AI fabricates evidence against Kim Soo-hyun, sparks scandal

According to BBC reporting and South Korean media coverage, an audio file presented during a 2025 controversy about actor Kim Soo-hyun was later concluded to be AI-generated. The Seoul Central District Court approved an arrest warrant on May 26, 2026, for YouTuber Kim Se-ui, who runs the YouTube channel HoverLab, after authorities alleged he circulated manipulated screenshots and an AI audio clip purporting to be the voice of the late actress Kim Sae-ron. Kim Se-ui was subsequently taken into custody. The clip was played at a press conference in May 2025 and depicted the actress's apparent voice describing a relationship with Kim Soo-hyun beginning when she was in middle school. The Korea Times cites a civic-group estimate that 2,026 news articles ran in the eight days between March 10, 2025, when the accusation first surfaced, and March 17, 2025. Kim Se-ui now faces charges of defamation, intimidation, attempted coercion, and distributing footage related to violations of the Act on Special Cases Concerning the Punishment of Sexual Crimes.

The Korea Times' June 22, 2026, analysis frames this as potentially the first of many such cases as AI voice synthesis becomes more accessible. Because the recording sounded authentic, mainstream journalists could justify covering it without independent verification, accelerating harm before verification was possible. Authorities encountered a significant forensic gap: the National Forensic Service reportedly could not formally confirm AI manipulation from the submitted file because the original file was unavailable. Police ultimately concluded fabrication based on circumstantial evidence - the distribution trail, existence of conflicting versions, and statements from those involved. Per the Korea Times, this evidentiary gap is one defendants in similar future cases could exploit.

Per public reporting, investigators concluded the audio was generated by AI, but sources do not specify which tools or models were used. Industry-pattern observations: modern voice-synthesis tools can reproduce speaker-specific timbre and prosody well enough to deceive untrained listeners, and readily available consumer tools plus minimal editing can produce clips that sound authentic. The core technical challenge is that detection methods are an arms race - watermarking and provenance approaches can help when platforms and creators adopt them, while forensic classifiers must continually update to match new synthesis techniques.

Trial proceedings and any forensic reports released publicly will be primary sources for assessing the technical claims. Observers should also watch for changes in platform takedown practices, adoption of provenance standards such as cryptographic provenance or embedded watermarks, and updates from forensic teams about detection performance against the types of audio synthesis implicated in this case. The Korea Times frames this as potentially the first of many similar fabricated-evidence incidents, raising the stakes for practitioners developing detection and incident-response tooling.

This summary draws on reporting from The Korea Times (June 22, 2026, and May 21, 2026), BBC (May 22, 2026), Quartz, and Seoul Economic Daily. The AI-audio conclusion and allegations about manipulated chat screenshots are attributed to police and prosecutors via media filings. No public technical report naming a specific synthesis model was cited in the coverage reviewed.