AI-Written Subject Access Requests: What the ICO Expects Organisations to Do

Privacy teams across the UK have noticed a change in their inboxes: subject access requests that run to several pages, cite laws from other countries, quote cases that do not exist and demand every scrap of metadata the organisation has ever generated. Many of them are written with the help of AI chatbots. On November 11, 2026, the Information Commissioner’s Office published short guidance on how to handle these “AI-assisted SARs.” The core message is simple, and some organisations will not like it: a request drafted by AI is still a valid request, and the usual rules apply in full.

What the ICO said

The ICO’s guidance on SARs generated by artificial intelligence defines AI-assisted SARs as requests that appear to have been generated using AI or drafted with its help. It acknowledges the problem plainly. Such requests can be lengthy, repetitive, broad or unclear; can focus on information of limited value such as metadata; and can include misleading information, irrelevant legal references, laws from other jurisdictions or a mix of different rights requests and complaints.

The ICO’s answer to “does using AI make a SAR invalid?” is a flat no. It compares AI to a spell checker or grammar tool: another form of help that does not mean the person isn’t genuinely seeking their information. What matters is whether it is clear the person wants their personal data. They do not need to use specific wording or cite the correct law.

The guidance uses the ICO’s must, should and could structure, which separates legal requirements from good practice. It covers requests under Article 15 of the UK GDPR, though much of it is also relevant to law enforcement requests under Part 3 of the Data Protection Act 2018.

Correcting a common reading

Some early summaries suggested that the ICO tells organisations to seek clarification whenever a request contains hallucinated content, and to extend the deadline when someone sends several AI-generated requests at once. The guidance is more careful than that.

• Inaccuracies do not invalidate a request. A SAR that misstates rights or cites non-existent case law is still valid if the person clearly wants their data. You do not need to spend time deciphering legal references that are wrong or irrelevant.
• Clarification is for unclear requests. You may ask for clarification if you do not understand what the person wants, which right they are using, or if the request covers a large volume of information likely to fall outside a reasonable search. You cannot force them to narrow it.
• Complexity is about substance, not style. A request is not complex simply because it is long, convoluted or full of legal references. The deadline can be extended by up to two further months only where there are genuine difficulties, for example hard-to-retrieve data, extensive exemption analysis, third-party information, or where the person makes multiple rights requests at the same time.
• Manifestly unfounded or excessive is not automatic. Using AI does not, by itself, make a request excessive. A clearly malicious, abusive or disruptive request may qualify, whether or not AI was used, and you must be able to justify your decision.

Where the Data (Use and Access) Act fits

The guidance leans heavily on two changes made by the Data (Use and Access) Act 2025, which is now moving from law to live obligation.

First, section 78 of the Act inserted a new Article 15(1A) into the UK GDPR. A requester is entitled only to the information the controller can provide “based on a reasonable and proportionate search.” The section came into force at Royal Assent on June 19, 2025, and is treated as having applied from January 1, 2024. This is the organisation’s main defence against sprawling AI-drafted demands. The ICO stresses that it is the controller, not the requester, who decides the parameters of the search, even when an AI-generated letter lists specific databases, backup tapes, platforms or CCTV.

Second, the “stop the clock” rule. When you reasonably need clarification and ask for it promptly, the one-month time limit pauses and resumes once the clarification arrives. For AI-assisted requests that are genuinely unclear, this is the fairest way to buy time without risking a breach of the deadline.

Courts have been moving in the same direction. A recent High Court ruling confirmed that controllers owe the data, not every document that contains it, which matters when an AI-drafted request demands copies of entire files.

The metadata problem

AI tools seem to love asking for metadata. The ICO’s position is balanced. A request for “all my personal information including metadata” does not automatically require you to search every category of system-generated audit data. But if metadata is reasonably likely to contain relevant personal information, or the person asks for something specific, such as records of when their file was accessed, you should consider searching for it and providing it.

The guidance gives examples of metadata that can identify a person: access timestamps, records that reveal working patterns, call logs showing who someone called and for how long, and GPS tags in photos. If metadata identifies the requester, it is within scope. Because people use the word “metadata” to mean different things, asking what they mean is legitimate, particularly where the request looks machine-written.

Fraud and false details

The guidance also flags a security angle. AI can produce convincing fake identity documents and letters of authority. If a request arrives through a third party with an AI-generated authority letter, the ICO suggests contacting the data subject directly. Correspondence addresses in AI-drafted requests may be wrong, so check them against your records before sending anything. The responsibility for disclosing data to the wrong person stays with you.

A practical workflow for AI-assisted SARs

Many organisations already deal with AI-written correspondence. Britain’s Financial Ombudsman found that AI-written complaints often contain hallucinations and turned to phone calls to get to the real issue. A similar workflow fits SARs.

1. Log it on receipt. Start the clock on the day it arrives, unless you need ID, proof of authority or a fee, as with any request. Our explainer on response times covers the calculation.
2. Find the core ask. Strip out the legal boilerplate and identify what personal data the person wants. Ignore wrong citations rather than rebutting them.
3. Verify identity and contact details. Compare addresses with your records and treat AI-generated authority letters with care.
4. Talk to the requester. An informal call or email often resolves scope faster than formal clarification. If you do formally ask for clarification, do it promptly and note the pause.
5. Set and document the search. Record whether you sought the requester’s views, what you included or excluded, and why. The ICO says this record supports accountability.
6. Decide on complexity or refusal carefully. Only extend or refuse on substantive grounds, and tell the person why within the first month.
7. Apply exemptions normally. There are no special exemptions for AI-assisted SARs; the usual exemptions apply.
8. Explain the law where helpful. The ICO says you can tell someone that their request appears to include misunderstandings, possibly caused by AI, and explain what the law does and does not require.

The ICO also recommends making it easy to request data in the first place: a clear SAR page, a contact point, optional example prompts or templates, and a reminder to review AI output before sending. Staff need training to recognise SARs and to not be intimidated by long, aggressive-sounding letters, which makes DSAR training worth refreshing.

Frequently asked questions

Can we refuse a request because it was obviously written by ChatGPT?

No. The ICO is explicit that AI use does not make a request invalid, complex or excessive on its own.

Do we need to respond to the parts asking about our search methods?

Not under the UK GDPR, since they are not the person’s personal data, but it may help to explain which parts fall outside scope. Public authorities may need to consider freedom of information rules for those parts.

Someone sent five AI-drafted requests in a week. Is that excessive?

It might be, depending on overlap, intent and burden. Assess the circumstances as you would for any repeated requests, and document your reasoning.

Should we warn people about using AI tools?

The ICO says you could remind people not to put sensitive personal information into AI tools, since those systems may retain or reuse it.

The author changed, the right did not

AI makes it easier for people to exercise their right of access and harder for organisations to read what they are asking for. The ICO has chosen not to let the second problem undermine the first. Teams that respond calmly, search proportionately and write down their reasoning will handle machine-drafted requests just as well as any other.