Telling Affected People When a Decision Involved High Risk AI
Article 27 requires public authorities to inform affected people when a decision concerning that person was based on high-risk AI. This transparency obligation recognises that individuals have a right to know that an AI system influenced a decision about them. If a local authority uses an AI system to make a social housing allocation and that system influences the decision about your application, you have a right to know that fact.
The notification must be clear and occur when the decision is communicated to the affected person. If a social housing authority tells an applicant the outcome of their application, that communication must also explain whether an AI system was involved and, if so, what role it played. The notification cannot be opaque or use technical jargon that would confuse the average person. It must be intelligible.
Consider a public authority using an AI system to assess eligibility for benefits. The authority sends a letter to an applicant stating the eligibility decision. The letter must include information that the decision was supported by an AI system, what information the system considered and whether the applicant can contest the decision based on the AI system’s assessment. The letter should be written in plain language and in the applicant’s language if required by law.
What the notification should include
The notification should explain that AI was used, what the AI system did (for example, “assessed your income eligibility against policy thresholds”), whether the AI recommendation was decisive or just one input to a human decision, and how the affected person can contest the decision. The notification might also explain relevant facts about the system, such as “this system was reviewed to ensure fair treatment of all applicants” or “human staff reviewed the AI recommendation before the final decision was made”.
The notification should not overwhelm the affected person with technical detail. The purpose is to give the person sufficient information to understand that AI was used and to know their rights, not to educate them about how machine learning works. A brief, clear statement is sufficient.
Practical challenges
Deployers often struggle with notification timing and method. If a decision is made entirely by an AI system and a human simply formalises it, does the human’s involvement change the notification requirement? The answer is no. If the AI recommendation is decisive (meaning the human would rarely override it), the decision is based on the AI system. If the human routinely reviews and sometimes changes the AI recommendation, human judgment has a real role and the notification should reflect that.
Some decisions are made in stages. A first-stage AI system recommends recipients of a benefit. A human reviewer considers appeals. Must the affected person be told about the AI’s role even if they do not appeal? The answer is yes. Notification should occur at the point the decision is communicated to the affected person, regardless of whether they later contest it.
Record-keeping
Deployers must keep records that they notified affected people about AI involvement. This is part of the broader logging requirement in Article 26. If a public authority is later accused of unfair decision-making and cannot produce evidence that it informed people of AI involvement, this is a breach of Article 27. Records should show the date of notification, to whom it was sent, the method and the content.
Providing recourse and contestation
Notification should also tell people how to contest or appeal a decision the AI system influenced. A public authority might provide a phone number, email address or online portal where people can request human review. The appeal process should be straightforward and free to use. People should not need to hire a lawyer or pay fees to contest a decision. Organisations should track appeals and analyse whether certain groups appeal more than others, which might indicate systemic bias.
Response times to appeals matter. If a person requests appeal after an AI system influenced their housing decision, they should receive a response within weeks, not months. Long delays are unfair to affected people and undermine their right to contest.
Accessibility and inclusion
Notification requirements must be applied inclusively. If a person is deaf, notification by letter might not be sufficient; the authority might need to provide the information by video with captions or sign language. If a person has intellectual disabilities, the notification must be in a format they can understand. Accessibility is not an optional add-on; it is a fundamental requirement of meaningful notification.
Organisations should provide notification in multiple languages if they serve multilingual populations. A local authority serving a neighbourhood with significant immigrant communities should provide notification in languages commonly spoken in that community. This is both legally required and ethically sound.