer chemotherapy decisions, the knowledge component was reliable, but the goal-matching (concordance) measure was limited and was not linked to confidence or regret (Lee et al., 2014). For some decisions no patient goals predicted which treatment was received, so concordance could not be scored; whether that reflects a measurement problem or patients having little input is unresolved (Massachusetts General Hospital Health Decision Sciences Center, n.d.).

A second line of work separates the quality of a decision from its outcome. A pre-registered replication (N = 692, between participants) reproduced outcome bias: identical medical decisions were rated better when the outcome was a success, and the bias persisted among participants who said outcomes should not matter (d = 0.64) (Aiyer et al., 2023). The standard it implies is a process standard: an unbiased judge uses only the information available to the decision-maker at decision time (Aiyer et al., 2023). Kaltoft and colleagues likewise propose assessing quality at the time of the decision, not by a later outcome, and argue that satisfaction is not an appropriate measure of it (Kaltoft et al., 2014). Judging forecasts, where the truth eventually arrives, is a different problem; see the forecasting section below.

Decision aids and decision coaching

Patient decision aids are the best-tested form of decision support. The 2024 Cochrane review summarises them as producing large increases in knowledge, accurate risk perception and an active role in decisions (Stacey et al., 2024), and as probably increasing congruence between informed values and care choices, with moderate certainty (Stacey et al., 2024). Reviews of decision coaching, in which a trained person helps a patient prepare for a decision, find better knowledge and no harm, but mixed effects on other outcomes (Ottawa Hospital Research Institute, 2024).

Values clarification

A meta-analysis of values clarification methods found that explicit methods reduced values-incongruent choices (risk difference -0.04) and decisional conflict (standardised mean difference -0.20) compared with control or implicit methods (Witteman et al., 2021). Multicriteria decision analysis, which asks the person to weight and score options against criteria, produced more values-congruent decisions than other methods, though the methods did not differ on decisional conflict (Witteman et al., 2021). The authors conclude that aids should include an explicit method, but they caution about high heterogeneity and broad method categories (Witteman et al., 2021).

Debiasing

A systematic review of health-related debiasing studies found most interventions at least partially successful, with consider-the-opposite and visual aids tested most; it also names weak transfer to real-life settings and weak conceptual rigour as the main problems (Ludolph & Schulz, 2018). Checklist-style tools have mixed results. In a randomized vignette study, diagnostic checklists reduced incorrect comorbid diagnoses and raised clinicians’ confidence, but made them less likely to diagnose depression even when criteria were met (Cwik et al., 2016). A randomized trial (n = 76) of a cognitive forcing tool found no difference in error rates (O'Sullivan & Schofield, 2019).

The premortem, which asks people to imagine a project has already failed and explain why, is often traced to experiments on prospective hindsight. Those experiments found that temporal perspective (future versus past framing) had little effect, while outcome uncertainty strongly shaped explanations (Mitchell et al., 1989).

Forecasting practice

In a randomized forecasting tournament, probability training, teaming and tracking improved both calibration and resolution of forecasts (Mellers et al., 2014). In the Good Judgment Project comparisons, probability training beat scenario training, which beat no training, and teams beat crowd-belief forecasters, who beat independent forecasters (Association for Psychological Science, 2014).

Taking advice

A meta-analysis of judge-advisor studies, in which a person gives an estimate, receives advice and may revise, found that people move their estimates only about 39% toward the advice, below the 50% of equal weighting, and the authors say publication bias may make the real discounting stronger (Bailey et al., 2022). Information suggesting advisor quality was the only unique predictor of how much advice was used, with weights of about 32% for low-, 37% for neutral- and 48% for high-quality advisors (Bailey et al., 2022). Advice is taken more for subjective and uncertain estimates than for objective ones when advisor information is neutral; the evidence base is mostly anonymous computer-mediated lab tasks with few studies of optional advice (Bailey et al., 2022).

Timing matters. In experiments by Rader, Soll and Larrick, people who formed their own judgment before receiving advice were more accurate than those who saw advice first, because advice-first people anchored on the advice and then pushed away from it; both orders beat no advice (Association for Psychological Science, 2015).

Questions, controversies, and new developments

Does AI assistance improve decisions?

A meta-analysis of 106 experiments found that, on average, human-AI combinations did worse than the best of the human or the AI alone (Vaccaro et al., 2024). Decision tasks showed a significantly negative synergy effect (g = -0.27; overall g = -0.23), while creation tasks showed a positive one (Vaccaro et al., 2024). Combinations gained when the human alone beat the AI alone and lost relative to the AI when the AI was the stronger party (Vaccaro et al., 2024). Two single trials agree. Physicians given GPT-4 did not diagnose significantly better than physicians with conventional resources (Goh et al., 2024), and in a randomized bail trial, algorithmic risk recommendations did not improve judges’ classification accuracy (Ben-Michael et al., 2024). A 1,500-person experiment found that chatbot advice moved people away from their initial leanings on average, even though the chatbot favoured considerations supporting those leanings(Conlon & Schwardmann, 2026). That bias toward agreement is the subject of Sycophancy in language models.

Nearly all of this evidence comes from classification and diagnosis tasks with a known right answer, many with older systems. Whether the findings carry over to personal or work decisions, where no answer key exists, is not established.

Designs that reduce overreliance

Interface design changes how much people lean on AI advice. Cognitive forcing designs, which make the user commit or deliberate before accepting advice, reduced overreliance compared with simple explainable-AI designs, but they were rated least favourably: people trusted and preferred less demanding systems (Bucinca et al., 2021). Forcing functions helped people high in Need for Cognition more, so motivation moderates their effect (Bucinca et al., 2021). Showing a confidence score improved trust calibration but not the accuracy of AI-assisted decisions (Zhang et al., 2020); local explanations did not improve trust calibration there, and accuracy trended down without significance (Zhang et al., 2020). A preprint reports that pairing uncertainty estimates with explanations gave better performance and model understanding than either alone (Papantonis & Belle, 2023). Always showing a second opinion cut overreliance on AI but raised underreliance, so the net benefit is not assured (Lu et al., 2024); letting decision-makers choose when to ask could reduce overreliance in some cases (Lu et al., 2024).

Order of work is another lever. Among 19 veterinary radiologists, recording a provisional decision before seeing the AI output did not lengthen task time, and the one-step workflow, with AI shown together with the case, showed more anchoring on AI flags regardless of AI accuracy (Fogliato et al., 2022). Its effect on final accuracy is unsettled, and the evidence is one small expert study in one domain. This parallels the advice-timing result above. The wider study of such human-automation problems belongs to Cognitive systems engineering.

Unconventional designs

Several designs that cast the AI as something other than a recommender have small tests. In a user study, an AI in a devil’s advocate role consistently reduced over-reliance on AI, while a recommender role reduced under-reliance but increased over-reliance; the authors state the sample was small (Ma et al., 2024a). A lab study (N = 80) found that a question-asking Thinking Assistant guided users to decisions better than agents that only asked, only advised, or did neither, and its authors recommend questions first, with advice after the context is understood (Park et al., 2023). Users of the Future You older-self chatbot (n = 344) reported less anxiety and stronger future self-continuity after about half an hour; decision quality was not the reported outcome, no effect sizes or durability were reported, and the authors warn that realism could backfire (Zewe, 2024). AI debate shown to human judges raised accuracy by 4-10% over a single AI’s argument on contested factual claims, in a preprint on known-answer claims rather than personal choices (Rahman et al., 2025). Questioning as a method of moral reasoning is treated in Moral education through questioning.

Evaluative AI

Evaluative AI shows evidence for and against each hypothesis instead of recommending one. Its proposers report promising improvements in human decisions in their own user studies, on housing prices and skin cancer (Le et al., 2025). An independent behavioural experiment (n = 375) found no significant improvement in decision performance and only superficial engagement with the evidence (Kornowicz, 2024). The question is contested, and the split tracks who ran the study, with proposers finding gains and an independent test finding none.

Does relying on AI erode decision skill?

Two 2026 preprints suggest it may. In a logic-puzzle experiment, people who requested AI help during an access phase did worse once help was removed, and more independent effort during access went with larger gains in ability (Wu et al., 2026b). Randomized trials (N = 1,222) found that AI assistance made people persist less and do worse once the AI was removed (Liu et al., 2026a). Neither used personal or work decisions.

Does more deliberation help or hurt?

A meta-analysis and a large replication found no reliable support that diverting thought (deliberation without attention, or unconscious thought) beats conscious deliberation; earlier positive reports came from underpowered studies (Nieuwenstein et al., 2015). A correlational study found that regret partly mediates the link between decision difficulty and life satisfaction, and that searching more alternatives related positively to life satisfaction, which complicates the view that maximizing always harms (Moyano-Diaz et al., 2014). For fast and slow processing generally, see Dual-process theory.

Broader connections

Decision support sits between the descriptive study of how people err, in Judgment and decision-making, and the design of tools that people work with, in Cognitive systems engineering. Several findings recur across the topic. Techniques that make the person do some work first (committing to a judgment, a forcing function, a question) tend to reduce anchoring on advice but are disliked. Where the AI is the stronger party, any technique that raises the person’s weight on their own judgment risks lowering accuracy. Disagreements between sources track who ran the study, whether the task has a known answer, and the measure used.

Further reading

• Vaccaro et al. (2024): the meta-analysis of 106 experiments on when human-AI combinations beat either alone.
• Stacey et al. (2024): the Cochrane plain-language summary of the evidence on patient decision aids.
• Witteman et al. (2021): the meta-analysis of values clarification methods.
• Bucinca et al. (2021): the cognitive forcing experiments, including the finding that the effective designs were the least liked.
• Bailey et al. (2022): the meta-analysis of how much weight people give advice.