Summary

  • Ofcom's original study tested eight generative-AI search APIs and three traditional search APIs. It assessed generated answer text and retrieved URLs with separate methods and denominators.
  • For queries actively seeking graphic violence, 29.76% of valid API responses retrieved at least one domain in Ofcom's highest-risk reference-list tier. That is a retrieval result, not a count of what a child saw on a consumer product.

A search result has a front desk and a filing room. The front desk supplies an answer. Behind it, a retrieval system gathers links. Ofcom's 22 October research paper inspected both. That simple separation is the story: a safety score for the answer cannot stand in for a record of the links that were fetched.

The regulator sent sets of 100 queries for each kind of harmful content through eight generative-AI search APIs and three traditional search APIs. Half of each set actively sought the harmful material; half used relevant terms without seeking it. Tests ran in three rounds on 3–5 August 2026. Ofcom anonymised the providers. It used three model judges to label generated answers, requiring unanimous agreement before calling an answer harmful. For the combined seeking and non-seeking questions, those judges marked 3.78% of GenAI responses about violent content and 3.07% about body stigma as harmful.

They could not agree on another 16.67% and 10.21%, respectively. A tidy single percentage would have concealed that undecided pile.

The other ledger tracks URLs, not prose. For both GenAI and traditional search APIs, Ofcom matched retrieved links against its reference lists of domains associated with pornography or graphic violence. Among responses to queries actively seeking graphic violence, 29.76% retrieved at least one domain in the highest-risk tier. The figure for questions on the same subject that did not seek harmful material was 7.27%. For pornography, the corresponding figures were 20.28% and 0.59%. Ofcom assigned each response the highest-risk domain it matched; these are response percentages, not percentages of all links or of all websites.

The gap between the ledgers is operational. A GenAI service may retrieve a URL without using it in its final answer, citing it or showing it to a user. Ofcom measured the retrieval trace; it did not observe children's consumer-facing searches. Its section 63 parent report uses this experiment as supporting evidence, while the statutory duty concerns children's encounters with harmful content. An API measurement and a child encounter belong in separate columns. The 12-year-old age cue could be supplied only to some APIs, and the test could not verify that every API treated the query as coming from the UK.

Ofcom's source list also has a clock. A domain's assigned risk tier may have been assessed before the August calls, so the study identifies matches to its reference list at that time rather than inspecting every site's safeguards on the day of each query. That record is still useful. It tells an auditor where to look next: at the link actually returned to a product, the date of the site's safeguard check, and the interface that a child could use. The answer drawer and the link drawer both need labels; neither should be quietly substituted for the other.

Official sources