UK Electoral Commission builds AI system to flag election deepfakes to candidates within three hours
The UK Electoral Commission is developing an AI-powered system to alert candidates and parties to significant deepfakes within three hours of first posting. Chief executive Vijay Rangarajan announced the plan at the Conservative Party conference and cited a deepfake of Irish presidential candidate Catherine Connolly during Ireland's 2025 election, which falsely claimed the vote was cancelled and that she was withdrawing. In a pilot during the May 2026 local elections, AI analyzed about three million posts and flagged 147 items. Human reviewers judged ten to be quite misleading and referred them to social media companies or the police. Rangarajan said the posts had little impact, but he expects "an absolutely certain attempt" to use deepfakes against the UK.
Detalhes técnicos
- The system scans social media in real time, flags suspicious audio and video for human review, and alerts the relevant candidates or parties without first attributing the content to a creator.
- Scope is limited to false statements that could undermine election administration, not disputes over political claims.
- Full Fact found AI-built campaign videos from a Glasgow candidate that fell outside the pilot because they were labelled as illustrative AI scenes, with the label at the end of the caption.
- Rangarajan said the Commission believes it detected algorithmic bias on at least one unnamed social network, and that UK law offers no clear remedy.
- The Commission reports 30,000 to 40,000 cyberattacks a month, sometimes 10,000 a day, including botnet attacks and phishing.
Remediação
- The Commission is seeking legally enforceable takedown powers and updates to election laws dating from 1983 and 2000, and possibly new offences.
- Rangarajan wants stronger, more independent security capabilities, and thinks another body could handle major social media monitoring.
- The Commission said it would publish the pilot results.
Linha do tempo
-
UK Electoral Commission builds AI system to flag election deepfakes to candidates within three hours
ler a análise desta fonte ocultar
The UK Electoral Commission is developing an AI-powered system to alert candidates and parties to significant deepfakes within three hours of first posting. Chief executive Vijay Rangarajan announced the plan at the Conservative Party conference and cited a deepfake of Irish presidential candidate Catherine Connolly during Ireland's 2025 election, which falsely claimed the vote was cancelled and that she was withdrawing. In a pilot during the May 2026 local elections, AI analyzed about three million posts and flagged 147 items. Human reviewers judged ten to be quite misleading and referred them to social media companies or the police. Rangarajan said the posts had little impact, but he expects "an absolutely certain attempt" to use deepfakes against the UK.
Technical Details
- The system scans social media in real time, flags suspicious audio and video for human review, and alerts the relevant candidates or parties without first attributing the content to a creator.
- Scope is limited to false statements that could undermine election administration, not disputes over political claims.
- Full Fact found AI-built campaign videos from a Glasgow candidate that fell outside the pilot because they were labelled as illustrative AI scenes, with the label at the end of the caption.
- Rangarajan said the Commission believes it detected algorithmic bias on at least one unnamed social network, and that UK law offers no clear remedy.
- The Commission reports 30,000 to 40,000 cyberattacks a month, sometimes 10,000 a day, including botnet attacks and phishing.
Remediation
- The Commission is seeking legally enforceable takedown powers and updates to election laws dating from 1983 and 2000, and possibly new offences.
- Rangarajan wants stronger, more independent security capabilities, and thinks another body could handle major social media monitoring.
- The Commission said it would publish the pilot results.