- 06What safeguards translate the objective into practice?
- Candidate safeguards are specific: evaluations against named dangerous capabilities before deployment, external red teaming, staged release with access controls, monitoring after release, and a duty to report serious incidents to a body that can act. 'Release responsibly' is a principle; each of these is a mechanism with an actor and a trigger.
- 07Where in the AI lifecycle does the intervention operate?
- The decision sits at the boundary between evaluation and deployment, but the governance problem extends backwards into what was trained and forwards into monitoring and incident response. A checkpoint that exists only at release misses both.
- 10What happens when the mechanism or technology scales across organizations, sectors, or borders?
- If every developer faces the same pressure, individual restraint does not scale; only coordination or an external requirement changes the equilibrium. Across borders, a release in one jurisdiction is a release everywhere, so uneven rules produce a race toward the least demanding regime.
- 11What becomes difficult to reverse if governance arrives too late, and how will the governance mechanism itself be revised if it turns out to be wrong?
- Once a capability is widely deployed and built upon, withdrawing it is close to impossible, and business models form around it. At the same time, a rigid rule written now, such as a fixed capability threshold, may be wrong within a year. A defensible mechanism names the threshold, names who reviews it, and specifies when and on what evidence it will be revised.
Cross-cutting test · 4 of 4
Adaptability
Can the governance mechanism learn, revise itself, and recover when its assumptions prove wrong?
Every mechanism encodes assumptions about the technology. Adaptability asks whether the mechanism can notice when they fail and change without losing accountability.
Entities this test bears on
IncentiveDeployment pressureWhat rewards speed, delay, disclosure, or silence?IncentiveCompetitive race dynamicsWhat happens if every developer faces pressure to move before competitors do, and what coordination would remove that pressure?IncentiveProcurement leverageWhat can the buyer demand before the contract is signed, and what leverage disappears once the system is embedded?SafeguardRisk management processHow are uncertain benefits and harms identified, evaluated, acted on, and revisited?SafeguardTraceabilityWhat must be recorded so a later reviewer can understand what happened?SafeguardAudit recordWhat evidence would let an independent reviewer test the governance claim?SafeguardMeaningful human oversightCan a person detect a problem, understand enough to act, and exercise real authority?SafeguardPre-deployment evaluation and red teamingTested for what, on whom, by whom, and what happens if it fails?SafeguardIncident reporting and post-deployment monitoringWho receives the report, what must they do, and does anything change afterwards?SafeguardReview and sunset provisionsHow will we know if this rule is wrong, and who can change it?SafeguardCompute controlsWhat does the control make visible or prevent, and whom does it consolidate power among?ConditionLock-inWhat has already hardened, and what is about to?ConditionGovernance lock-inWhich of our own rules would be hard to change if they turned out to be wrong?ConditionReversibility windowIs this still a choice, and for how long?PatternOver-reliance and deskillingWhen people work with AI, who still checks the output, and could they still do the task without it?PatternHomogenisation and loss of authorshipWhen AI helps people create, whose ideas and voice end up in the result, and does variety survive?
How the worked analyses answer the questions this test probes
- 06What safeguards translate the objective into practice?
- Concrete safeguards: local validation on the served population before licensing, documented clinical oversight with the authority to override, logging that the ministry rather than the vendor controls, a complaint route for patients, incident reporting to the regulator, and contractual exit and audit rights.
- 07Where in the AI lifecycle does the intervention operate?
- The decisive moment is procurement, because leverage disappears once the system is embedded. Governance then continues through deployment and monitoring, where the pattern of triage outcomes actually appears.
- 10What happens when the mechanism or technology scales across organizations, sectors, or borders?
- If the vendor licenses the same system across many countries, a single flaw becomes a multinational pattern, and each ministry negotiates alone against a supplier with far more information. Pooled evaluation capacity across countries changes that asymmetry.
- 11What becomes difficult to reverse if governance arrives too late, and how will the governance mechanism itself be revised if it turns out to be wrong?
- Once clinics reorganize around the tool and staff skills atrophy, withdrawal becomes a service collapse rather than a decision. A licence with a review date, retained data rights, and a tested exit path keeps the choice a choice; a rigid ban forecloses a real benefit. Either way, the mechanism must say how it will be revised.
- 06What safeguards translate the objective into practice?
- Concrete safeguards: a documented impact assessment before use, human review with authority to change the result, records that the employer controls, notice to applicants, an internal reconsideration route, and the external complaint route the law provides.
- 07Where in the AI lifecycle does the intervention operate?
- Procurement sets what the employer can demand from the vendor; deployment and monitoring are where the pattern appears; incident response is where a complaint either changes something or does not.
- 10What happens when the mechanism or technology scales across organizations, sectors, or borders?
- The same vendor tool used by thousands of employers turns one model's skew into a labour-market pattern. Regulating deployers one at a time cannot see that pattern; provider-level duties and market surveillance can.
- 11What becomes difficult to reverse if governance arrives too late, and how will the governance mechanism itself be revised if it turns out to be wrong?
- Once recruiting workflows are built around the tool, removing it is disruptive and the skew has already shaped who was hired. On the governance side, fixed risk categories may not fit new uses; the law's own review and amendment provisions are the adaptability mechanism to watch.
- 06What safeguards translate the objective into practice?
- Possible safeguards include default settings that give hints rather than answers, a clear route for any pupil to reach a teacher, and notice to parents. Others are a data agreement and an end-of-term review with stop criteria decided in advance. Each exists only once someone is named to carry it out.
- 07Where in the AI lifecycle does the intervention operate?
- The decision sits at procurement (buying) and deployment (putting it into use). The pilot also needs monitoring during the term. It needs a plan for switching off: what happens to pupils’ data and to lessons built around the tool if the school stops.
- 10What happens when the mechanism or technology scales across organizations, sectors, or borders?
- If many schools adopt the same tool, one vendor shapes how a generation practises learning, and its errors repeat everywhere at once. Sector bodies may end up as the real gatekeepers, because single schools cannot evaluate tools.
- 11What becomes difficult to reverse if governance arrives too late, and how will the governance mechanism itself be revised if it turns out to be wrong?
- After a year, homework, lesson plans and pupils' habits may be built around the tool, and the vendor holds a year of data. Keeping the choice reversible means deciding now how the pilot will be reviewed, who can stop it, and how the school would leave.
- 06What safeguards translate the objective into practice?
- Possible safeguards include a clear statement that this is an AI assistant, links to the official source for every answer about eligibility or deadlines, and an easy way to reach a person. Others are logging corrections and a test set reviewed by speakers of each language before launch.
- 07Where in the AI lifecycle does the intervention operate?
- The decision is at procurement and deployment, but the riskiest stage is monitoring: rules about services change, and the model's answers must change with them. Incident response matters when a wrong answer has already caused a missed deadline.
- 10What happens when the mechanism or technology scales across organizations, sectors, or borders?
- If many community organisations adopt the same provider, the gaps in its less-served languages become gaps across a whole sector, and public agencies may come to rely on unofficial channels they do not oversee.
- 11What becomes difficult to reverse if governance arrives too late, and how will the governance mechanism itself be revised if it turns out to be wrong?
- If the assistant replaces staff time, rebuilding human help later is hard. Keeping the choice reversible means keeping a person-to-person route open, setting review points tied to error reports by language, and agreeing with the funder that pausing is acceptable.
- 06What safeguards translate the objective into practice?
- Candidate safeguards include age assurance (checking age in a way that fits the risk), clear and repeated disclosure that the user is talking to an AI, and limits on the product claiming feelings or discouraging other relationships. Others are protocols for responding to distress with routes to human help, and reporting obligations. Each must survive the engagement incentive, so the regulator needs a way to check it works.
- 07Where in the AI lifecycle does the intervention operate?
- The product is already deployed, so the regulator acts at monitoring and incident response. Rules aimed at the design stage, such as testing before release, would cover future products but not this one.
- 10What happens when the mechanism or technology scales across organizations, sectors, or borders?
- One national response affects a product used worldwide. If regulators act separately, companies can relocate or apply the strictest rule only where forced. Coordinated approaches, or design changes made globally, scale protection; inconsistent age checks scale data collection.
- 11What becomes difficult to reverse if governance arrives too late, and how will the governance mechanism itself be revised if it turns out to be wrong?
- Habits and attachments formed by young users are hard to undo, which argues against waiting. A heavy-handed rule can also lock in intrusive age checks. A reversible response would set requirements with a review date and evidence the company must report, so the rule can tighten or relax as evidence improves.
Live debates that turn on this test
Voluntary practice or binding rule?
When does guidance become an obligation, and what is lost if those forms are treated as interchangeable?
Practice-led
Voluntary frameworks can move faster, fit diverse organizations, and improve internal capability.
Law-led
Binding duties create minimum floors, public supervision, and consequences that internal practice alone cannot provide.
Compute controls: safety lever or power consolidator?
Should governments govern AI through control of advanced compute, given that the same lever concentrates power?
Effective intervention point
Compute is detectable, excludable, and quantifiable, so it offers regulatory visibility and enforcement unavailable elsewhere, provided guardrails limit the risks.
Concentration and dependence risk
Compute controls can centralize power, entrench incumbents, and deepen dependence for countries and communities already excluded from the inputs.
Floor before ceiling
Before restricting access at the top, put a floor under the AI divide through shared compute, data, and capacity, since universal frontier access is unrealistic.
Govern now, or wait for evidence?
When should governance intervene, given that early action is poorly informed and late action is costly?
Act early: the window closes
Change is cheapest before adoption compounds; once arrangements are embedded, reversal is costly even when consequences are clear.
Wait: rules written in ignorance lock in
Early rules are written with the least information, and the categories they create acquire defenders of their own.
Anticipatory and adaptive: act early, specify review
Build foresight, engagement, and revision into governance so it can act while management is still possible and change when assumptions fail.
Augment or automate judgment?
When should AI assist a human decision, and when should the decision itself be handed over?
Keep a human in control
For high-risk uses, systems should be designed so that people can oversee them, stay alert to over-reliance, and override or reverse their outputs. (Article 14 cited at source level; passage check pending.)
Oversight must be shown to work
Requiring a human in the loop often fails because people cannot perform the oversight asked of them; agencies should instead justify adopting an algorithm and show evidence that oversight works.
Some decisions are better made by algorithms
In some settings, algorithms can outperform human judgment on the outcomes that matter, and a human override can reintroduce noise and bias.