Enforcement begins, deepfakes tested in court, and how to prove what your AI did
The AI Office has sent its first formal requests under the Act, courts are grappling with deepfakes, and builders are asking how to make agent logs hold up.
- 01
The AI Office starts enforcing transparency rules
On 2 August 2025, the European Commission's AI Office began enforcing the AI Act's requirements for general-purpose AI providers. That includes machine-readable marking of AI-generated content and transparency documentation.
Why it matters for evidence: Providers must show what their systems generate and mark it, which means keeping a record that holds up under scrutiny.
Source: ai.gov.gr
- 02
First formal requests sent to OpenAI, Anthropic and Google
The EU AI Office has sent its first formal information requests under the AI Act. The recipients include OpenAI, Anthropic and Google, marking the start of active supervisory enforcement.
Why it matters for evidence: When the supervisor asks what happened, logs and records that can be shown unchanged make the conversation simpler.
Source: linkedin.com
- 03
Canadian court paper asks how to weigh deepfake evidence
A courthouse library analysis examines how courts should assess video evidence when witnesses can honestly authenticate manipulated footage. The paper highlights the challenge generative AI poses to traditional authentication in legal proceedings.
Why it matters for evidence: Courts apply free assessment. A record of when and how video was captured, sealed at the time, is harder to challenge than memory alone.
Source: courthouselibrary.ca
- 04
AI-generated fake nude images of students sold and shared
Fake nude images of students were generated with AI and distributed without consent, then sold and used for harassment. The incident raises questions about content marking, deepfake disclosure and proof that safeguards against misuse were in place.
Source: OECD AI Incidents Monitor
- 05
A technical guide to tamper-evident audit trails for agents
A step-by-step guide on building tamper-evident logs for AI agents, using structured records, canonicalisation, hashing and chain-linking. It addresses the core question: how do you prove what an autonomous system actually did?
Why it matters for evidence: Logging what an agent did is one thing. Showing later that the log hasn't changed is another, especially when you kept it yourself.
Source: hackernoon.com
- 06
Marine insurance: proving a photo hasn't changed, months later
In marine insurance, a photo of damaged cargo can determine who pays. The challenge comes months later, when someone has to show the image has not been altered since it was taken.
Why it matters for evidence: A sealed record at capture makes the photo easier to rely on in a dispute than a file anyone with access could edit.
Source: linkedin.com