AI chats may be discoverable
U.S. lawyers are warning clients that chatbot conversations could be used in court after a recent ruling raised the prospect that AI chats may not be confidential. (reuters.com). At the same time OpenAI is facing investor scrutiny over its $852 billion valuation as it pivots toward enterprise customers, and has begun a limited release of a GPT‑5.4 'Cyber' model to trusted partners — moves that show firms are both gating capabilities and confronting legal uncertainty. ( )
Lawyers in the United States are warning clients that what they type into ChatGPT or Claude may be obtainable in court. (reuters.com, usnews.com) The warning follows a February 17, 2026 ruling by Judge Jed S. Rakoff of the Southern District of New York in *United States v. Heppner*, where a defendant’s exchanges with Anthropic’s Claude were found not protected by attorney-client privilege or the work-product doctrine. (harvardlawreview.org, hsfkramer.com) Reuters reported on April 15 that some law firms have started putting the warning directly into client agreements, telling people that sharing legal advice with a chatbot could waive confidentiality protections. (reuters.com, usnews.com) Attorney-client privilege is the rule that usually keeps private conversations between a lawyer and client out of court. The Heppner ruling treated a public chatbot more like a third party than a member of the legal team. (harvardlawreview.org, crowell.com) That legal uncertainty is colliding with a business shift inside OpenAI. The Financial Times reported, and Reuters matched, that some investors are questioning OpenAI’s $852 billion valuation as the company pushes harder into enterprise products and code tools. (reuters.com, usnews.com) OpenAI said the criticism does not reflect investor demand. In a statement carried by Reuters on April 14, a company spokesperson said its $122 billion fundraise was oversubscribed and backed by a broad group of global investors. (reuters.com, finance.yahoo.com) At the same time, OpenAI is limiting access to some of its most sensitive new systems instead of releasing them broadly. On April 14, it began rolling out GPT-5.4-Cyber, a version of GPT-5.4 tuned for defensive cybersecurity, to selected participants in its Trusted Access for Cyber program. (bloomberg.com, openai.com) OpenAI said on its own site that it launched Trusted Access for Cyber in February as an identity-based program for vetted defenders, and said this week it is expanding that program to thousands of individual defenders and hundreds of teams. (openai.com, openai.com) The split is becoming sharper: companies want users to rely on chatbots for work, but courts are starting to test whether those conversations are private, and model makers are fencing off higher-risk tools behind vetting programs. (reuters.com, openai.com, reuters.com) For users, the immediate rule is simpler than the policy debate: if a fact would be damaging in a lawsuit, lawyers are increasingly saying not to hand it to a chatbot first. (reuters.com, natlawreview.com)
