Chatbots, recruitment systems and watermarks: proving what AI did
This morning, four stories on systems that made promises or decisions, and one on how to leave a trail anyone can follow.
- 01
OpenAI on watermarking AI text
OpenAI explains its approach to the EU's machine-readable watermarking rule for generated text. The company says watermarks can't measure human contribution and absence of a watermark doesn't prove human authorship.
Why it matters for evidence: Article 50(2) requires the provider to mark AI output, not the user. The limitation matters: the mark says what made it, not who wrote it.
Source: openai.comOur explainer →
- 02
US court ruling on AI hiring
A federal court has ruled in a lawsuit against Workday after its AI recruitment system allegedly rejected an applicant based on race, age and disability. The case shows the risk when organisations can't prove what their AI did and why.
Why it matters for evidence: High-risk AI under Annex III includes recruitment, requiring automatic logging (Article 12) and the ability to explain the system's role in a decision (Article 86).
Source: linkedin.com
- 03
Chatbot medical advice causes poisoning
Researchers report a 60-year-old hospitalised due to poisoning after following AI chatbot medical advice. The incident raises whether organisations can prove what their AI recommended and whether it met medical-device safety requirements.
Source: OECD AI Incidents Monitor
- 04
Law firm recruiting wrongful death cases
A US law firm is recruiting families for wrongful death lawsuits against OpenAI, alleging ChatGPT encouraged suicidal thoughts. The cases turn on whether AI providers can prove what their systems actually said to vulnerable users.
Why it matters for evidence: Under the new product liability rules, software becomes a product. Courts can order evidence, and proving what a conversational system said matters.
Source: lanierlawfirm.com
- 05
Vercel guide on auditing AI agents
Vercel publishes a technical guide on building audit trails for AI agents, connecting identity, tool calls, approvals and results. The patterns let teams reconstruct what an autonomous system did and investigate errors.
Why it matters for evidence: High-risk AI systems require automatic logging (Article 12). For agents, that means recording what they did, on whose instruction, and with what result.
Source: vercel.com
- 06
EU deepfake labelling Code of Practice
The European Commission has published a Code of Practice on transparency, including guidance on labelling deepfakes and AI-generated text, and a set of EU icons for marking AI content.
Source: digital-strategy.ec.europa.eu