The ICO's AI and Data Protection Guidance: What It Adds

Most of the ICO’s AI guidance is UK GDPR applied carefully, which is useful but not distinctive. Three parts are not: its treatment of statistical accuracy as separate from the accuracy principle, its fairness annexes, and its insistence that you document the trade-off you chose rather than pretend there was not one.

The shape of the guidance

The Information Commissioner’s Office published its Guidance on AI and data protection in July 2020 and substantially updated it in March 2023 to align with the risk-based structure the regulator now uses. It is organised around four questions rather than around the articles: what the data protection implications of AI are and how accountability and governance work; what is required for lawfulness, fairness and transparency; how to assess security and data minimisation; and how to ensure individual rights. The 2023 revision added Annex A on lawfulness and Annex B on fairness, which are the substantive additions. The guidance sits alongside the AI and Data Protection Risk Toolkit, which is the version most teams actually work from because it is structured as risk statements with suggested controls. Both are published on the ICO site.

Generative AI was addressed separately, through a consultation series run across 2024 covering the lawful basis for web-scraped training data, purpose limitation, accuracy, individual rights and controllership across the supply chain, with an outcomes report published at the end of that year. Treat the consultation outputs as the ICO’s current thinking rather than as settled guidance; several of the positions were expressed as provisional.

The fairness chapter and statistical accuracy

The most useful distinction the guidance draws is between the accuracy principle in Article 5(1)(d) — personal data must be accurate and kept up to date — and what it calls statistical accuracy, meaning how often a model’s predictions are correct. These are different obligations with different remedies. A model can be statistically excellent and still process inaccurate personal data; a record can be accurate and the inference drawn from it wrong. The guidance says that an inference recorded as fact engages the accuracy principle, while an inference recorded as a prediction with its confidence does not in the same way — which is a design instruction, not just a classification.

On fairness, Annex B walks through the sources of unfairness the ICO expects you to consider: imbalanced training data, proxies for special category data, differential performance across groups, and feedback loops where the model’s own decisions generate its next training set. It expects you to have chosen a fairness measure knowing that several mutually incompatible definitions exist, and to have recorded why. Where testing for bias requires processing special category data, the guidance points to the substantial public interest conditions in Schedule 1 to the Data Protection Act 2018 — the equality of opportunity or treatment condition being the one usually reached for — with the appropriate policy document that condition requires. Generic GDPR advice does not get you to that paragraph.

Lawful basis across the lifecycle

The guidance insists that training and deployment are separate processing operations that may need separate lawful bases, and that you must identify each. That single instruction resolves a lot of confusion: consent obtained for providing a service does not automatically cover training on the resulting data, and legitimate interests may be available for one and not the other. Where legitimate interests is relied on, the three-part test — purpose, necessity, balancing — must be recorded, which is what a legitimate interests assessment for training is for.

The guidance is also firm that a data protection impact assessment will be required for most AI involving personal data, because such processing typically meets the Article 35 criteria and appears on the ICO’s own list of operations likely to result in high risk — including innovative technology, denial of service, and large-scale profiling. See the triggers for which limb catches which system. Where the residual risk stays high after mitigation, Article 36 prior consultation with the ICO is not optional.

Automated decisions after the 2025 reform

This is the part of the guidance that has moved and where dates matter. Under UK GDPR Article 22 as originally enacted, a decision based solely on automated processing producing legal or similarly significant effects was prohibited unless one of three exceptions applied, with safeguards including a right to obtain human intervention, to express a point of view and to contest the decision.

The Data (Use and Access) Act 2025, which received Royal Assent on 19 June 2025, restructures that. It replaces the general prohibition with a permissive framework for decisions not involving special category data, subject to safeguards that must be in place — broadly, informing the data subject, enabling representations, providing human intervention, and enabling the decision to be contested — while retaining a stricter regime for special category data. The text is on legislation.gov.uk. Commencement is by regulations and is phased, so the operative question for any given date is which provisions have been commenced; check that before relying on either version. The ICO has been updating its guidance to match, and the Article 22 page tracks the detail.

What has not changed is the meaning of “solely”. A human who rubber-stamps a model output does not take the decision out of the regime; the ICO’s position, consistent with the pre-reform guidance, is that the human involvement must be meaningful, exercised by someone with the authority and competence to change the outcome.

Using it as evidence

The guidance is written to be used as an audit trail, and the highest return on reading it is to mirror its structure in your own documentation. Where it says to record a trade-off — accuracy against explainability, data minimisation against bias testing, statistical performance against a fairness constraint — the record of having considered and chosen is the compliance artefact. An organisation that can produce a DPIA, a legitimate interests assessment, a fairness testing record and a documented trade-off has answered most of what an ICO enquiry opens with. One that has excellent controls and no documents has not.