There’s a growing assumption that anything you say to an AI platform is private. But there is also a growing body of law that calls into question that assumption when it comes to litigation and judicial proceedings. At the very least, the law is unsettled when it comes to discovery of parties’ chats.
But one thing is clear. When it comes to search warrants served on AI providers, there is little protection from the government’s ability to access it under the Stored Communications Act (SCA).
The Kim Decision
A recent decision by Judge Lorna Schofield from the Southern District of New York underscores that the law is indeed pretty clear. In the case of United States v. Kim, Kim sought to quash a search warrant served on OpenAI under the SCA. The warrant sought certain content in connection with a securities fraud criminal proceeding. Citing a Supreme Court decision and other lower court decisions, Judge Schofield ruled that Kim had no right to seek to quash the subpoena. His only remedy was to try to suppress use of the evidence later on.
That makes some sense when you are talking about tangible things. If the police obtain a search warrant of your residence, you don’t usually get advance notice so you can argue against it. Your protection (if you can call it that) is that to obtain the warrant, the government must show probable cause to a judge or magistrate. If it turns out the warrant was not supported by probable cause, or the material obtained is irrelevant to the issue or otherwise protected, then you can seek suppression of whatever is found.
Don’t Count on the Provider
But what about the holder of digital evidence, in this case OpenAI, which was not a party to the proceeding? Can’t it step in to protect you? Probably not. The SCA was enacted in 1986 and is a mechanism for the government to compel providers to disclose communications. It governs disclosure of things like emails, cloud-stored documents, chat histories, and the like. The provider must thus produce things like communications, account records, stored files, and meta data. The Act requires a warrant but the ability of the provider to object is pretty limited. It can argue the information is voluminous and therefore unduly burdensome to produce, but that’s about it. Judge Schofield noted those circumstances were not relevant in the Kim case.
Bottom line: AI conversation histories with large language models are going to be treated like emails or cloud storage under the SCA. That’s a twofold problem.
So, What’s the Problem?
First, I know from experience that once someone obtains documents and data from a party or third-party provider, it’s hard to ensure that they will be kept out of the proceeding. Many times, that material is arguably relevant or arguably protected by attorney client or work product privilege. If the material is not disclosed, the argument about its producibility is in a vacuum.
But once the secrets are out, it’s much harder to keep it from being used. And if it’s a bench trial, the fact finder will see it in any event. Plus, it always gives me pause that the government has largely unfettered access to documents and data.
And as far as the need to show probable cause to get the warrant in the first place: those are routinely provided based only on what the government puts forward to the judge or magistrate without anyone else present.
Which leads to the bigger problem and that is one of expectations. I fear that litigants and, for that matter, the general public, have been lulled into thinking what they put in a chatbot is private and won’t see the light of day. Or they think even if it’s produced, it won’t come into evidence. (An assumption that lawyers sometimes make as well.)
The result is they are not careful what they put in and what they say to the chatbot. They muse about their fears or suspicions. They say things they would never say publicly. People without legal training (and even some with legal training) don’t think about privilege issues and that what they say may be used against them later.
There was a similar problem in the early days of email. People would say things in emails they would never say in a letter. Which is why I would often give training sessions to clients to sensitize them to the fact that what they say in emails was likely discoverable. And we have seen the disastrous impact of problematic email language over and over again by someone who didn’t think before they hit send.
It’s even worse with AI and large language models. The providers have gone out their way to design systems that make you think you are talking to a human. A human who cares about you and is your friend. A human who would never disclose what you say in private. But that’s not the case. At best, the law with respect to whether AI conversations are covered by the work product and attorney client privileges is not settled. The conversations may be privileged. Then again, depending on the judge or jurisdiction, they may not be. And as the Kim case shows, the government will always have an easy time getting to your AI conversations.
Our Responsibility
So, where does that leave us? As I have previously discussed, it’s all about education. Those of us in the legal profession have a responsibility to sound the alarm about what our clients and the public as a whole put in chatbots. To let people know that what is said should not be assumed to be private. And that the providers often can’t and won’t be there to protect us. Without that education, we are going to see a tsunami of prejudicial AI conversations admitted into evidence that will make what happened in the early days of email look like a ripple in a pond.
Stephen Embry is a lawyer, speaker, blogger, and writer. He publishes TechLaw Crossroads, a blog devoted to the examination of the tension between technology, the law, and the practice of law.