THE SIGNAL IN ONE SENTENCE
Europe has drafted a form for asking what a high-risk AI system could do to people's rights. That is progress. It is also exactly the sort of thing that can become excellent paperwork attached to a lousy decision. The European Commission's AI Office published a draft questionnaire on November 16 for fundamental-rights impact assessments under Article 27 of the EU AI Act. Certain organizations that deploy high-risk AI systems will have to complete one before putting the system into use. The Commission is asking for public feedback until January 15, 2027. It also wants affected deployers to volunteer by December 4 to test the draft on one of their own systems. That second invitation is the useful part. A rights-impact assessment should not be a writing exercise performed after a procurement team has already fallen in love with the product. It should be a field test capable of changing the design, narrowing the use, delaying the launch or stopping the deployment. Start with the scope. Article 27 does not require every person who opens a chatbot to complete a government questionnaire. It applies to specified deployers of certain high-risk systems listed under the AI Act, including public bodies and private entities providing public services in relevant cases. The Commission says the obligation is scheduled to apply from December 2, 2027. The draft is intended to guide those deployers through the elements named in the law and to clarify how their responsibilities relate to the provider's responsibilities. Once final, the template is also supposed to become an automated tool on the AI Act's Single Information Platform. Automation will make the form easier to complete. It will not make the answers true. Imagine a city assessing an AI system used to prioritize housing inspections. The form may ask for the purpose, the people affected, the foreseeable harms and the mitigations. A polished answer could say that a human makes the final decision and that staff can correct errors. Now walk into the actual office. Does the inspector see the reason behind the score? Is there enough time to challenge it before the day's route is locked? Does correcting one address update the data source or only today's screen? Can a tenant find out that a system influenced the inspection? Can someone appeal when a low score delays attention to mold, heat or structural danger? Does the queue behave differently in neighborhoods with weaker historical records? The form needs evidence from that room. For every important answer, require a supporting artifact. Attach the data map, interface screenshot, test result, policy, training record, complaint path, meeting note or decision log that makes the statement inspectable. If the answer says human oversight, show the person, the authority and the time available. If it says people can appeal, show how they learn about the decision, where they appeal, how quickly someone answers and whether the original harm can still be repaired. If it says the system is accurate, show the relevant population, error measure, comparison process and failure slices. A single average can hide the group that pays for the mistakes. If it says the risk is mitigated, name the owner, deadline, budget and test that will prove the mitigation works. An impact assessment without attached evidence is an opinion with numbered fields. The people affected need a seat before the answers harden. Bring in workers who operate the system, people who receive its decisions, advocates familiar with the setting and specialists who can identify accessibility, discrimination, privacy and due-process problems. Pay community participants for their time. Provide material in usable languages and formats. Explain the proposed system without vendor theater. Then give them scenarios. Do not ask only whether the system sounds fair. Ask what happens when the address is wrong, the identity is duplicated, the person cannot use the digital channel, the model is uncertain, the record reflects historic discrimination, the network is down or a worker has twenty seconds to review the recommendation. Run the ugly cases. The assessment also needs a baseline. The relevant comparison is not an imaginary world in which the existing process is perfect. Human systems can be biased, inconsistent, slow and opaque. Measure the current process honestly, then test whether the AI system improves the outcome without creating a new harm that the old process did not have. For a benefits decision, compare error rates, waiting time, explanations, appeals, reversals and burdens on claimants. For hiring, compare who advances, why qualified people are rejected, how accommodations work and whether recruiters become less attentive after seeing a model score. For education, measure access, false flags, teacher override and the consequences of being labeled at risk. Keep the baseline and the new system on the same scoreboard. The draft consultation arrives at a useful time because deployers still have room to shape the instrument before compliance becomes a yearly ritual. The AI Office says selected volunteers will apply the template to one of their own systems with support, then discuss the results in a closing session. Those pilots should publish what they can. Remove personal and security-sensitive details, but share which questions were confusing, which evidence was unavailable, how long the work took, who participated, which risks changed the deployment and which mitigations failed during testing. A template improves faster when people can see where the first version met reality. The hardest field should be the stop condition. Every assessment should name the result that blocks deployment. That could be an unexplained error gap, an inaccessible appeal path, an inability to trace the data, an override that workers cannot actually use, a security weakness, an unacceptable false-negative rate or a harm that cannot be repaired after the fact. Without a stop condition, the assessment can become a catalog of concerns that never changes the launch date. There also needs to be a restart condition. The Commission notes that deployers must update the assessment when relevant elements change or are no longer current. Turn that legal idea into an operational trigger. Reopen the assessment when the provider changes the model, the deployer adds a new data source, the system moves to another population, the interface changes, the threshold moves, a serious incident occurs or complaints show a failure that testing missed. Version the assessment alongside the system. Record the model and product version, deployment setting, data period, decision threshold, connected tools, owner, approval date and next review date. If any of those move, the organization should be able to tell whether the earlier assessment still describes the thing in use. Procurement contracts must support that record. A deployer cannot assess what a vendor refuses to explain. Contracts should require documentation, change notices, evaluation access, incident support, logging, correction mechanisms and a practical exit path. The provider's own conformity work is relevant, but it cannot replace the deployer's assessment of how the system behaves in a specific institution with specific people. That difference is the heart of Article 27. The same model can create different harms in different settings. A ranking tool used to suggest optional training is not the same system when it helps determine who receives an investigation, interview, loan, benefit or service. Purpose, authority, time pressure, recourse and the cost of error change the rights impact. So does the surrounding bureaucracy. A person may technically retain final authority while facing an interface that makes disagreement slow and agreement effortless. An appeal may formally exist while requiring a document the affected person cannot obtain. A notice may be legally present while written in language nobody understands. The field test catches those details because it follows the decision, not just the diagram. Europe's draft form can become a valuable common instrument. It can give public agencies and other deployers a disciplined way to identify affected people, expose assumptions, compare evidence and assign responsibility before a high-risk system starts making consequential suggestions. But the finished questionnaire should resist the ancient administrative temptation to confuse completion with care. Make every claim point to evidence. Put affected people in the test. Compare against the real baseline. Name a stop condition. Give each mitigation an owner. Reopen the work when the system changes. Publish lessons from the pilots. The form is the map. The field test is where you find out whether the bridge holds.
01
WHAT ACTUALLY CHANGED
The European Commission AI Office published a draft fundamental-rights impact assessment questionnaire and explanatory notice on November 16, 2026.
The draft is designed for certain deployers of high-risk AI systems that will be subject to Article 27 of the EU AI Act.
The Commission says the assessment must take place before deployment and be updated when relevant elements change or are no longer current.
The draft follows the structure of Article 27 and was developed with input from the EU Agency for Fundamental Rights.
Once finalized, the template is expected to become an automated tool on the AI Act Single Information Platform.
Public feedback on the draft is open until January 15, 2027.
Affected deployers can volunteer by December 4, 2026 to test the questionnaire on one of their own systems.
The AI Office plans to support selected testing participants and hold a closing discussion about the results.
The Commission says Article 27 obligations are scheduled to apply from December 2, 2027.
Participation in the testing exercise is voluntary and does not alter a deployer's legal obligations.
02
WHY THIS MATTERS
A common template can make rights-impact work more consistent across sectors and member states.
A completed questionnaire can still conceal weak evidence, impractical oversight and inaccessible appeals.
The effect of an AI system depends on its setting, authority, users, affected population and cost of error, not only on the model.
Frontline workers and affected communities can identify operational harms that a provider or procurement team may miss.
Evidence attachments make claims about accuracy, oversight, notice, recourse and mitigation inspectable.
A baseline comparison can show whether the AI system improves a flawed existing process or merely changes the shape of the harm.
A named stop condition gives the assessment power to delay, narrow or block a deployment.
Restart triggers keep an assessment current when the model, data, threshold, interface or use changes.
Versioned records connect the approved assessment to the actual system in operation.
Public lessons from pilot assessments can improve the final template before it becomes a compliance routine.
03
WHERE IT COULD HELP
- Attach an evidence artifact to every consequential claim in the assessment.
- Map the full decision path from data collection to notice, review, appeal and correction.
- Include frontline workers, affected people, advocates and accessibility specialists in field testing.
- Pay community participants and provide materials in usable languages and formats.
- Run scenarios involving missing data, duplicate identity, uncertainty, outages, inaccessible channels and time pressure.
- Measure the current process before comparing it with the proposed AI-supported process.
- Break performance down by relevant population, setting and type of harm.
- Define a stop condition that blocks deployment when a critical test fails.
- Give each mitigation a named owner, budget, deadline and verification test.
- Record the model version, data period, threshold, connected tools and interface version.
- Reopen the assessment after material changes, serious incidents or repeated complaints.
- Require vendor change notices, documentation, logs, evaluation access and an exit path in the contract.
- Test whether human reviewers have the authority, information and time to disagree with the system.
- Measure whether notice and appeals work for the people most likely to need them.
- Publish de-identified pilot lessons about confusing questions, missing evidence and failed mitigations.
- Schedule a review date instead of treating approval as permanent.
KEEP A HAND ON THE WHEEL
The questionnaire and explanatory notice are drafts under consultation. They are not the final Article 27 template or a complete statement of how every legal obligation will be interpreted. Article 27 applies to specified deployers and high-risk uses. The release should not be read as a requirement for every AI user or every AI application. The Commission says deployers will need to comply from December 2, 2027 under the current timeline. The implementation schedule and surrounding guidance can still change. Volunteer testing is intended to improve the template. Participation does not prove that a system is safe, lawful or ready to deploy. An automated form can improve access and consistency while also encouraging checkbox behavior. The quality of the assessment depends on evidence, affected-person participation, authority to stop the deployment and follow-up after launch. Watch for the full draft documents, feedback from civil-society and worker groups, the sectors selected for pilot testing, public lessons from those pilots, revisions to the template, connections to data-protection impact assessments and a final tool that makes evidence and update triggers visible rather than burying them in free text.
04
TERMS WORTH KEEPING
SOURCES AND VERIFICATION STATUS
This article was written from the materials below. Product claims and dates were checked against those sources on November 17, 2026.
THE PUBLICATION ENGINE
WANT A SIGNAL OF YOUR OWN?
We build source-grounded publications, private briefings, and editorial systems for organizations with something useful to say.
WORK WITH US