AI AND ATTORNEY-CLIENT PRIVILEGE
The Questions Everyone Missed
Everyone is talking about AI and attorney-client privilege right now. Why?
In the U.S., one of the sparks was USA vs. Heppner.
In that case, the Southern District of New York held that materials generated through a free public AI tool were not protected by attorney-client privilege or work product doctrine in the way the submitting party expected. The reasoning turned on a simple but uncomfortable point: if information is disclosed to a third-party system without clear confidentiality safeguards, privilege can evaporate. Also, that attorney-client privilege requires a human attorney-client relationship and interactions with AI software do not satisfy that requirement.
That’s the moment the conversation shifted from “interesting ethics issue” to “this could lose you protection in litigation.”
But what is more interesting than the panic? Most commentary since then has focused on:
– Don’t paste client data into public AI (editor´s note: why was this not obvious in the first place dear colleagues?)
– Check your vendor contracts
– Implement internal policies
All sensible. All necessary.
At the same time, in Europe, the long shadow of Akzo Nobel Chemicals Ltd v Commission still hangs over everything. On September 14, 2010, the European Court of Justice issued a judgment in the case (C-550/07P), refusing to modify or overturn prior precedent that communications with in-house lawyers are not accorded legal professional privilege under European law in the context of an investigation of violations of competition law by the European Commission. So when AI enters the picture, European lawyers aren’t just asking “is this safe?” They’re asking, “Are we about to narrow privilege even further?”
So yes, the alarm bells make sense.
What if privilege doctrine itself has to evolve?
We’ve adapted privilege before, to paralegals, translators, forensic accountants, cloud storage providers. At some point, AI may stop looking like an “external third party” and start looking like infrastructure. If we refuse to even entertain that possibility, we’re freezing doctrine in a pre-AI world.
Another blind spot: clients.
While lawyers debate policies, CEOs and board members are already pasting legal memos into AI tools to “make them clearer.” Is that a waiver? Implied waiver? Negligence? Most commentary barely touches it.
Then there’s inequality.
If preserving privilege in the AI era requires private, ring-fenced enterprise models, how can smaller law practices afford that? Are we quietly building a system where only large firms and large corporations can safely use modern tools? What consequences will this have for the legal market? For clients’ ability to afford legal services? Ultimately, for the principle of providing equal and accessible legal representation?
And we can’t even stop there. There’s also the cross-border chaos.
If AI processing occurs across jurisdictions, what does that do to privilege claims in multinational investigations? Especially when EU doctrine already takes a stricter structural view?
Right now, most writing is defensive. Risk alerts. Vendor checklists. Compliance hygiene.
But structural thinking is missing:
- Should certain AI systems qualify as agents of counsel?
- If a company runs a fully internal model, is that disclosure at all?
- Should we really end up with a privilege regime that only firms with private AI infrastructure can safely navigate?
- Are we focusing on sanctions risk while ignoring long-term doctrinal consequences?
Privilege is what makes candid legal advice possible. AI is not just challenging lawyers’ caution but also testing whether our doctrine is flexible enough to survive technological inevitability.
However, in our opinion, these technological developments are only a “challenge” if we choose to face them unprepared. Let’s not make that mistake.
If you would like to know more about the world of startups, or have any questions regarding starting one, do not hesitate to contact us, or book a consultation with one of our colleagues by clicking here.
