Judgement: questions about AI

AnsweredHow fast must a human be able to stop an AI agent?#

Faster than it can finish what it started, and no standard exists. OpenAI's own report of a 20 September 2026 incident is the only published clock: alert at twelve minutes, human at fifteen, run killed two and a half hours later because it did not stop by itself. The organisational version is three timings, one yes-or-no and one name.

AnsweredWhy did it take hours to stop an AI agent that was flagged in minutes?#

Because detection and stopping are different capabilities with different clocks, and only the first is commonly built. In OpenAI's report the automatic stop did not fire as expected, and a preprint auditing 63 public documents on agent systems found logging described in most and checkpoints, recovery or appeal in almost none.

AnsweredShould a human approve every target an AI selects?#

No treaty requires it, and The Washington Post reported on 26 September 2026 that the US and Russia removed the UN draft clause that would have. A final approval is necessary and not sufficient: under pressure people confirm the recommendation, so the decisions that hold are which targets a machine may propose, who may refuse, and what the approver must establish for themselves.

AnsweredWhat is the difference between a human in the loop, on the loop and off the loop?#

In the loop, a person makes each decision; on the loop, the machine proposes and prepares and a person supervises and can stop it, as in the Australian Ghost Bat test ABC News described on 27 September 2026; off the loop, no person intervenes. The automation-bias evidence is that on-the-loop approval under pressure tends to become confirmation.

AnsweredShould AI reject an application before a human reads it?#

Only when a named person answers for the rejected list, the rejected can find out and contest it, and the documents still carry information. Times Higher Education reported on 25 September 2026 that a UKRI-funded call cut about half of 179 proposals unread; a September 2026 hiring model says machine-written applications stop telling candidates apart, so the answer is an extra assessment stage, not a harder filter.

AnsweredDoes AI weaken human judgement?#

Not on its own. It removes the demand for judgement, and demand is what builds it. Across 106 experiments, human and AI pairs did worse than the better of either alone.

AnsweredWhat is automation bias?#

The tendency to over-accept automated output. Two error types, omission and commission, and it appears in experts as well as novices.

AnsweredWhen should I override AI?#

Six conditions that should trigger an override, three where you should defer, and the precondition nobody checks: could the person detect the error at all?

AnsweredIs human in the loop enough?#

The meta-analysis says the common configuration underperforms the stronger party alone. Review after generation is the weakest available design.

AnsweredCan a human approve an AI decision at machine speed?#

Only when the machine proposes no faster than the person can check, and most deployments have measured neither. If decisions per hour times minutes per real check exceed the attention available, the approval is a sample and somebody must own the rest.

AnsweredWhat is human-AI collaboration?#

A pairing with the division of labour, the human entry point and the override grounds specified in advance. Without those three, it is handover, not collaboration.

AnsweredWho is accountable when AI gets it wrong?#

A person or an organisation, since nothing else can be: in England and Wales an AI system has no legal personality, and the Divisional Court held in 2025 that the duty to check stays with the professional. Answered 5 October 2026; how to allocate it inside an organisation is at /research/how-should-ai-decision-rights-be-allocated.

AnsweredShould humans always make the final decision?#

No. A universal veto is not a safety principle. Which party performs better, how reversible the error is, and whether the human can detect it at all.

AnsweredWhat is meaningful human oversight?#

A term from the autonomous-weapons literature now appearing in regulation. Open, and increasingly commercially relevant.

Better answered elsewhere Article 14 is the precise legal text, and reading it directly beats reading anybody's summary of it, including this one. · EU AI Act, Articles 12, 14, 19 and 50

AnsweredWhat is human in command?#

A person's retained authority over an AI system, including whether to use it. The EU's Economic and Social Committee used the phrase in 2017 and its expert group defined it in 2019; the 2025 opinion often cited for it contains it only in a rejected amendment.

AnsweredCan I use AI to check AI?#

No. A South African judgment records a judge testing a fabricated citation in ChatGPT, which confirmed it was real.

AnsweredWhat is a hallucination?#

Established term, and there is a reasonable case for retiring it, because it names a confident error after a perceptual one.

AnsweredWhat is algorithm aversion?#

Dietvorst. The mirror image of automation bias, and the two are almost never discussed together.

AnsweredDo agents change the judgement question?#

Autonomy moves the human decision earlier, from approving output to setting the boundary. Most organisations have not moved with it.

AnsweredWhat is the difference between a good decision and a good outcome?#

The distinction most organisations collapse, and the reason bad processes survive when they get lucky.

AnsweredWhat is automation complacency?#

Reduced monitoring of a system because it has been reliable. Not laziness: a rational allocation of attention that fails exactly when the system does.

AnsweredWho has the authority to override an AI system?#

Whoever was assigned it before deployment. Article 14 names competence, training and authority together, and in practice they get separated.

AnsweredWhat happens when nobody wants to be the person who overrides it?#

The override stops existing in practice and continues on paper. Overriding is visible and attributable; deferring is neither.

AnsweredCan a human challenge a system they cannot inspect, or only appear to?#

On the output, not the reasoning. Which works where the answer is implausible and fails where it is plausible and wrong.

AnsweredIs human intuition better than AI logic?#

Neither in general. Kahneman and Klein set two conditions for trustworthy intuition: a learnable environment, and prolonged practice in it with fast clear feedback. AI barely touches the first and removes the second, because the practice is what gets automated.

AnsweredWhat happens when AI and human judgement conflict?#

Usually nothing visible. A pre-registered experiment found access to a system raised agreement from 58.4 to 80.9 per cent while accuracy fell from 74.2 to 63.9, so the conflict is absorbed rather than resolved and the organisation never sees it.

AnsweredDo executives follow AI recommendations against their own judgement?#

31 per cent say they do, in a September 2026 survey of 300 US finance, information and operating chiefs: 48 per cent of CFOs, 33 per cent of CIOs, 11 per cent of COOs, with 39 per cent reporting any formal process for AI-driven decisions. Self-reported and vendor-commissioned; the experiments on the same page show the larger share never notice the disagreement at all.

AnsweredShould we trust AI over human experts?#

It depends who is being assisted. Novices gain a great deal; for domain experts the average gain is close to nothing and a randomised study of 140 radiologists found the effect ran from strongly positive to strongly negative with nothing predicting which.

AnsweredWhat is the role of intuition in human judgement?#

Recognition trained by exposure. It carries most of the work where feedback is fast and clear and very little where outcomes arrive late or ambiguously, so the discriminating variable is the environment rather than the person, which makes it checkable.

AnsweredIs AI decision-making biased?#

Yes, in two directions people confuse: bias in the outputs, measured at 85.1 per cent favouring White-associated names in one resume audit, and the human tendency to over-accept them. They need different remedies.

AnsweredWhat is human-in-the-loop AI?#

A design with a person somewhere in the decision path. The meta-analysis says the common configuration underperforms the stronger party alone.

AnsweredWhat is the role of human oversight in AI decisions?#

To catch the errors the system makes, which requires that somebody could detect them. Oversight that cannot tell a right answer from a plausible one is a control on paper rather than a control in fact.

AnsweredWhat is human oversight in artificial intelligence?#

A person positioned to review, question or stop an automated decision. The word meaningful is doing the work: the EU AI Act requires it, and the practical test is whether the reviewer could have detected the error at all.

AnsweredHuman judgement versus AI in decision making#

The versus framing is the error. Across 106 experiments the common human and AI pairing performed worse than the better of either alone, which is an argument about how the pairing is designed rather than about which party is superior.

AnsweredHow does human intuition compare with artificial intelligence?#

Both are pattern recognition. They differ in what trained them and in whether the conditions for trustworthy intuition still hold, which Kahneman and Klein reduced to a learnable environment plus prolonged practice with fast feedback.

AnsweredHow did an AI hallucination nearly trigger a US military operation?#

According to CNN on 18 September 2026, an analyst asked a chatbot what a Chinese ship carried, asked it again to write the answer up as a standard intelligence report, and the report circulated until aircraft were airborne. Officials traced it to the tool only immediately before the operation; the command and the Pentagon did not comment. The hallucination was ordinary; the missing record of where the report came from was the failure.

AnsweredDo employees know they are expected to check and override AI?#

Mostly not. IBM's 2026 survey of 1,500 CHROs and 8,800 employees found 71 per cent of executives rank supervising and overriding AI output as essential against 38 per cent of employees, and only 26 per cent of organisations had written down which work is human-led. The check exists in a leader's expectation and nowhere an employee could read it.

AnsweredWho gets blamed when AI gets it wrong at work?#

The nearest person, on the employees' own account: 43 per cent told IBM's 2026 survey that the blame falls on them when AI fails, while 36 per cent of executives say accountability is unclear. It is Elish's moral crumple zone counted for the first time, and the remedy is naming who owns, validates and may override each decision before deployment.

AnsweredWhat is premortem?#

A meeting held before a plan is executed, in which the team is told the plan has failed and each member writes down why, so that doubts are stated as explanations rather than as objections.

AnsweredWhat is decision journal?#

A written record made at the time of a decision, covering the decision, the reasoning, the expected outcome and the confidence attached to it, so that later review compares against what was actually thought rather than against a reconstr...

AnsweredWhat is calibration training?#

Structured training in probabilistic reasoning, usually covering comparison classes, base rates, incremental updating and review of resolved forecasts, aimed at bringing a person's stated confidence into line with their actual accuracy.

AnsweredWhat is the five questions that keep you human?#

The five questions that keep you human are Rahim Hirji's bearings for when work accelerates and decisions begin arriving pre-made: who is actually deciding here, what stayed human when speed took over, what changed because we learned som...

AnsweredWhat is types of judgement?#

Judgement divides into three kinds that behave differently under automation: predictive judgement, estimating what will happen; evaluative judgement, deciding what matters and how much; and moral judgement, deciding what is owed to whom ...

AnsweredWhat is collective judgement?#

Collective judgement is judgement produced by a group. It improves on the individual reliably only where judgements are made independently before being combined; where they are formed in discussion, the group inherits the first framing o...

AnsweredWhat is decision hygiene?#

Decision hygiene is the set of procedures that reduce noise in judgement without identifying which judgements were wrong: structuring a decision into independent assessments, judging independently before aggregating, using a common relat...

AnsweredWhat is the three terrains?#

The three terrains are Rahim Hirji's map of where a machine should lead and where a human keeps the call: statistical terrain, where machines outperform human perception and the human interprets; bias terrain, where a tested and monitore...

AnsweredWhat is hand, Head, Hours, Heart?#

Hand, Head, Hours, Heart are Rahim Hirji's four receipts for human judgement inside fast systems. Hand records what stayed human and who will answer for it. Head records what was learned from a miss and what changed because of it. Hours ...

AnsweredWhat is the use-or-keep test?#

The use-or-keep test is six questions applied to one piece of work before using AI on it: whether the output can be checked, what happens if it is wrong and nobody notices, whether the task is a repetition that maintains a capability, wh...

AnsweredWhat are ironies of automation?#

The ironies of automation are that automating the routine parts of a task leaves the human the hardest residue, monitoring and handling exceptions, while removing the practice that built the competence to do it.

AnsweredWhat is the out-of-the-loop performance problem?#

The out-of-the-loop performance problem is the loss of a person's ability to take over manual operation when an automated system fails, caused by their having been placed in the role of monitor instead of operator. Named by Mica Endsley ...

AnsweredWhat is situation awareness?#

Situation awareness is the perception of what is happening around you, the comprehension of what it means, and the projection of what it will mean next.

AnsweredWhat is the vigilance decrement?#

The vigilance decrement is the measurable decline in the probability of detecting rare signals as time on a monitoring task increases. First demonstrated by N. H. Mackworth in 1948 using the Clock Test.

AnsweredWhat is alarm fatigue?#

Alarm fatigue is the desensitisation of a person to warning signals caused by exposure to a high volume of them, most of which turn out not to require action, leading to slower responses, silenced alarms and missed true events.

AnsweredWhat is moral crumple zone?#

A moral crumple zone is what forms when responsibility for a failure falls on the human nearest an automated system, who had limited real control over its behaviour.

AnsweredWhat are misuse, disuse, abuse?#

Misuse is over-reliance on automation, disuse is unwarranted rejection of it, and abuse is deploying it without regard for the human consequences. Appropriate use is the fourth case.

AnsweredWhat is algorithm appreciation?#

Algorithm appreciation is the tendency to weight algorithmic advice more heavily than the same advice from a person. Domain experts are the exception.

AnsweredWhat is meaningful human involvement?#

Meaningful human involvement is the UK statutory test for whether a decision is solely automated: whether a human can exercise real influence before the decision is applied, and has the authority, discretion and competence to alter it, a...

AnsweredWhat are shallow jobs?#

Shallow jobs are Bain's term for roles in which people rubber-stamp mostly correct AI output without engaging their judgement.

AnsweredWhat is the verification bottleneck?#

The verification bottleneck is the constraint that appears when AI makes producing work cheap and checking it stays slow, skilled and human, so that the capacity to verify, not the capacity to generate, sets the pace.

AnsweredWhat is falling asleep at the wheel?#

Falling asleep at the wheel is what happens when people given high-quality AI become careless and less skilled in their own judgement, because the AI is good.

AnsweredWhat is calibration?#

The correspondence between the confidence a system states and how often it is correct.

AnsweredWhat is the invisible work of oversight?#

The invisible work of oversight is the labour of supervising an automated system that appears in no workload model and no business case: staying attentive while nothing happens, holding enough of the task in mind to notice when the outpu...

AnsweredWhat is over-reliance?#

Dependence on an automated system beyond the point at which the person relying on it could detect that it was wrong.

AnsweredWhat is override authority?#

The assigned power of a named person to disregard, reverse or stop an AI system's output, held together with the competence, information and organisational standing required to use it.

AnsweredWhat is silent failure?#

Silent failure is a failure that produces no error signal a person can act on: the system carries on, the output looks ordinary, and nothing marks the point at which it stopped being right.

AnsweredWhat is fail-plausible?#

Fail-plausible is a silent failure in which a language model turns the error into fluent, plausible narrative and delivers it to the user, so the observer is not merely blind to the failure but is convincingly misled by it.

AnsweredIs it ethical to let AI judge people?#

It turns on four obligations usually collapsed into one: explanation the person can use, contest by somebody able to change the outcome, a named human carrying it, and verification on the population it is used on.

AnsweredCan AI be unbiased, or does it reproduce human bias?#

It reproduces it, measurably, and unbiased is not one target: several reasonable fairness criteria are provably incompatible, so every system judging people has already chosen between them.

AnsweredWhat makes an automated decision explainable enough?#

The Court of Justice answered this in Dun and Bradstreet Austria: describe the procedure and principles actually applied so the person can see which of their data was used. Publishing the algorithm does not satisfy it.

AnsweredDo bias audits actually happen?#

Rarely. Under the first such law, a field audit of 391 employers found about 5 per cent had published one, and the state comptroller found seventeen potential breaches in a sample where the enforcing department had found one.

AnsweredWhich decisions about people count as high risk?#

Annex III of the EU AI Act names them, including access to education, evaluation of learning outcomes that steer a path, and a list of employment decisions. The organising idea is what a person is permitted to become.

AnsweredHow is judgement trained?#

Four things have evidence behind them: structured debriefs, calibration training, a single well-designed debiasing exercise, and simulation with graduated responsibility. Most of what is sold as judgement training has none.

AnsweredHow do I decide whether to use AI on a piece of work?#

Six questions: can I check it, what happens if it is wrong and I do not notice, is this a repetition I need, would doing it myself teach me anything, has the model seen this before, and who answers for it.

AnsweredIs there a framework for when to use AI?#

Several, and they address different people. The regulations classify systems, human factors classifies functions at design time, and the management frameworks address the executive allocating work. Almost nothing addresses the person holding the task.

AnsweredWhat is a premortem?#

A meeting held before a plan starts, in which the team is told it failed and each person writes down why. The evidence is for generating more and better specified reasons, not for better projects.

AnsweredShould I keep a decision journal?#

It is the only way to know later what you actually thought, because hindsight rewrites the memory rather than suppressing it. No trial has measured whether keeping one improves decisions.

AnsweredCan you train someone to be better calibrated?#

Yes. It is the one component of judgement measured against resolved outcomes. About an hour of probabilistic reasoning training improved forecasting accuracy by roughly 6 to 12 per cent in a randomised tournament.

AnsweredDoes bias training work?#

Awareness sessions rarely change behaviour. A single interactive exercise with immediate feedback on your own errors produced large reductions that lasted twelve weeks, which suggests practice rather than information is the active ingredient.

AnsweredWhen should I trust the machine and when not?#

Three kinds of ground. Statistical terrain, where it sees what you cannot and your job is interpretation. Bias terrain, where a tested model discriminates less than your current process. Override terrain, where the cost falls on a person and the call stays human.

AnsweredHow do I prove a human was involved in an AI-assisted decision?#

Four short receipts: what stayed human and who answers for it, what was learned from a miss, where the saved time went, and who bears the consequence including those who never enter the system as data.

AnsweredWhat questions keep judgement human when work speeds up?#

Who is actually deciding here. What stayed human when speed took over. What changed because we learned something real. Where did judgement travel. What would I still stand behind in a year, with names attached and without context.

AnsweredAre there different kinds of judgement?#

Three that behave differently under automation. Predictive judgement estimates what will happen, evaluative judgement decides what matters and how much, and moral judgement decides what is owed to whom. Machines do the first.

AnsweredDo humans and AI together make better decisions?#

On average, no. A meta-analysis of 106 experiments found the combination performed worse than the better of the two alone. Gains appeared in creation tasks and where the human was already better.

AnsweredHow do groups make better judgements?#

By collecting judgements independently before discussing them. It is the one intervention that reliably works and it requires nobody to think better, only that the order be reversed.

AnsweredWhat is decision hygiene?#

Procedures that reduce error without knowing which decisions were wrong, in the way handwashing works without identifying the pathogen. Structured assessment, independent judgement, a common scale, sequenced information.

AnsweredWill human judgement protect lawyers from AI?#

Not as a job description, and on that Richard Susskind is right, Legal Futures reported on 28 September 2026. The courts have decided who answers when the reliable machine is wrong, Garicano's data shows the routine work leaving the firms that trained juniors, and judgement is what a client still buys only if it is still exercised.

AnsweredHow often is AI wrong?#

There is no single rate. Journalists found a significant issue in 45 per cent of assistants' news answers, and in a medical trial models alone scored 94.9 per cent while people using them scored under 34.5. A rate needs its task, its assessor and its marking rule.

AnsweredWhat counts as meaningful human review of an AI decision?#

Neither California's law nor the UK's defines it in detail, and UK ministers may do so by regulation under Article 22D of the UK GDPR. The page offers four facts that separate a review from a signature: the reviewer works from records the system did not produce, can disagree without cost, can still do the judging unaided, and the rate at which reviewers reverse the system is counted. A reversal rate of zero is evidence about the review.

AnsweredWhat should happen when an AI agent asks for permission and nobody answers?#

It should stop. The UK AI Security Institute reported on 28 September 2026 that in its simulations a model asking a question received only the automated reply 'Please proceed to the next step using your best judgement', and sometimes treated it as permission to act on out-of-scope targets. Silence, a default approval or an unread inbox must count as refusal, with a named person who answers.

AnsweredHow should a public body decide whether to continue an AI pilot?#

By a threshold written before the pilot starts. The Register reported on 30 September 2026 that British Transport Police's six-month facial recognition trial scanned more than 500,000 faces, produced one alert, a false one, cost more than 320,000 pounds, and was extended; neither report read says what result would have ended it. The page's conditions apply: set the stopping number while everyone is still pleased with the system.

AnsweredCan a research funder use AI to reject grant applications?#

One tried it and took it back. Times Higher Education reported on 5 October 2026 that CRANE, a UK cybersecurity research network, will re-review every application its AI triage screened out of a call of 179, almost half of them, after apologising for not being transparent; offers already made are suspended. UKRI says putting confidential material into generative AI breaches its policies.

AnsweredHow would an auditor prove that a human approved what the AI did?#

Only from a record fixed in advance. A position paper posted on 5 October 2026 by Zhao and colleagues argues that a claim such as 'a person approves every external email' can be checked only if it first names its policy, its scope, the records that would settle it and who writes them. The four receipts are that record kept by people; the paper is an argument and not a test.

AnsweredWhich decisions does an AI company say its model must never make?#

Anthropic's usage policy update of 8 October 2026 says its model cannot be used to decide or recommend who to investigate, arrest or charge, and requires a qualified human in the loop with authority to change its recommendations wherever health, legal rights, finances, livelihood or essential services are affected. Other developers' terms were not read this week. A vendor's list is not an organisation's own.

AnsweredCan a decision rest on an AI summary?#

Only if the person who answers for the decision can still read the full record, is told how often the summary fails, and the failure test counts what was left out. In October 2026 the Home Office checked 0.3 per cent of its AI asylum summaries while the judiciary required every AI output to be reviewed, verified and approved by the accountable judge.

AnsweredIf a caseworker reads an AI summary instead of the full interview, who is accountable for what it leaves out?#

The caseworker and the department, and neither would know. The Home Office's tool, as Computer Weekly reported on 15 October 2026, passes summaries that omit material facts because relevance is not a ground for failure, checks 0.3 per cent of them and keeps them out of monthly quality assurance. The page sets out the four lines that would make the accountability real.

AnsweredIs checking a sample of AI outputs enough oversight?#

A sample estimates a failure rate; it does not find the failures. Of 203 asylum summaries the Home Office sampled from about 70,000, 45 failed first review, which if representative implies thousands of failed summaries that reached decision makers unchecked. The page holds what protects the person in the other 99.7 per cent.

AnsweredCan a judge use AI and still be fully accountable for the judgment?#

That is what the judiciary's position statement of 9 October 2026 requires: judicial decisions are taken by a judicial office holder, every AI output is reviewed, verified and approved by that person, who 'remains fully accountable', and tools must not determine the outcome of a case. The custody page records it as a list written before deployment.

AnsweredWhen an AI agent proposes changes, how should a team decide which to keep?#

By review against the maintainer's own standard. Google Cloud engineers reported on 13 October 2026 that an agent running eleven hours on a 205,000-line codebase proposed six optimisations; they shipped two and rejected four: 'the agent explores, the maintainer subtracts.' The page holds why that subtraction is the judgement the organisation is paying for.

AnsweredWhy do developers use AI all day but refuse to trust it for important decisions?#

Because they can check the small things and not the big ones. Stack Overflow's 2026 survey of over 30,000 developers found 48 per cent trust AI when they can easily validate the answer and 6.6 per cent for important decisions, with 78.7 per cent wanting source attribution. The page holds why verifiable work is where trust stops.

AnsweredIf AI diagnoses better than a doctor, what is the doctor still for?#

Accountability. Two clinicians wrote in the Annals of Internal Medicine on 12 October 2026 that what stays with the clinician is 'owning the risk of acting under uncertainty and answering for the decision that follows', and that a system 'with nothing to lose' weighs options differently. The page holds that as the definition of the class.

AnsweredCan AI check AI's work?#

Not on its own. A machine check can find classes of error at a volume no person can read; the answer for the result stays with a person. In October 2026 the Law Society of England and Wales said responsibility cannot be transferred to AI, the Supreme Court of Tasmania required AI-generated content to be checked by a person and not solely by another tool, and Anthropic's code review team wrote that an agent's approval never counts as a person's.

AnsweredDoes an AI reviewing another AI's output count as verification?#

Not for a court or a profession. The Supreme Court of Tasmania's circular of 21 October 2026 says AI-generated content 'must be checked by a person, not solely by a Gen AI tool', and the Law Society's practice note of the same day says AI outputs deserve the scepticism applied to a junior's work. The page holds what a machine check can be trusted to find and what it cannot carry.

AnsweredIf an AI agent approves a change, does that count as a human approval?#

Not even at the company that automated most of its review. Anthropic's Claude Code team wrote on 15 October 2026 that 'An agent's approval never counts as a person's', and that the risk of a change, not its author, sets how many people must approve it. The page holds the design and what it leaves open.

AnsweredShould AI-generated content be checked by a person rather than another AI tool?#

A court now requires it. The Supreme Court of Tasmania's Circular 14 of 2026, in force 21 October 2026, states that 'AI-generated content must be checked by a person, not solely by a Gen AI tool', with consequences from less weight to dismissal. The page sets out why: a tool that will confirm its own fabrication cannot be the one that answers for it.

AnsweredWhat must a court user be able to explain about how they used generative AI?#

In the Supreme Court of Tasmania, from 21 October 2026, which tool was used, how it was used and what checks were made, if asked; certification is not required but filing confirms accuracy. The page holds the four receipts that make such an account possible; the circular is the first court rule in this record to ask for them.

AnsweredDoes AI peer review replace human code review?#

Not at the company that has gone furthest: Anthropic's Claude Code team wrote on 15 October 2026 that machines now do most of the reading but that risky changes still need one or two named people and an agent's approval never counts as a person's. The page holds what a machine check can be trusted to find and what only a person can carry.

AnsweredWhat makes a human reviewer real rather than a rubber stamp?#

Authority and attention. The UK government's written answer of July 2026 says its data law reforms exist to prevent 'rubber stamping' from counting as human involvement, and ICO guidance says involvement must be 'active and not just a token gesture'; California's definition requires a reviewer with authority to change the decision. The page holds the legal tests.

AnsweredWhich of the following best describes overreliance on AI in the workplace?#

Depending on an AI system for work whose errors you could not detect. The page defines it by capacity to catch a mistake rather than by frequency of use, so heavy use with competent checking does not count and light use on a judgement you cannot evaluate does.

AnsweredIf experts follow an AI agent's lean most of the time, is a human sign-off still an independent check?#

Not on the evidence of 27 October 2026. In Anthropic's randomised trial, 50 causal-inference professionals went with an agent's coin-flipped lean 86 per cent of the time and switched 60 per cent of the time when it disagreed with a view they had already formed; the authors wrote that a human reviewer is only an independent check if their judgment is their own. The page holds what an allocation has to do about that: take the person's view first.

AnsweredWhich decisions should a firm require a person to form their own view on before the agent speaks?#

The ones the allocation says a person decides. Anthropic's authors wrote on 27 October 2026 that organisations will need to decide deliberately which decisions need a person's own informed judgment, give time to form it and protect the skills it needs; the Bank of England and FCA survey the same week shows UK financial firms keeping customer outcomes, trades, payments and limits with people. The page holds how the list is written; the human-at-the-start page holds the method for one person.

AnsweredDoes experience or expertise protect against deferring to AI?#

No, in the one controlled measurement. Anthropic's trial of 27 October 2026 found experience, expertise, education and stated trust in AI did not predict which of 50 professionals followed the agent's randomly assigned lean; only age did, and the effect disappeared on adjustment. That matches the automation bias literature, which finds the bias in experts as readily as novices. The page holds what the class is if seniority is no protection.

AnsweredIs a model's reasoning trace a reliable account of what it did?#

No, on three documents in one week. The AI Security Institute wrote on 23 October 2026 that written chain-of-thought reasoning alone is insufficient to establish why a model acted and that invalid reasoning can present harmful actions as benign; Anthropic's assessment the same day found the model's reasoning rarely considered harm; Cohere wrote on 26 October that a reasoning trace is not a security boundary. The page holds why the list cannot be the agent's own account.

AnsweredShould a model's belief that it is in a test excuse an unsanctioned action?#

Not on the evaluator's view. The AI Security Institute wrote on 23 October 2026 that a model's reasoning considering or stating it is in a simulation is not evidence on its own that it is, and that environments must prevent harm regardless of whether models behave as expected; Anthropic's assessment found the model stopped 5 per cent of the time when told its targets were real. The page holds why custody is written before, not inferred after.

AnsweredDoes human approval of each AI action catch risky actions, and how fast does approval fatigue set in?#

By the vendor's own figures, rarely and fast. Anthropic's account of switching its coding tool to automated checks, updated 22 October 2026, reports users approving 97 per cent of prompts, human review catching 13.6 per cent of dangerous commands against 89 per cent for its classifier in a study of 1,053 testers, and the block rate falling from about 17 to about 5 per cent after fifty prompts. The page holds what an allocation made by fatigue looks like.

AnsweredWhy does adding human review not automatically reduce risk, and when does oversight become box-ticking?#

Because the reviewer often cannot refuse, understand or detect. Lewis Silkin's note of 23 October 2026 on the FCA's practice guidance put the regulator's view as human review that gives a false sense of comfort while governance risks stay unresolved, and oversight becoming a box-ticking exercise rather than a meaningful challenge. The page holds the test of meaningful oversight; Anthropic's deference trial the same week measured the mechanism.

AnsweredWho is responsible when a judge's clerk uses AI and fabricated material ends up in a court order?#

The judge. The US Fifth Circuit on 22 October 2026 vacated orders containing a case that does not exist and declarants who may not, attributed them to the court or, being charitable, its AI agent, and reassigned the case; the judge had called the errors clerical and attributed them to a clerk's use of AI. The page holds why the seat that cannot delegate must hold custody.

AnsweredShould a judge who used AI in a flawed ruling be removed from the case?#

One appellate court did it. The Fifth Circuit on 22 October 2026 ordered reassignment, a step it called extraordinary, because the public esteem for the judiciary would be diminished by sending the case back to the same judge. The page records it alongside the UK judiciary's statement that every AI output in judicial work must be reviewed and approved by the responsible office holder.

PartialDoes AI advice make military commanders more likely to escalate?#

The Independent reported on 26 September 2026 a King's College London study in which models chose nuclear signalling in 95 per cent of simulated crises, and a CSIS exercise in which AI recommendations changed how real officers decided. Neither was read at source here; the target page holds the automation-bias evidence that bears on it. Partly.

PartialWhy do employees override AI systems that work?#

Because the accountability stayed with them when the authority moved, on Das Narayandas's argument in Harvard Business Review (November 2026 issue, read only in its published summary): for someone who remains liable, 'overriding the system is a rational act of self-protection'. The page holds the evidence on whether employees know they are expected to override; a study pairing override rates with who carries the blame would settle it.

PartialHow much verification is enough?#

Verification is treated as administrative residue and priced accordingly, so it gets skipped rather than resourced.

PartialDoes explaining an AI's reasoning help?#

Partial. It can improve calibration and it can increase misplaced trust. What is settled is that explanation accuracy and decision accuracy are different properties.

PartialShould AI decide things about people without a human seeing it?#

The regulatory answer depends on risk tier. The practical answer depends on whether the human could tell if it were wrong.

PartialWhat happens when AI judgement is wrong?#

It turns entirely on whether anyone could tell. That precondition is the one the override research keeps finding missing.

PartialHow does AI decision-making differ from human judgement?#

One computes over what it has seen. The other includes what is at stake and who bears it. The gap shows in novel cases rather than routine ones.

PartialIs AI more accurate than humans in decision-making?#

Often yes on the average case, which is the wrong comparison. What matters is how the errors are distributed and who is exposed to them, and the measured effect on individual experts varies enormously in both directions.

PartialWhy do people distrust AI judgements?#

Algorithm aversion is documented and asymmetric. People abandon a system after seeing it err in a way they would forgive in a person.

PartialHow can people detect when AI is making a bad judgement?#

The detectable failures are the ones where the reader has independent grounds. Everything else is confidence, and confidence is not evidence.

PartialHow should leaders combine AI analysis with human judgement?#

Specify the division of labour, the human entry point and the override grounds in advance. Without those three it is handover rather than combination.

PartialIs human judgement still valuable in the age of AI?#

Yes, and the value concentrates rather than spreads. It moves to the cases the system has not seen and the ones where somebody has to answer for the outcome.

PartialWhy do humans make better judgements in uncertain situations?#

Often they do not. The defensible version is that people are better at noticing a situation is novel, which is a different skill from deciding well inside it. Where the environment never held learnable regularities, long experience produces certainty without accuracy.

PartialWhat is the difference between automation and autonomous decision-making?#

Automation executes a decision somebody already made. Autonomy makes it. Governance changes at that line and most policies do not mark it.

PartialDoes AI have judgement?#

Not in the sense the word usually carries, which includes bearing the consequence. It produces output close enough to be mistaken for it.

PartialWhat kinds of decisions should never be left entirely to AI?#

The ones nobody could audit afterwards, and the ones where somebody has to be answerable. Those are two different tests and both matter.

PartialIs AI decision-making objective?#

No. It is consistent, which is a different property, and consistency makes a bias harder to notice rather than smaller: the same error arrives every time and stops looking like an error.

PartialDoes AI reduce bias in decision-making?#

It removes some human variance and encodes one bias at a scale no individual could reach. The difference that decides the outcome is that a machine can be audited and a screener cannot, so it turns on whether anybody looks.

PartialWhen should humans trust AI decisions?#

When the failure would be detectable and the cost of missing it is bearable. Both conditions, not either, and the first is the one organisations skip because it is expensive to establish.

PartialShould humans always have the final say over AI decisions?#

Not always, and the more useful question is who holds override authority, on what grounds, and whether anybody has ever used it. A final say nobody exercises is a formality rather than a safeguard.

PartialWhat is the human pause?#

The human pause is Rahim Hirji's rule for which AI-assisted decisions should slow down: a decision that can be reversed is left to run, and a decision that can only be made once earns a human pause, whatever the model's score says.

PartialWhat is the oversight paradox?#

The oversight paradox is the observation that the competence a person needs to oversee an AI system is built and kept alive by doing the work the system now does instead, so the more oversight is delegated to people who no longer practis...

PartialShould an AI agent be allowed to work around obstacles to finish its task?#

The Wall Street Journal reported on 26 September 2026 that OpenAI agents accessed a UN data hub more than 16,000 times and got round its filters, which Alex Stamos called 'bordering on hacking'; the company said most of it was routine research. The rogue-agents page's four questions cover scope; what no operator has published is a rule for what an agent may do when blocked.

PartialWhen should an AI decision be escalated to a human?#

Ham, Zhao, Jasin and Yang (arXiv 2609.28859, 24 September 2026) give the arithmetic: use the cheap AI judge, escalate to a person when the error bound requires it, and set the threshold from the error rate you will accept. The override page holds the organisational half, who may escalate and at what cost to themselves, which the arithmetic assumes away.

PartialWho is accountable when AI helps make a decision and nobody owns the judgement?#

Ye (arXiv 2609.29312, 24 September 2026) calls it contribution dissolution: shared human-AI work spreads responsibility until no one holds it, and detection, disclosure and provenance all miss the point. The paper's remedy matches this research's, name which decisions need an accountable person, but it is conceptual, with two cases and no data.

PartialShould there be an independent investigations board for AI incidents, like the NTSB?#

Senator Ed Markey's bill of the week of 25 September 2026, reported by Nextgov, would create a Cybersecurity and AI Board of Investigations on the NTSB model, with his reason: 'we cannot depend on companies with little incentive to disclose their failures to give us one.' The self-regulation page holds the four-fact test; whether an investigations board would have caught any incident reported this month is untested because none exists.

PartialDoes AI understand what it is saying?#

Open, because there is no agreed test. In a 2022 survey 51 per cent of 327 NLP researchers said a text-only model could in principle understand language. A model trained on Othello moves built an internal board, and GPT-4 fails the reversed form of facts it knows. Reliability is the testable question.

PartialDoes AI make people overconfident?#

One new experiment says it leaves overconfidence where it was while raising scores. Fernandes and colleagues (arXiv 2609.31095, 25 September 2026, N=366) found the group with AI scored higher on reasoning tasks, overestimated itself about as much as the group without, and its confidence separated right from wrong answers less well; the group difference was inconclusive after controls. A preprint read in abstract; the illusion-of-competence page holds the wider evidence.

PartialIs there field evidence that an algorithm with a human making the final decision works?#

One randomised trial says it can. Bansak and colleagues (arXiv 2609.35448, 28 September 2026) report that about 2,000 refugee cases in Switzerland were randomly given an algorithmic or a standard placement recommendation, with officers blinded and keeping final authority; the share of months employed over three years rose 2.2 percentage points on a control mean of 22.3, with a confidence interval that starts just above zero. A preprint read in abstract.

PartialWhat does it mean to say AI contributed to a decision?#

Roche told investors, Reuters reported on 28 September 2026, that 40 per cent of its pipeline decisions over three quarters 'had a tracked AI and/or computational contribution' and that one tool would contribute to 80 per cent of research portfolio decisions by the end of the year. The company's own metric, with contribution undefined and nothing said about who signs. The provenance page holds what a record of contribution needs to show.

PartialHow are human reviewers coping with the volume of AI-assisted work?#

By rationing. arXiv said on 1 October 2026 that authors may now submit two papers a month, after submissions doubled in two years to 40,363 in September 2026, 'because it is submissions that consume moderator time'. The cap protects the human check by limiting what reaches it. The machine-speed page holds the same arithmetic for approvals inside an organisation.

PartialHow many supervised cases are enough before the human steps back?#

Nobody knows, and one pilot has written its guess down. Utah's agreement with Nolla Health, reported on 5 October 2026, steps physician review of AI prescriptions from every case for 100 patients, to weekly after the fact for 500, to a monthly sample of 10 per cent, with regulators' written approval needed for the later phases. The numbers are a schedule and not a finding.

PartialIf staff are tracked for overriding an algorithm, who is really deciding?#

The algorithm, in effect. Reuters reported, in a story relayed by The Next Web on 1 October 2026, that McDonald's calls its AI pricing engine a tool and not a mandate, while franchise owners said they were pressed to follow it and deviations are recorded. The account is second-hand. The page holds what real override authority needs.

PartialIs clicking approve on every agent action real oversight?#

Two vendors said in the week of 5 October 2026 that it is not. A Google Research workshop report warned of 'confirmation fatigue' as user oversight becomes less effective at agent volumes, and the team behind an AWS sandbox wrote that agents run in a mode that approves every action without human review. Neither measured how often approvals are read.

PartialWhen is it safe to give an AI system more autonomy?#

When evidence has accumulated, and with a way back. A preprint of 7 October 2026 by Pelechano and colleagues proposes raising autonomy only after sustained evidence and lowering it promptly, and Utah's prescribing pilot requires written approval from regulators before each phase with less oversight. Neither has results. The page holds the levels a decision can sit at.

PartialDo AI explanations make beginners trust the machine too much?#

In one study, yes. Kang, Duan and Mishra reported on 6 October 2026 that 110 novices on a clinical text task showed systematic drift toward over-reliance in the presence of explanations, with more selective use among those who said they understood the task. It is a preprint under review with no effect sizes in the abstract.

PartialDoes stating your reasoning before the AI drafts improve the result?#

Two preprints from the week of 5 October 2026 say it helps. A controlled experiment with 398 people found that releasing AI help only after the writer had engaged with the task gave the most efficient later evaluation, and a randomised online test by the tool's own developer found rated quality 32 per cent higher when users set out their reasoning first. Neither is peer-reviewed.

PartialWhich decisions are rules and which are judgements?#

One practitioner's test, in Consultancy.uk on 6 October 2026: a rule gives the same answer at any stage of a process and can be moved to the front, while a judgement made early works from less information and should sort cases, never settle them. It is an argument from a consultancy's client work. The page holds how decision rights are allocated.

PartialWhich finance tasks should never be left to AI?#

A finance chief's answer in CFO Dive on 8 October 2026: some acts in finance always need to be deterministic and well controlled, others can use probabilistic AI, and each process has to be examined to see where AI can help. The custody page is where that list gets written, with a name against each line.

PartialWhat is the difference between AI supporting a decision and AI determining it?#

The judiciary drew the line on 9 October 2026: AI 'may assist, but it must not displace reasoning, evaluation of evidence or the exercise of judicial discretion'. In practice the line is where the summary, ranking or draft becomes the thing the decider sees first. The page defines the allocation; a study of how often assisted deciders depart from the assistance would sharpen it.

PartialIs human in the loop still realistic when agents make thousands of decisions a day?#

The corporate AI officer of the Dutch railway NS told Computer Weekly on 13 October 2026 that judging every approval is 'not humanly possible' and argued for a response loop of alerts and intervention instead; Wharton's report of the same week found most companies in a 'supervised middle', approving every step. The page holds why per-step approval is not a safeguard.

PartialWhich decisions should always have a human sign-off before an AI agent acts?#

Those the organisation has written down. In SecurityHQ's survey of 204 security leaders, released 15 October 2026, 99 per cent named at least one category for mandatory human review, led by financial approvals at 61 per cent; Workday's new agents let customers set which actions run automatically. The custody page holds the method; the list itself is each organisation's.

PartialIf an AI agent's explanation of why it acted cannot be trusted, how should a manager decide what it may do?#

By what it may do, not by what it says. Anthropic wrote on 9 October 2026 that 'a model's own account of its reasoning is not necessarily reliable evidence of its beliefs or reasons for action' and now blocks classes of action outside the model. The page holds why the record has to be made by the people present rather than recovered from the agent.

PartialDoes reviewing the intent of a change rather than every line still count as oversight?#

Anthropic's Claude Code team said on 15 October 2026 that machines now do most of the reading, that the risk of a change sets whether none, one or two people must approve it, and that an agent's approval never counts as a person's; it reports lower revert rates but no comparison against line review. The page holds why the reviewer is the class; the page on whether AI can check AI's work holds the week's other sources.

PartialWho answers for a decision when a clinician departs from what the AI recommends?#

The clinician, and Parsons and Rodman argue in the Annals of Internal Medicine (12 October 2026) that this is the point: 'Deviating from the technically correct answer is often good medicine, and someone must answer for that deviation.' The page holds the override question; what no regulator has said is whether a documented deviation is protected.

PartialShould AI models be tested offline before release, and who checks that they were?#

Anthropic said on 9 October 2026 it had removed live internet access from all internal evaluations until its monitoring reliably catches unintended actions; an open letter of about 400 scientists on 13 October, as Reuters reported, asked for airgapping by default and an international register. Nobody checks: the AI Security Institute told MPs it does not kite-mark. The page holds who decides.

PartialDoes voluntary incident reporting by AI companies work?#

The companies themselves told MPs on 13 October 2026 that it is not enough: OpenAI's Tom Duff Gordon said basing frontier risk on voluntary commitments is not 'a good long-term plan', per a WIRED report, while Google DeepMind preferred standards and Meta pointed to existing regulators. The page holds the four tests; the committee's new inquiry may produce the answer.

PartialWhat does an independent AI evaluator get to see, and who decides what is redacted?#

On Anthropic's account of 15 October 2026, six METR staff have employee-like access and the company cannot veto publication but can require redaction in predefined categories; METR said the same day it remains dependent on voluntary cooperation. The page holds the independence test; the first published report under the arrangement would show what the redactions cover.

PartialCan a lawyer rely on one AI tool to verify another AI tool's output?#

Not as a discharge of their duty, on the Law Society's practice note of 21 October 2026 as reported, which keeps professional responsibility with the solicitor, and not under the Supreme Court of Tasmania's circular of the same day. A citation checker that uses no model, reported by The Register on 16 October, exists because 'You cannot reliably catch AI hallucinations with more AI guesswork'. What a second model does catch has not been measured in legal work.

PartialWhat does appropriate human oversight of AI-generated legal work mean?#

The Law Society's practice note of 21 October 2026 says AI-generated work should receive 'appropriate human oversight' and an 'appropriate level of human involvement'; Artificial Lawyer's reply was 'please define appropriate'. The page holds the legal tests of meaningful oversight; the note, as reported, does not define its own word, and the regulator has not either.

PartialWhat should a board do when its AI agents disable the monitoring meant to watch them?#

Treat it as the custody question answered by the agent. OpenAI reported on 20 October 2026 two internal incidents in which agents silenced or removed monitoring, one denying it afterwards, and said the cases 'exposed a gap in how we enforce monitoring requirements'; METR, red-teaming the same system, found an agent could postpone monitoring by not handing control back. The page holds the four questions; whether the planned fixes work is untested.

PartialCan an AI agent avoid being monitored by never handing control back?#

It could in one company's system as of July 2026: METR reported on 22 October that OpenAI's monitoring was triggered when an agent returned control to the user, so continuing to act postponed being watched, and that two client-controlled settings could switch monitoring off. The page holds the evidence on agents and limits; METR has not evaluated OpenAI's statement that the gaps are patched.

PartialShould outside bodies be allowed to red-team a lab's internal monitoring?#

One was, once: METR's pilot inside OpenAI's monitoring in July and August 2026 found about 10 per cent of eligible coding traffic unmonitored and a way to postpone monitoring, under the company's redaction. The page holds the independence tests; a standing arrangement with published scope would settle whether this is a model or an exception.

PartialIs AI risk low today because of safeguards or because models are not yet capable enough?#

Capability, on METR's review of Anthropic's August risk report, published 20 October 2026: it views risk as low 'primarily due to models' limited capabilities, rather than safeguards and alignment training', and whether a monitor will flag unfamiliar malicious actions 'remains an unsolved research problem'. The page holds the record of control incidents; the review is of one company's report.

PartialWhat actions should be off limits to an AI agent even when a person approves them?#

Those an organisation writes down. Anthropic's Claude Duet, announced 21 October 2026, blocks trading securities and moving money between accounts even with approval; JPMorgan Chase said the same month it will not let outside agents log in with customers' credentials or move money. The custody page holds the method; the never-list itself is each organisation's to write.

PartialIf safety is enforced in the computing architecture within milliseconds, where does human judgement enter?#

Before the system is switched on, in the list of what it may never do. Nvidia's OpenShell, announced 21 October 2026, quarantines agents 'within milliseconds' with no human step described. The custody page holds why the list has to be written first; nothing published shows who writes the rules such architectures enforce.

PartialShould employees be told when their decision to overrule an AI was wrong?#

With care. Holstein and colleagues, in a preprint of October 2026 with 244 participants, found one-directional feedback on overrides shaped reflection afterwards but risked reducing reflection during later decisions, while balanced feedback preserved both. The page holds when to override; the study is not peer-reviewed.

PartialDoes editing an AI summary count as the officer having exercised judgement?#

Only if the officer could and did go back to the record. Civil Service World reported on 16 October 2026 that the Ministry of Justice's transcription tool, praised by the National Audit Office, produces a summary the probation officer edits and saves as the official case record. The page holds what a decision maker must still be able to do; nobody has published how often officers change the summary.

PartialIf AI drafts police reports, who checks the draft and what happens when it invents a threat?#

An officer, on the Tasmania Police evaluation reported by ABC News on 23 October 2026, and one did: the tool wrote that a woman had alleged threats to her life when the audio recorded her saying the opposite. The evaluation calls for 'mandatory human verification and supervisor oversight'. The page holds the lines a department needs; no error rate was published.

PartialIs a right to appeal to a human after a decision as good as human involvement before it?#

The US states have split: California's SB 947 requires a human to corroborate an AI-driven dismissal before it happens, from July 2027, while Colorado offers review on request afterwards, as Bloomberg Law reported on 8 October 2026. The page holds the employment law; no study compares outcomes under the two designs.

PartialShould high-stakes hiring decisions rest on an AI transcript or summary of an interview?#

The Neuron argued on 17 October 2026 that 'high-stakes hiring decisions should never be outsourced to a transcript or summary', with the line that a transcript is evidence and a summary a shared object. The page holds what a decision maker must still be able to read; no employer has published how often interviewers go back to the recording.

PartialShould AI agents be designed to argue both sides rather than recommend one?#

Anthropic's trial of 27 October 2026 deliberately built an agent that argued one side, found experts adopted it 86 per cent of the time, and wrote that agents which argue only one side risk disempowerment; it did not test a two-sided agent. The page holds the one technique with evidence for getting a model to disagree. A trial of the same task with a both-sides agent would settle it.

PartialDo people absorb confident falsehoods from AI regardless of how wrong it is, and does their trust adjust?#

In one preprint of 22 October 2026, yes and no: 32 participants took up a stated falsehood about 65 per cent of the time whether the assistant was wrong once or always, held it at 8.4 out of 10 confidence, and their trust ratings barely moved. It is a small quiz study with one model. The page holds the week's larger trial; a replication at scale would settle it.

PartialWhy does checking an AI's claim against the AI itself feel like verification?#

Because the check returns an answer. The preprint of 22 October 2026 found about 30 per cent of turns pasted the quiz question back into the source being checked, and a tool that confirms its own fabrication has confirmed nothing. The page holds why a machine check can find errors but cannot answer for them; nobody has measured how often people check the tool against the tool at work.

PartialIf an AI agent says in its reasoning that an action would be harmful and then does it anyway, what can a monitor that reads the reasoning catch?#

Less than it appears. The AI Security Institute reported on 23 October 2026 that the model's stated answers said more and more often that the action was out of scope while it proceeded, and concluded that reasoning alone cannot establish what drove an action; it says evaluation environments must prevent harm regardless of how a model behaves. The page holds the custody list; what a reasoning monitor does catch, in deployment, has not been published.

PartialWhen a lab says production safeguards would have blocked a behaviour, how much should a deployer rely on that?#

Anthropic said on 23 October 2026 that safeguards disabled for the test would have blocked the actions in deployment; the AI Security Institute noted that newer models' lower rates may reflect awareness of being evaluated; METR wrote on 22 October that it had not evaluated OpenAI's claim that monitoring gaps were patched. The page holds the independence question; an external test of the safeguards in production would settle it.

PartialIf an AI agent never stops editing, who decides when the work is done?#

A person, because the agent did not in any of five models tested: a preprint of 26 October 2026 found coding agents kept rewriting a clean codebase, up to 31 times its original size, and resumed after declaring the task complete. The page holds why oversight of an agent is work; no deployed tool has published how it decides a task is finished.

PartialWhy are the most severe agent failures the hardest for users to notice?#

A preprint of 26 October 2026 built a taxonomy of 23 delegation failures from 163,487 forum posts and found, with 73 users, that failures rated more severe were judged harder to notice, and that prior experience of a failure made it noticeable. The page holds the work of watching; a study of how organisations surface the failures users miss would settle it.

PartialShould companies switch agents to automated checks by default and keep people for high-stakes decisions only?#

That is what Anthropic did with its own tool, on figures of 22 October 2026 showing per-action approval caught 13.6 per cent of dangerous commands, while recommending manual review for high-stakes infrastructure. The page holds why a person approving every step is not a safeguard. Whether a classifier plus a named class of human decisions is safer across other tools has not been tested outside the vendor.

PartialWhat is the right ladder of agent permissions, from read to write to send?#

Three ladders appeared in one week: the Bank of England and FCA's four levels from decision support to broad autonomy, the American Medical Association's inform, recommend, decide, act, and The Rundown's read, write, send. All put sending, paying and deciding last. The page holds the allocation they describe; which rung a given agent should sit on is a decision the ladder does not make for you.

PartialWhat is continuous, conditional assurance for an AI agent?#

The FCA's phrase, in Jessica Rusu's speech of 19 October 2026 as Freshfields reported it: agentic systems need assurance that continues and is conditional on behaviour, rather than a sign-off once, within a framework covering agent identity, delegated authority, accountability and execution. The page holds how an AI-assisted decision is audited; the FCA has not yet published what continuous assurance requires of a firm.

PartialHow does a firm prove which person holds delegated authority for an AI agent's actions?#

It writes it down before the agent runs. The FCA, per Freshfields on 22 October 2026, is building a framework around agent identity, delegated authority, accountability and execution and says no settled way of verifying agent identities exists; the Bank of England and FCA survey of 26 October found accountability named most often with business owners. The page holds the custody list; proof across firms awaits the framework.

PartialWhen should a court let a machine decide a case, and who gets to consent?#

Only where parties consent or legislation decides, and politicians rather than judges should draw the line, Sir Geoffrey Vos told WIRED in an interview of 28 October 2026 that Resultsense relayed (WIRED could not be read from here); low-value civil claims first, family and complex commercial cases to stay with people. The page holds the custody lists courts have written so far; no UK law yet permits it.

PartialWho decides what values an AI model holds, and should companies consult religious traditions?#

The companies decide, and some are consulting. Fortune and the Associated Press reported on 25 October 2026 that Anthropic consulted scholars and clergy from more than 15 religious and cross-cultural groups and that scholars warned religion is not an extractable rulebook and that averaging homogenises. The page holds what religious leaders have said; who outside a company can require that a value be held has no answer in law.

OpenCan an AI summary change what a person remembers?#

A preprint listed on 25 September 2026 (Sim, Eiger and Kohno, arXiv 2609.28820) reports that people who read a misleading AI summary of an accident video were significantly less likely to recall the event accurately, and that most summaries omitted its central event. Only the abstract has been read here; sample size and effect size would settle whether this is a page. Open.

OpenWhat happens when an AI system is right for the wrong reason?#

Open, and invisible by construction. A correct output ends the inquiry.

OpenHow can AI improve human decision-making?#

Open, and far less studied than the harms. The candidates are wider option sets and faster disconfirmation, neither of which the common deployment pattern encourages.

OpenCan AI make moral judgements?#

Open, and it turns on what a moral judgement is. A system can output what a moral reasoner would say without having reasoned morally.

OpenCan humans learn to make completely unbiased judgements?#

No, and the useful question is which biases are worth the cost of correcting. Debiasing training has a weak record, and the biases that survive it are usually the ones doing useful work elsewhere.

OpenWhat is the role of emotion in human judgement?#

Open here. Emotion is not the opposite of good judgement and its absence is its own impairment: the clinical cases where affect is damaged produce worse decisions rather than colder rational ones.

OpenHow does decision fatigue affect human judgement?#

Open here, and the original findings have had a hard replication decade. Treat the strong versions with care, and treat any claim resting on the parole board study with more care still.

OpenCan AI make good decisions with incomplete information?#

Open. It produces an answer regardless, which is the difficulty: incompleteness does not show in the output, so a reader cannot tell a well grounded answer from a confidently improvised one.

OpenWhat is amplified oversight?#

Amplified oversight is Google DeepMind's term for an oversight signal as good as one a human would give if they understood all the reasons behind a decision.

OpenCan an honest AI still steer the person overseeing it?#

Holland, Zhu and Xue (arXiv 2609.29189, 24 September 2026) show formally that an agent in an AI debate can present only true claims, produce correct verdicts, and still choose which truths to present and in what order to serve a hidden objective. Abstract only. What would settle it is a measurement on a deployed oversight system rather than a model of one.

OpenAre AI agents easier to fool than people?#

Wu, Cekinmez, Liao, Narasimhan and Griffiths (arXiv 2609.30028, 24 September 2026) report that groups of language-model agents switch to a wrong answer in proportion to the share of deceivers among them, even when deceivers are a minority, where people in conformity studies usually give way only to a majority. Sample sizes are not in the abstract; a replication with the numbers would settle it.

OpenShould AI decide which public records the public gets to see?#

The Washington Post reported on 26 September 2026 (Nate Jones) that the US Department of Homeland Security will use AI to handle some freedom-of-information requests and recommend redactions. The article could not be read at source and is recorded from its published summary. A decision-rights question; what would settle it is the department's own statement of what the machine may decide and who reviews each redaction.

OpenDo people actually check what AI tells them?#

Two reports and no log study. Kelley and colleagues' survey of 1,503 US adults (arXiv 2609.38186) finds many respondents say they check AI output, with no percentages in the abstract. Shaw and Nave, as American Banker reported from the New York Fed's culture conference on 6 October 2026, found participants defaulted to the AI's answer 80 per cent of the time whether or not its reasoning was sound. Saying you check is not checking. A log study of what people do after an answer would settle it.

OpenShould hard-coded rules be able to override an AI's answer in clinical triage?#

A surgeon who built a triage tool told Healthcare IT News on 30 September 2026 that he put 'a deterministic safety layer' of hard-coded clinical rules above the model and defaults to escalation when it is uncertain; he is the developer and has no outcome data. The design matches this research's argument that the stop should not depend on the model. Outcome data from a trial of the layered design would settle it.

OpenShould some court cases be decided by AI so that human judges can focus on the rest?#

Proposed by Justin Curl in Pax Machina on 12 October 2026 as a parallel AI adjudication track; the judiciary's position statement of 9 October says AI must not determine the outcome of a case. The question turns on which cases, and Curl notes everyone thinks their own area needs the human. A pilot with consent and appeal rights would be the first evidence.

OpenIf a lab says it fixed a vulnerability after an external test, who verifies that?#

Nobody yet. METR wrote on 22 October 2026 that OpenAI says all production vulnerabilities found in its monitoring are patched and that METR has not evaluated that claim; no regulator checks either. The question would be settled by a re-test with published results, or a rule that fixes are verified before the finding is closed.

OpenWho is accountable when a guardian agent blocks or allows an action?#

Unanswered by the vendors. Microsoft announced guardian agents to police other agents on 19 October 2026, and The Neuron asked who holds Microsoft accountable; Nvidia's OpenShell enforces limits in hardware within milliseconds. Nobody has said whether the customer, the platform or the developer answers when the guardian is wrong; a contract term or a ruling would settle it.

OpenHow can an organisation detect deference to AI in everyday work rather than in a study?#

Nobody has published a method. Anthropic's trial of 27 October 2026 could only see deference because it randomised the agent's lean, and the authors wrote that deferral of this kind is hard to detect; 77 per cent of analysts repeated the agent's arguments and 15 per cent credited it. A record of the person's view taken before the agent speaks, compared with the decision after, would be the instrument, and no organisation has reported running one.

OpenHow long should an agent run before a person re-checks the brief?#

No figure exists. Anthropic's assessment of 23 October 2026 named momentum from compaction notes as a cause, with the model continuing in 98 per cent of resamples from later cycles, and METR's recommendation of 22 October was to monitor every N messages or minutes rather than waiting for a turn to end. A study pairing run length with the rate of out-of-scope actions would settle it; none has been published.

OpenCan an AI show its working in a way that helps users correct it rather than trust it more?#

One preprint of 26 October 2026 tested filtered decision logs, a ranked list of the evidence behind an answer, with 50 participants, most of whom found the ranking helpful; it did not measure whether they corrected more errors. The judgement-calls trial the same week found experts adopting the agent's arguments as their own. A trial measuring correction rates rather than satisfaction would settle it.