The source record
Evidence at a glance
Dates identify the source or documented event. A published claim is not automatically a demonstrated outcome.
| Record and date | What it establishes | What it does not establish |
|---|---|---|
| 2021 Team performance experiment | Human–AI combinations improved performance; the tested explanations added no further improvement. [4] | Three tasks, not a universal result. |
| 2021 Review design experiment | Buçinca and colleagues found lower overreliance with cognitive-forcing interventions. [7] | Benefits came with usability and participant-difference tradeoffs. |
| 30 June 2021 Accountability framework | GAO organized assessment around governance, data, performance and monitoring. [8] | A framework is neither a certification nor an evaluation of the reader’s system. |
| 8 February 2022 Public-sector experiments | Alon-Barkat and Busuioc did not find general automation bias relative to equivalent human advice. [5] | Dutch experimental settings; results differ across samples and measures. |
| January 2023 Voluntary risk framework | NIST AI RMF 1.0 specifies responsibilities and training for risk management. [1] | Adoption does not establish successful implementation. |
| 7 January and 9 February 2024 Timing and adjustment experiments | Agudo and colleagues studied advice timing; Sele and Chugunova studied adjustable recommendations. Both identified conditions in which human decisions deteriorated. [6] [3] | Different experimental designs cannot supply a single real-world failure rate. |
| 12 July 2024 publication Original EU AI Act | Articles 14 and 26 describe oversight capabilities and personnel requirements. [2] | Historical text; subsequent amendments, applicability and transition rules require separate legal analysis. |
Research and interpretation
Begin with the decision that people are making
A school reviewing an AI-generated lesson, a public agency assessing an application and a board approving an automated process face different decisions. Our proposed starting point is a written description of the decision, who may be affected, which part the AI performs and who can change the result. These are illustrative contexts, not findings about any named institution.
NIST AI RMF 1.0 is a voluntary framework. Its GOVERN provisions distinguish operational responsibilities, staff training and executive accountability; Appendix C discusses the varied results of human–AI interaction. Assigning a reviewer therefore raises a further research question: what resources and evidence make that role effective here? [1]
Research and interpretation
Acceptance and accuracy are different outcomes
Sele and Chugunova’s 2024 online prediction experiment involved 292 participants. Allowing adjustment made algorithmic delegation more attractive, but the human-in-the-loop design reduced decision accuracy in that experiment. Participants were less likely to intervene on the least accurate recommendations. This supports scrutiny of review design; it does not show that all human oversight is harmful. [3]
Bansal and colleagues tested three tasks with AI accuracy comparable to human performance. Combined performance improved, but the tested explanations added no further improvement and increased acceptance of incorrect advice. Confidence in an explanation should not substitute for measured outcomes. [4]
Research and interpretation
The evidence does not support a single bias story
Alon-Barkat and Busuioc’s three Dutch experiments did not find a general tendency to follow algorithmic advice more than equivalent human-expert advice. One citizen study found greater adherence when advice matched group stereotypes; the civil-servant study did not reproduce that selective-adherence pattern. This counterevidence matters: automatic deference is a risk to test, not a universal description of public employees. [5]
Agudo and colleagues manipulated when participants saw erroneous automated support in simulated judgments about defendants. Incorrect advice was particularly damaging when presented before participants formed their own judgments. These were experimental tasks, not an audit of courts or a demonstration that one interface sequence is always safest. [6]
Research and interpretation
An intervention can have costs as well as benefits
Buçinca and colleagues tested interventions designed to make participants think before relying on AI advice. In their 199-person experiment, these interventions reduced overreliance relative to simpler explainable-AI approaches. The strongest designs received less favorable subjective ratings, and benefits differed with participants’ motivation for effortful thinking. A promising intervention still needs evaluation in its intended setting. [7]
Our practical inference is to test review under realistic workload, information access and time constraints before treating it as a safeguard. Record when reviewers disagree correctly, when they introduce errors and when the process delays a needed decision. These are proposed evaluation questions, not measured outcomes from a deployment reviewed here.
Research and interpretation
Separate written requirements from demonstrated performance
Articles 14 and 26 of the original 2024 EU AI Act connect high-risk AI oversight with proportionality, interpretation, intervention and supported personnel. This is a historical governance reference. We do not determine current applicability, subsequent amendments, which organizations are covered or when duties apply. [2]
GAO’s 2021 framework includes assessment questions and audit procedures across four principles. A useful next step is to ask for the records behind an assurance of oversight: the assigned responsibility, evaluation plan, monitoring results and corrective action. A policy can establish an organizational commitment; performance records are needed to examine whether it is fulfilled. The latter is our analytical distinction. [8]
Keep the distinctions visible
Claims, findings and unknowns
A human signs off, so the AI is adequately supervised.
What the evidence supports: The signature describes a process step. The experimental evidence does not establish that such a step reliably improves accuracy. [3] [6]
What remains open: Whether reviewers detect the errors that matter in this deployment.
An explanation makes the recommendation safer.
What the evidence supports: The tested explanations increased acceptance of incorrect advice. [4]
What remains open: Whether the proposed explanation improves decisions here.
People inevitably defer to algorithms.
What the evidence supports: The Dutch studies provide counterevidence to that generalization. [5]
What remains open: How this workforce responds over repeated use, under its actual incentives.
Recommended oversight questions
Questions for decision-makers
These questions are our recommendations. They identify the records needed to test a claim or evaluate a decision.
01Who can pause, override or reject the AI-supported decision?
Why ask: Our recommendation is to demonstrate practical authority rather than merely identify a job title. [1]
Who can answer: Operational owner and accountable executive.
Records to request: Role descriptions, access permissions, escalation procedure and a demonstrated intervention.
Link to this question →02Does the combined process outperform each baseline?
Why ask: Our recommendation is to test added value. [4]
Who can answer: Evaluation lead and independent assessor.
Records to request: Protocol, representative cases, human-only and AI-only results, error severity and uncertainty.
Link to this question →03Can reviewers form an independent judgment before seeing a recommendation?
Why ask: Timing changed susceptibility to incorrect advice in an experiment; its relevance here should be tested. [6]
Who can answer: Service designer and operational supervisor.
Records to request: Interface sequence, alternative-workflow test, reviewer reasoning and case-level outcomes.
Link to this question →04Which errors reach people despite review?
Why ask: Our proposed monitoring approach examines missed errors and consequences, not only how often staff press an override button. [8]
Who can answer: Quality lead, complaints owner and internal audit.
Records to request: Incident log, error sampling plan, appeal outcomes, corrective-action record and privacy protections.
Link to this question →05Does review remain effective when demand rises or staff change?
Why ask: Interventions have tradeoffs. Our recommendation is to evaluate workload and accessibility, then specify conditions that require redesign or suspension. [7]
Who can answer: Service manager and staff responsible for review.
Records to request: Staffing assumptions, training assessments, timed exercises, accessibility testing and review capacity plan.
Link to this question →Source register
Sources reviewed 14 November 2026. Dates below identify the cited publication or version; undated documentation is labelled. Links open the original record.
- 01
National Institute of Standards and Technology · January 2023
Artificial Intelligence Risk Management Framework AI RMF 1.0 ↗
Primary voluntary framework. GOVERN 2 and 3.2; Appendix C. Source version is 1.0, not an assertion that it is the newest guidance.
- 02
Official Journal of the European Union via EUR-Lex · 12 July 2024 publication
Regulation EU 2024 1689 original act Articles 14 and 26 ↗
Original enacted text. Article 14 on printed pages 60–61 and Article 26 beginning on printed page 67. No current-applicability assessment is made.
- 03
Daniela Sele and Marina Chugunova in PLOS ONE · 9 February 2024
Putting a human in the loop Increasing uptake but decreasing accuracy of automated decision making ↗
Original experiment, 292 participants. Findings concern the tested prediction task.
- 04
Gagan Bansal and colleagues at CHI · May 2021
Does the Whole Exceed its Parts The Effect of AI Explanations on Complementary Team Performance ↗
Original research on three tasks; corporate-hosted publication page.
- 05
Saar Alon-Barkat and Madalina Busuioc in JPART · 8 February 2022 online publication; January 2023 issue
Human AI Interactions in Public Sector Decision Making Automation Bias and Selective Adherence to Algorithmic Advice ↗
Original study. Three Dutch experiments, including a civil-servant sample. Includes findings that qualify a simple automation-bias narrative.
- 06
Ujué Agudo and colleagues in Cognitive Research Principles and Implications · 7 January 2024
The impact of AI errors in a human in the loop process ↗
Original two-experiment study using simulated judgments; the second experiment was preregistered.
- 07
Zana Buçinca Maja Barbara Malaya and Krzysztof Z Gajos · 19 February 2021 preprint; 2021 CSCW publication
To Trust or to Think Cognitive Forcing Functions Can Reduce Overreliance on AI in AI assisted Decision making ↗
Authors’ research manuscript; related publication DOI 10.1145/3449287. Experiment with 199 participants.
- 08
U.S. Government Accountability Office · 30 June 2021
Artificial Intelligence An Accountability Framework for Federal Agencies and Other Entities ↗
Primary framework GAO-21-519SP. Supports evaluation and audit questions, not a claim about a particular institution’s performance.
Method and corrections history
Sources were selected for original empirical evidence or authoritative governance text, including findings that challenge a simple pro- or anti-automation account. This is a focused review, not a systematic review or meta-analysis. Study outcomes remain attached to their tasks and samples. Proposed questions and cross-source inferences are labelled as our recommendations or analysis. We did not audit an operational system, measure staff performance, or determine legal compliance. Corrections should identify the affected claim, source and revision date.
- 14 November 2026 · Edition 1
- First publication of this brief, its connections companion and matching PDF and Word editions. No subsequent corrections are recorded in this edition.
New evidence and substantive corrections will be dated here. Read the corrections policy →