Good morning. The core of AI governance is “when an AI system does something, who is responsible?”. When an AI generates something false, harmful or commercially significant, where does liability ultimately land?
So far, the answer remains the lawyer's favourite refrain: “it depends”.
This week, we examine three different attempts to answer it. A Munich court attributing liability to Google for AI-generated summaries despite disclaimers. An AI-powered law firm winning a real trial in London, with automation handling much of the process but regulated professionals remaining essential. And new survey data suggesting most professionals believe they should remain accountable for AI-assisted work, even as a third admit to using shadow AI.
The technology may be moving quickly, but responsibility continues to follow people, institutions and incentives. Exactly where those boundaries sit is becoming harder to predict. 🎯
— Philip
P.S. Over the last two months, Alexis Goldfinch and I met with the founders and CEOs of 17 businesses to undrstand how they are adopting AI. We uncovered consistent themes around client pressures, the shift from deployer to provider status, and anxieties around defending value. We've compiled the findings into a short report: AI Governance 2026: The AI Reality Gap. Reply “report” if you'd like a copy.
BRIEFING ROOM
Owning the answer: German court holds Google liable for false statements in AI overviews
Google is appealing a recent Munich court ordering it to stop its “Overview with AI” from repeating false claims about two German publishers. Asked about the publishers, Google’s AI summary had opened with a flat “Yes” - yes, the publisher is known for dubious business practices and is “often perceived as a scam” - and then laid out a tidy structure of “characteristics of the alleged fraud scheme,” complete with subscription traps and debt-collection threats. The connections it drew appeared in none of the sources it linked: the AI had, on the court’s account, blended the publishers with genuinely dubious companies and produced something new.
Earlier German law gave search engines a wide shield: an engine merely makes third-party content findable, so it is liable only once it is told of an obvious infringement. The Munich court viewed the AI overview differently, seeing it as representing Google’s “own statement made by its own AI offered to users”. The court said: “since the defendant introduced the artificial intelligence itself and offers it to users, it must also be held accountable for its results, as it alone has influence over the AI’s offering and the algorithms it uses”.
On whether Google should be able to rely on disclaimers that AI overviews may make mistakes, the court said the output was “self-contained and understandable... so that users regularly have no reason to additionally check the displayed answer”
Google’s view is that the case “focuses on specific and narrow errors, not the foundational way AI Overviews display web content”, noting that the “overwhelming majority” of AI Overviews are accurate.
Why it matters
Legal teams will be reluctant to draw generalised conclusions as the decision is interim relief and is not yet final. The departure from earlier search-engine case law might be reversed. But this case is less “Google lost” than evidence that AI outputs are being challenged, visibly and with success. That principle is far-reaching.
🛑 The disclaimer isn’t the shield you think. “AI-generated, may contain errors” did not save Google, because the court treated the output as a self-contained statement a reader has no reason to double-check. Legal teams may flag to product developers that boilerplate warnings pushing risks onto users may not always be watertight.
🪞 The more your AI rewrites, the more it speaks for you. The court introduced a novel dividing line between passing content through (a link) and generating “independent, new, and substantive statements”. A tool that quotes and cites sits on the safer side; one that summarises, synthesises and asserts, sitting closer to “your words” might scale risk depending on how the tool rephrases.
🗣️ Don’t assume free-speech defences cover the machine. The court suggested the AI's output may merit less protection because it reflected no human conviction. Arguments that work for a journalist or a reviewer may not transfer to an automated output, which changes how you would defend one. Strip the human out of the loop and you may strip out the defence with them: honest opinion and fair comment lean on a person's belief and judgment, and an algorithm offers a court neither.
📨 A complaint you sit on is a risk you keep. Google received a cease-and-desist and didn’t resolve it; the unbroken repetition fed the “urgency” the court relied on. A more prosaic reminder of the value of decent legal operations systems.
If the injunction holds or the reasoning is picked up in other courts, businesses deploying AI will face a product design challenge to revisit the trade-off between usefulness with higher risk of hallucinations; and accuracy with more friction in UX. Primary source publishers may welcome the latter.
The hallucination was Google’s to fix. The principle is everyone’s to absorb.
FROM THE SIDEBAR
Quick signals worth clocking
🇬🇧 AI law firm Garfield AI won a contested trial at Wandsworth County Court — believed to be the first court case won using an AI lawyer anywhere. It drafted the documents and witness statements for a freelancer chasing a £7,000 unpaid invoice, for about £400 in fees; a human barrister did the advocacy, and the SRA-regulated firm carried the file. AI changed the cost of the claim, not who was accountable for it.
🌎 Thomson Reuters’ Future of Professionals report finds AI everywhere and governance lagging. Seventy-four percent use AI weekly, but 34% admit to using tools their organisation cannot see, and of those whose firm has a named AI strategy, 35% say it doesn't match what happens at their desk. 47% say final responsibility for an AI-assisted error rests with the individual professional.
🇪🇪 Estonia says it will be the first country to issue AI agents their own personal ID numbers, so a person delegating to a bot need not hand it their whole digital identity. Framed as convenience, it is really a support to the liability question the other stories raise: give the agent a traceable identity, and an autonomous action has a name attached to it. The policy does not yet have a start date or substantive details.
Enjoying the signal?
If this edition would help a colleague thinking through AI governance, enterprise risk or legal operating change, feel free to forward it on.
💬 Forward to a colleague
🧠 Was this forwarded to you? Subscribe here to get Profiles in Legal every Wednesday.
Here’s how I can help
I advise technology businesses and leadership teams on AI, product and regulatory strategy and enterprise readiness to move from informal AI adoption to scalable governance.
If your organisation is navigating AI deployment, maturity or commercial negotiation challenges, feel free to reply directly.
- Philip
Profiles in Legal examines how AI, governance and technology are reshaping modern businesses and legal teams.
This publication is for general information only and does not constitute legal advice. Seek professional advice for specific situations.