Welcome to the AI Brief from the YPO Technology Network. I'm Stephen Forte. Somewhere in your company right now, someone is typing your most sensitive thinking into an AI and it is being written down. Here is what I will give you in the next few minutes. The court case where a CEO's own chatbot history cost him his company, why the law already treats those logs the way it treats your email, and then the practical part. The retention rules I would set this quarter and the one place I would put genuinely confidential work so that no subpoena can ever reach it. Here's the lens for today. Every conversation your people have with an AI is a business record discoverable in a lawsuit. Usually not privileged and in most companies quietly set to auto-delete right up until the day that deletion becomes illegal. We spent the last episode on a risk you can no longer insure. Today is the twin of that. The record of what your AI did is now evidence a court can compel, and almost nobody has a policy for it. Let me start with the proof because it is the most CEO-specific story I have told this month. This spring, a Delaware court removed a chief executive and handed his company back to the founder he had replaced. The dispute was a $250 million earnout, the extra payment owed to a company's former owners if the business hits its targets after an acquisition. The court found the CEO had deliberately worked to undermine those targets so the payment would not come due. And the evidence that proved it was not an email, not a memo, not a whistleblower. It was his own conversations with ChatGPT, including messages he had deleted. He had used the AI to think through the plan step by step, and the AI wrote it all down, time-stamped, and kept it. The court found bad faith, reinstated the prior CEO, and extended the earn-out by 258 days. Sit with that. The AI did not leak the scheme. The AI was the scheme in writing with a clock on every line. He deleted the messages and they surfaced anyway. That is not a frontier lab problem or an engineering edge case. That is a CEO using the same chatbot your executives use every day, building the case against himself one prompt at a time. Now, why is this the law and not just bad luck? Here is the doctrine in plain English because you do not need to be a lawyer to run this playbook. You just need to know the three moving parts. The first part, an AI chat, is what the courts call electronically stored information in the same legal bucket as an email or a text message. If it is relevant to a lawsuit and your company controls it, the other side can demand it. There is nothing special about the fact that a chatbot generated it. A leading firm put it flatly this year. AI chats are another form of electronically stored information, much like email and text. The second part, and this is the one that bites, the duty to preserve does not start when you are sued. It starts the moment you reasonably anticipate being sued. A demand letter, an incident, a regulator's phone call. From that instant, anything potentially relevant has to stop being deleted And our system deletes it automatically or we cannot search it stops being an explanation and becomes a confession That is exactly the fight OpenAI is in right now A court ordered it to preserve chat GPT logs The news publishers suing it allege OpenAI then deleted billions of them anyway and told the court for two years it could not search its own data until a court-ordered deposition of its own privacy engineer earlier this year allegedly revealed the company had already built an internal database of some 78 million conversations to check its own exposure before the lawsuit was even filed. 17 news organizations are now asking a judge for sanctions. I want to be careful and fair. That motion is pending. The allegations are unproven and OpenAI calls them false. But notice the shape of it because it is not new. The third part is the reason the shape matters. When a court decides a company destroyed evidence it should have kept and did it intentionally, it can do something close to fatal. It can instruct the jury to simply assume the destroyed material was damaging. It is called an adverse inference instruction, and it can decide a case before the facts are even argued. This is not hypothetical. A federal judge did exactly that to Google a couple of years ago over a chat setting. Google had left its internal chat on history off, letting employees quietly not preserve messages even under a legal hold. The judge called it what it was and told the jury to assume the worst. The AI log version of that identical fact pattern is what is on the table today. Same rule, same remedy, new technology, and one more so no executive mistakes the chatbot for a confidant. A federal judge ruled this year that a defendant's chats with a consumer AI tool were not protected by attorney-client privilege. because the AI is not a lawyer, it owes you no duty of confidentiality, and the provider's own terms say the conversation is not private. Whatever your people type into a consumer chatbot, assume opposing counsel can read it out loud in court. Now, let me give the other side its due because this is not a simple story, and you should not act on the panic version of it. The courts are genuinely split. In two other cases this year, judges protected AI conversations, but only where the AI was used under a lawyer's direction on an enterprise tool with a real governance policy around it. That is a meaningful signal, not a loophole. Second, judges are not rubber-stamping phishing expeditions. In the OpenAI case itself, the magistrate refused to force the New York Times to hand over its own internal AI content, calling it disproportionate. And third the privacy objection is real Sweeping millions of ordinary users chats into a lawsuit they have nothing to do with is a genuine cost not a technicality So the honest read is not AI logs are automatically evidence It is AI logs are ordinary evidence and which is worse because ordinary is exactly what gets subpoenaed Here is why this lands on your desk and not your IT departments A survey this year found 43 of employees are pasting confidential or sensitive company data into AI tools their employer never approved. Most of those tools auto-delete, which feels safe and is the opposite of safe, because the deletion is doing nothing but destroying your ability to know what was said while creating no protection at all. And the tools you do sanction quietly keep more than you think. Microsoft's co-pilot, for instance, tucks every prompt and response into a hidden folder that your own e-discovery hold will preserve and make searchable, including in the mailbox of an employee who already left. So, if I were sitting in your seat this quarter, here is the playbook. Four moves. First move. Set a records retention rule for AI logs, but do not default to keeping everything forever and do not default to deleting fast. Both are traps. Keep everything and you have built the most candid, unfiltered archive of your company's thinking and handed it to your future adversary on top of a bigger breach target. Delete on instinct and you are one lawsuit away from that adverse inference instruction. The right rule is neither. It is this. An AI log is not a new kind of record. It is evidence of how an existing record was made. So match its retention to the rule that already governs that record. Concretely, and confirm this with your counsel. If you are in financial services, that is roughly three years for client communications and six for the books under the SEC and FINRA rules. In healthcare, six years for compliance records, and often six to ten for anything that touches the patient record, under HIPAA and state law. In general business, seven years for anything audit-related, one to three for routine HR, and indefinite, the second litigation is reasonably anticipated. Ignore the tidy, keep chat for 30 or 90 days figure you will hear. There is no authority behind it. Anchor to the record, not to a round number someone made up. Second move, build the switch that stops the deletion and turn it on before you need it, not during a subpoena. Every serious platform has it. Microsoft Purview for Copilot, Google Vault for Gemini, Legal Holds and Slack. These take days to weeks to configure and test. The value is not in what you keep by default. It is in being able to freeze the right data, scope to the matter the instant the duty triggers. Third move, and this is the one I most want you to take away because it is the positive answer, not just a warning. There is a right place to do genuinely confidential AI work, legal strategy and acquisition a personnel problem a board matter And it is not a cloud service with a good privacy promise It is a model that runs on hardware you own where the conversation never leaves the building so there is no vendor no server log, and nothing for anyone to subpoena in the first place. This used to require a data center. It no longer does. Today, an executive can run a serious open-weight model, meaning one whose weights you download and run yourself on a single Apple Silicon Mac through a free one-click app called LM Studio. My concrete recommendation, a MacBook Pro or Mac Studio with 32 to 64 gigabytes of memory running OpenAI's GPT-OSS, an open weight model it released under a clean, unrestricted license or Alibaba's QN if you want more horsepower and your board is comfortable with the provenance. Ask it your most sensitive question. Close the window and there is no record anywhere but your own machine. Two honest caveats because I am recommending this for real. One, a model running on your laptop is a notch below the frontier. The gap has narrowed two single digits on the public benchmarks, but it is real. So treat this as your private study, not the workhorse you use for everything. Two, keep it actually local. Do not let anyone wire that model up to the open internet, which is the one way people have turned these local tools into a vulnerability. Run it offline and it is the most private thinking space money can buy. And not much money at that. For work that does not rise to crown jewels but still shouldn't sit on a consumer account, step up to an enterprise tool with a zero retention agreement. For the everyday 90%, a sanctioned enterprise chat tool and the old email rule, don't type it if you wouldn't want it read aloud. Fourth move, and it is free. Tell your people the truth in one sentence. The prompt is not a diary. It is a document. The reason the Delaware CEO is out of a job is that he treated a chatbot like a private strategist instead of a business record. Your people are making the same mistake right now at scale, and no tool fixes a habit. Here's my read on the whole thing. A records policy was never about remembering everything. It is about being able to stand in front of a judge and explain why you kept what you kept and let the rest go. On a schedule you set before anyone sued. The company that saves every AI chat forever is not careful. It has built the perfect case against itself. The company that deletes on instinct is one motion away from losing before it argues. Best practice is the narrow ridge between them, a normal schedule you can instantly provably freeze, plus one genuinely private place that you own for the work that can never leak. Because in the end, the only conversation nobody can subpoena breach or read back to you in court is the one that never left your laptop. That is the YPO Tech Network AI Brief for Thursday, July 16th. I am Stephen Forte. If this was useful, send it to a fellow member, especially the one still pasting board material into a free chatbot. I will be back Friday with more. Until then, stay sharp.
