UK OSA Minimum Standards of Accuracy
Minimum Standards of Accuracy in the Detection of Terrorism and CSEA Content (Online Safety Act 2023, section 125)
Statutory minimum standards of accuracy that terrorism-content and CSEA-content detection technology must meet before Ofcom can accredit it and require its use through a technology notice under Chapter 5 of Part 7 of the Online Safety Act 2023. Accreditation is an audit-based assessment across four principles (technical performance, fairness, robustness, maintainability), each scored from evidence submitted by the technology developer.
Jurisdiction
United Kingdom
Enacted
Sep 4, 2026
Effective
Sep 4, 2026
Enforcement
Ofcom (accreditation and technology notices); Secretary of State for Digital, Culture, Media and Sport approves and publishes the standards
Approved and published by the Secretary of State for Digital, Culture, Media and Sport on 4 September 2026, following advice from Ofcom under section 125(13) of the Online Safety Act 2023. In force from 4 September 2026.
GOV.UK: Minimum standards of accuracy in the detection of terrorism and CSEA content (published 4 September 2026)Why It Matters
Technology notices are the mechanism in the Online Safety Act for compelling proactive scanning, including of privately communicated CSEA content. These standards define the accuracy, fairness and robustness a classifier must demonstrate before it can be mandated, and are the first UK statutory benchmark for evaluating content-detection models.
Recent Developments
Published and brought into force on 4 September 2026 as a DCMS notice. This is the first set of minimum standards approved under section 125 of the Act, which is the precondition for Ofcom to accredit technology and issue technology notices for terrorism and CSEA content. Accreditation does not by itself mean Ofcom will consider a notice requiring that technology necessary and proportionate. Ofcom sets the evidence questions and grades submissions.
At a Glance
Applies to
Harms addressed
Requires
Who Must Comply
- Developers of terrorism-content or CSEA-content detection technology seeking accreditation from Ofcom
- Providers of regulated user-to-user services and regulated search services that receive a technology notice from Ofcom requiring use of accredited technology
Obligations fall on:
Safety Provisions
- Technology must score at least 80 out of 100 on Objective 1.1 (performance metrics: evaluated against the false positive rate and other appropriate metrics for its intended deployment environment) and at least 40 out of 100 on every other objective
- Technology must score at least 60 out of 100 on each of the four principles and at least 60 out of 100 overall
- Dataset quality: development, training and test data must be comprehensive, representative of the harm being detected and diverse enough to test generalisation to unseen data
- Reproducible performance across the environments for which the technology is designed
- Secondary validation: outputs evaluated against expert human judgement where possible, particularly in complex or nuanced cases
- Fairness: identification and mitigation of unfair bias across different groups of people
- Robustness against adversarial attacks such as evasion attacks, injection attacks and input perturbations
- Maintainability: ability to modify or update the technology to preserve accuracy against new threats
- Evidence is graded 5 points (robust and comprehensive), 1 point (notable limitations) or 0 points (missing, misleading or inaccurate) per question set by Ofcom
Compliance & Enforcement
Key Dates
Sep 4, 2026
Minimum Standards come into force; Ofcom (or a person it appoints) may accredit detection technology against them for use in technology notices
Penalties
None attach to the standards themselves. Failure to comply with a technology notice is enforceable by Ofcom under the Online Safety Act 2023, with fines of up to 18 million pounds or 10 percent of qualifying worldwide revenue.
View on map
United Kingdom
Focus Areas
Cite This
APA
United Kingdom. (2026). Minimum Standards of Accuracy in the Detection of Terrorism and CSEA Content (Online Safety Act 2023, section 125).
Related Regulations
Ofcom Children's Codes
Ofcom codes requiring user-to-user services and search services to protect children from harmful content including suicide, self-harm, and eating disorder content. Explicitly covers AI chatbots that enable content sharing between users. Requires detection technology, content moderation, and recommender system controls.
UK OSA
One of the most comprehensive platform content moderation regimes globally. Creates specific duties around suicide, self-harm, and eating disorder content for children with 'highly effective' age assurance requirements.
KCSIE
Statutory safeguarding guidance issued by the Department for Education for schools and colleges in England, setting out duties to prevent and respond to harm to children, including abuse, neglect, and online safety risks. Reissued annually; requires a whole-school approach to online safety organised around a four-part risk framework (content, contact, conduct, commerce), together with a designated safeguarding lead, staff training, and a child protection policy.
UK AI Chatbot OSA Extension
Section 248 of the Crime and Policing Act 2026 confers a delegated power to bring standalone and generative AI chatbot providers (ChatGPT, Grok, Gemini, etc.) within scope of Online Safety Act illegal-content duties - the mechanism to close the loophole where AI-only chatbots fell outside the OSA. The substantive duties on AI chatbots await secondary legislation.
UK OSA CSEA Reporting Regs 2026
Statutory instrument implementing section 69 of the UK Online Safety Act 2023, requiring regulated user-to-user service providers (UK and non-UK) to report child sexual exploitation and abuse (CSEA) content to the National Crime Agency under tiered priority-based timeframes.
AR HB 1071
Amends Arkansas publicity rights law to explicitly include AI-generated reproductions of voice and likeness. Covers simulated voices and 3D generation.
Last updated September 6, 2026. Verify against primary sources before relying on this information.