Room to Disagree

The framework text is fixed at v1.0 and will not be revised structurally before February 2028. Findings that would change a future version are recorded here as they arrive, dated and credited, so that the framework is visibly answerable to work published after it. Most entries below enlarge v1.0; the note of August 31, 2026 on what a detection rate can establish corrects its wording.

September 27, 2026 — The base condition’s statutory footing: Article 16(l)

The base condition — that the oversight interface must itself be accessible — was argued in v1.0 on practical grounds: a person assigned to oversight may have a disability, and if the interface fails them, the oversight obligation fails with it. It also has a statutory footing that v1.0 does not cite. Article 16(l) of the AI Act requires providers of high-risk AI systems to “ensure that the high-risk AI system complies with accessibility requirements in accordance with Directives (EU) 2016/2102 and (EU) 2019/882”.

Article 50(5) makes the corresponding demand of the disclosures in Article 50: the information “shall conform to the applicable accessibility requirements”. Accessibility therefore reaches both sides of the boundary this framework is concerned with — what the system discloses to the people it affects, and the interface through which the designated person oversees it.

Two limits belong alongside it. The first is scope. The directives Article 16(l) points to were written for the websites and mobile applications of public sector bodies, and for a defined list of products and services. An internal review console — a credit-review queue, a claims-handling screen — is often neither, and what Article 16(l) requires of such an interface has not, to my knowledge, been settled by guidance or case law. This note does not settle it. The second is that the design conclusion does not wait for the legal one: a designated person who cannot operate the interface cannot exercise any of the five capabilities, whatever the scope of Article 16(l) proves to be. The base condition was stated as a design criterion and remains one; the statute adds weight to it, not a new requirement.

One consequence for the seeded-error audit. Where a deployer’s reviewers work through more than one input method — mouse, keyboard only, screen reader — detection should be reported for each. A gap between them is evidence that the base condition fails in operation, whatever conformance testing reported. The audit’s ground rules apply with more force here, not less: below thirty seeds, counts rather than percentages, and the result is a finding about the interface, never about the people using it.

Regulation (EU) 2024/1689 (Artificial Intelligence Act), Articles 16(l) and 50(5) — EUR-Lex, consolidated version of 27 July 2026.

Directive (EU) 2016/2102 of the European Parliament and of the Council of 26 October 2016 on the accessibility of the websites and mobile applications of public sector bodies.

Directive (EU) 2019/882 of the European Parliament and of the Council of 17 April 2019 on the accessibility requirements for products and services.

September 2, 2026 — What friction symmetry claims, and what it does not

Buçinca, Malaya and Gajos found that cognitive forcing functions reduce overreliance by requiring an explicit analytical act before a recommendation can be accepted. That is compatible with friction symmetry rather than opposed to it: raising the cost of accepting moves an asymmetric interface toward parity. The criterion is that rejecting costs no more than accepting, not that rejecting is frictionless. Requiring a justification for a high-consequence reversal is also not purely a cost, since the record carries accountability and learning value, and treating it as friction alone is one-sided.

Two limits belong alongside it. In the authors’ words, people assigned the least favourable subjective ratings to the designs that reduced overreliance the most, and the benefit concentrated among participants high in need for cognition — meaning the intervention helps least the users most at risk. Separately, Bo and colleagues evaluated three interventions intended to improve reliance on large language models and found that while they reduce over-reliance, they generally fail to improve appropriate reliance.

Friction symmetry therefore claims less than it may appear to. It removes a bias toward acceptance that the interface itself introduced. It does not claim to make reviewers better at judging, and the evidence does not support that claim. Detection rate alone is correspondingly incomplete: a reviewer who rejects often and wrongly scores as well as one who rejects correctly. Appropriate reliance — the established construct in this literature — is the measure that belongs alongside it, and its absence from v1.0 makes detection rate gameable on its own.

Buçinca, Z., Malaya, M. B., & Gajos, K. Z. (2021). To Trust or to Think: Cognitive Forcing Functions Can Reduce Overreliance on AI in AI-Assisted Decision-Making. Proc. ACM Hum.-Comput. Interact. 5(CSCW1). DOI 10.1145/3449287.

Bo, J. Y., Wan, S., & Anderson, A. (2024). To Rely or Not to Rely? Evaluating Interventions for Appropriate Reliance on Large Language Models. arXiv:2412.15584.

September 2, 2026 — Seams, and where this framework’s Interpret criterion comes from

Ehsan, Liao, Passi, Riedl and Daumé argue that seamless design is an unquestioned virtue that conceals a system’s mismatches, and that hiding those seams disempowers users to mitigate the consequences of AI mistakes. Their proposal is to reveal seams strategically as a resource for explainability and user agency, rather than treating them as defects to be smoothed away.

The argument this framework makes about interaction cost is the same move at a different level. Concealing a system’s mismatches disempowers the person affected by a decision; concealing the cost of disagreement disempowers the person reviewing it. In both cases seamlessness is credited as a design virtue while doing the damage.

The overlap is sharpest on the Interpret criteria. A reviewer who can only accept or reject a verdict has nothing to take hold of. What makes specific disagreement possible is somewhere to locate the failure — this input was misread, this threshold is wrong, this case sits outside the validation set. Those are seams in their sense, and their paper accounts for why they matter more fully than the Interpret section of this framework does.

The paper was published in April 2024 and is not cited in v1.0. It was brought to my attention by one of its authors after publication, and it will be cited in the related work at the next revision.

Ehsan, U., Liao, Q. V., Passi, S., Riedl, M. O., & Daumé III, H. (2024). Seamful XAI: Operationalizing Seamful Design in Explainable AI. Proc. ACM Hum.-Comput. Interact. 8(CSCW1), Article 119. DOI 10.1145/3637396.

August 31, 2026 — What a detection rate can and cannot establish

The framework describes the seeded-error audit as producing evidence about whether an interface enables oversight. That should be read asymmetrically. A seeded-error audit cannot establish that oversight is effective. It can establish that it is not: where reviewers fail to catch errors they were equipped to catch, under the conditions they actually work in, oversight is not happening on that surface. A high detection rate is weaker evidence than a low one, because seeded errors are drawn from failure modes already known, and the failures that cause harm are frequently the ones nobody anticipated.

Functional safety has a name for this structure. Under IEC 61508 a proof test is a periodic test performed to detect dangerous hidden failures in a safety-related system, and proof test coverage is the established measure of what fraction of dangerous undetected failures a given test can reveal. Coverage below one hundred percent means a passing test does not demonstrate the absence of faults. The seeded-error audit has the same shape and the same limit, and saying so plainly is more useful than claiming more. Wording on this page and on the method page has been corrected accordingly.

August 31, 2026 — Planted-error detection is an established research paradigm

Liu, Zhou, Shen, Liu, Wu and Chen collected interaction logs from seventy-seven participants working with a language model into which plausible misinformation had been injected, across quiz solving, article summarization and trip planning, and assessed overreliance by whether participants detected and corrected those errors. They identify five behavioral patterns that separate low- from high-overreliance users.

Their method section records that injecting known errors and measuring whether people catch them is an established approach in this literature, citing four prior studies as precedent. The uniqueness dossier accompanying this framework traced the seeding technique to error seeding in software validation and to Threat Image Projection in aviation security. Those precedents are real, but the nearer one is the standard experimental paradigm of human-computer interaction research on overreliance, and the dossier will be corrected at its next version.

What remains particular to the seeded-error audit is narrower than that dossier states, and is limited to four things: it runs in service on a deployed system rather than on recruited participants in a study; it runs unannounced (the practice disclosed to the workforce, the timing not), under a consent and governance regime that makes that lawful; it exists to produce conformity evidence rather than research findings; and it reports at the level of the interface against a threshold committed in advance.

Two things in their paper support the argument made here. Their opening observation is that overreliance cannot easily be captured by post-hoc task outcomes, which is the distinction this framework draws between records and evidence. And their closing caution — that future work must separate genuinely overreliant behavior from rational strategies and from interface- or user-driven effects — applies directly to detection rate. A missed seeded error is not automatically a failure of oversight. It may be a rational allocation of attention under load, and the framework treats the measure as cleaner than it is.

Their behavioral indicators are continuous and passive where this audit is periodic and active. Instrumentation of the kind they describe could reasonably tell a deployer when an audit is worth running.

Liu, C., Zhou, Q., Shen, X., Liu, X. B., Wu, T., & Chen, X. A. (2026). Behavioral Indicators of Overreliance During Interaction with Conversational Language Models. CHI 2026. DOI 10.1145/3772318.3790332.

August 27, 2026 — The field’s consensus framework, and where this one attaches to it

Keeping an Eye on AI (Gaube et al., 2026, arXiv 2605.16278) is the output of Dagstuhl Seminar 25272 and carries twenty authors. It sets out an oversight architecture of two layers — the task layer being overseen and the oversight layer doing the overseeing — a process model of the information flow between them, preconditions for monitoring and intervention, and a documentation template applied to four worked domains. It was published in April 2026 and is not cited in v1.0 of this framework. That is an omission rather than a judgment, and it is recorded here rather than quietly corrected.

The two are complementary in a specific way. Gaube et al. conceptualize monitoring as a signal detection task: overseers attend to cues, accumulate evidence, and intervene when the evidence warrants it. Signal detection requires knowing which items carried a signal, since a hit rate cannot be computed without it, and their framework does not supply a way to obtain it in a live deployment. The seeded-error audit is one: planted errors are known signal, and detection rate is the hit rate. Their architecture is the model; this contributes an instrument for it, grounded in the same signal-detection account already cited here through Langer, Baum and Schlicker.

Two things in this framework remain absent from theirs, verified by full-text search: the accessibility of the oversight surface as a condition of the obligation, and cost parity between accepting and rejecting an output. Neither is a criticism of a paper working at a different level.

Gaube, S., Langer, M., Miller, T., et al. (2026). Keeping an Eye on AI: A Framework for Effective Human Oversight of AI Systems. arXiv:2605.16278. CC BY-SA 4.0.

August 27, 2026 — Over-reliance is one of two failure directions

Faas, Kerstan, Uth, Langer and Feit ran four co-design workshops with domain experts overseeing an AI grading system. Their most prevalent finding across all groups was that participants were drawn to performing the AI’s task themselves rather than overseeing it — re-checking every item, finding it inefficient, and reporting that they felt useless and bored.

That is under-reliance and role confusion, and it is the opposite of the mechanism Resist is built around. Their conditions removed the pressures that produce automation bias: no throughput target, no queue depth, and participants who had not used AI support for the task before. The two findings are compatible. But Resist follows Article 14(4)(b), which names automation bias, and a reader could take the framework to mean that over-reliance is the only way oversight fails. It is one of two, and the interface can produce either.

Faas, C., Kerstan, S., Uth, R., Langer, M., & Feit, A. M. (2025). Design Considerations for Human Oversight of AI: Insights from Co-Design Workshops and Work Design Theory. arXiv:2510.19512.

August 27, 2026 — Removing the per-item decision, as an alternative to cost parity

In the same study, two groups designing their ideal oversight interface removed the per-item accept and reject decision entirely, leaving only a control to finish the task. One group separated detecting errors from fixing them.

Friction symmetry equalizes the cost of the two paths. This is a third option: remove the per-item confirmation, so that there is nothing to rubber-stamp and the reviewer’s job becomes finding problems rather than confirming each item. Where the reaction timeframe permits it, it is a stronger intervention than making rejection cheaper, and the framework does not currently name it.

Credit: Faas et al. 2025, workshop groups CS2 and Psy2.

August 27, 2026 — The cost of disagreeing is not only interaction cost

Friction symmetry counts actions. Two sources point at a cost that actions do not capture.

Faas et al. report participants feeling bad about themselves when reducing a student’s marks — one observed that had they not intervened, the student would have kept the points, and “I screwed them up.” Separately, an analysis of what makes disagreement possible names blame asymmetry: where deferring to the system and being wrong is attributed to the system, while overriding and being wrong is attributed to the reviewer, no rational reviewer overrides.

These are the same phenomenon from opposite directions, and both operate even where the workflow is perfectly symmetric. A future version should carry a criterion under Resist covering the psychological and liability cost of disagreeing, distinct from its interaction cost.

This also touches the one condition for effective oversight the framework does not turn into criteria. Sterz et al. name four: causal power, epistemic access, self-control and fitting intentions. The framework covers the first three and is silent on whether the reviewer is motivated to do the job properly, which Faas et al. address directly.

August 27, 2026 — Automation bias ranked fourth in a corpus of documented failures

Hardy’s Oversight-to-Control analysis coded nineteen public incident and deployment cases. Automation bias appeared in four of them and was not the leading failure mechanism; missing escalation or recourse, inadequate validation, and organizational governance failure appeared more often.

The corpus is drawn from public records of institutional failure, which will under-represent interface defects that never generate a case file, and the author states the counts are not prevalence estimates. It is recorded here as a reasonable corrective to any reading of this framework that treats the interface as the only place oversight goes wrong. Of the several places it can go wrong, the interface is the one a design team can change.

Hardy, K. (2026). Research Square preprint, DOI 10.21203/rs.3.rs-10151972/v1.