Persons, Selectively
A German court ruled Google liable for AI-generated defamation. The slippery slopes are real. So is the principle behind the ruling.
By Geordie Everitt
A German court ruled recently that Google is liable for defamatory content in its AI-generated search summaries. The tech industry's response was predictable: slippery slopes, unintended consequences, the practical impossibility of holding any platform accountable for algorithmically generated content at scale.
Those concerns are real. They don't address the question the ruling actually raised.
The Original Bargain
Corporate personhood is one of the more remarkable legal fictions in modern history. The corporation-as-person was developed, over two centuries, primarily as a shield: an entity that could enter contracts, own property, and absorb liability on behalf of its owners without exposing them to personal consequence. That was the deal. Shareholders risked their capital. They did not risk their freedom, their reputation, or their personal assets beyond their investment.
The deal has since been renegotiated, mostly without announcement.
Corporations now claim freedom of speech — Citizens United v. FEC, 2010, established that political spending is protected speech, and that corporations are entitled to it. They claim religious freedom — Burwell v. Hobby Lobby, 2014. They claim Fourth Amendment protection against unreasonable search. They reach, in short, for every right developed by and for individuals: rights that exist specifically because individuals have skin in the game, reputations to protect, and bodies that can be incarcerated.
The liability side of that ledger has not kept pace. Corporations externalize costs — to workers, to communities, to the environment — with a consistency that would be criminal if a person did it. Section 230 granted platforms broad immunity for user-generated content. Legal departments exist specifically to make accountability expensive enough to be impractical. The entity originally designed to absorb liability has become, in practice, a mechanism for avoiding it.
The German ruling is, at its most basic, an attempt to rebalance the ledger. If you present AI-generated summaries as authoritative information about real people — if you claim, in effect, the standing of a publisher — then you accept the obligations of a publisher. Including the one where you don't defame people.
Free Speech Requires Skin in the Game
Free speech is an individual right. It was developed for individuals because individuals have something at stake in what they say. A person who lies publicly faces consequences: reputational, social, and in extreme cases legal. Those consequences are not incidental. They are the mechanism that makes speech a right rather than a weapon — the friction that gives the right its shape.
An entity with no reputation of its own, no personal relationships, no stake in the social fabric, has a different relationship to speech than an individual does. When a corporation says something false, the individual it was said about bears the consequence. The entity that generated it does not. The asymmetry is structural, and by design.
Extending speech rights to corporations was always a category error — the assertion that an entity created to limit personal liability deserves the rights developed for persons who accept personal liability. The German court, somewhat inadvertently, demonstrated one practical consequence: when an AI system generates a false claim and a corporation presents it as authoritative, someone is accountable for that claim. In the absence of a person, it must be the entity.
This is not a new principle. It is the original one, applied to a new substrate.
The Slopes Are Slippery. The Floor Is Still a Floor.
The complications are genuine. If AI-generated summaries create publisher liability, does that extend to search results? To autocomplete? To outputs that reflect patterns in training data the company neither created nor reviewed? The line between generating content and surfacing content has never been clean, and large language models make it significantly murkier.
The ruling will be appealed, narrowed, and argued over for years. The EU's AI Act layers additional regulatory complexity on top of it. Some version of this litigation will wind through European courts for a decade.
None of that changes the underlying question: if a system makes authoritative claims about real people, and those claims are false, who pays the cost? For the greater part of recent history, in the United States especially, the answer has been the person the claim was made about. Corporate-friendly governments — and courts — have spent a century ensuring that the costs of corporate behavior land on individuals rather than shareholders. Citizens United extended that logic into the political domain, granting corporations political voice while insulating them from political accountability. The AI context is the same argument, run one more time.
The instinct to hold corporations accountable for their outputs is not universal, and it faces powerful institutional resistance — the same resistance that has protected corporate behavior from consequence across every domain it has touched. That resistance will show up in this case, too.
The principle the German court applied is not exotic. It is legible, and it is old: you do not get to claim the rights of a person while systematically avoiding the obligations of one. The corporation exists to absorb liability. The novelty is that it is finally being asked to.
What the ruling cannot do is make AI output trustworthy. And creating the impression that it can is its most dangerous unintended consequence.
The Other White Meat
We have been here before, in a different industry.
The FDA and FTC once placed strict restrictions on health claims in food advertising. Manufacturers could not assert that eggs lowered cholesterol, that milk built strong bones, or that beef was a health food — because the evidence did not support those claims, and regulators said so.
What followed was one of the more effective campaigns of implied benefit in commercial history. "The Incredible Edible Egg" made no health claim — it simply positioned eggs as remarkable. "Got Milk?" implied you needed it, without specifying why. "Beef — It's What's for Dinner" normalized consumption as cultural identity. "Pork — The Other White Meat" borrowed health associations from chicken without claiming any directly. "Coca-Cola Is the Real Thing" offered authenticity as a proxy for goodness.
None of those slogans said anything specifically false. All of them moved consumption in a direction that decades of epidemiological evidence would eventually characterize as harmful.
Now look at what replaced them. Eggs are no longer incredible — they are "Stronger by the Dozen," a wellness implication delivered without a wellness claim. Milk no longer does a body good — dairy now "Does More," a benefit so vague it cannot be falsified. Pork retired "The Other White Meat" in 2025 after 36 years and replaced it with "Taste What Pork Can Do" — pure sensory language, nothing to regulate, nothing to refute. Beef has used "It's What's for Dinner" since 1992 without modification, because it never made a claim to begin with. Identity and appetite; nothing a regulator can touch.
This is not a retreat. It is a strategic upgrade. The old slogans implied health benefits and were therefore vulnerable to health science. The new slogans imply nothing falsifiable. They are more durable precisely because they claim less. Restricting one dimension of deception did not eliminate deception. It refined it into a form that regulation cannot reach.
The German ruling closes a specific door: Google cannot present AI-generated lies about real people as authoritative fact. The door it inadvertently opens is the implication that regulated AI output is trustworthy output. The population that internalizes that implication will be worse off than the population that stayed appropriately skeptical.
The Sycophant Problem
There is a category of AI harm the ruling cannot touch at all: the AI that tells you what you want to hear.
Ask a sufficiently accommodating AI to evaluate your business idea — the one that is, in your closest friend's private assessment, genuinely terrible — and it will find the angle that makes it sound viable. It will identify the addressable market. It will cite comparable cases. It will generate the encouragement you came for. Nothing it produces is defamatory. Nothing it says is false in a way a court could adjudicate. It simply agrees with you, fluently and at length, in the direction of your existing beliefs.
Confirmation bias as a service. Not regulatable, because the AI is not making claims about a third party — it is reflecting your premises back at you with professional polish. The harm is to your judgment, not to anyone Google could be sued over.
Which is where the burden has always ultimately landed. Regulating corporate deception is necessary and insufficient. The food industry found its way around explicit health claims. Corporations will find their way around AI defamation liability. The residual obligation — the one that cannot be delegated to a regulator or discharged by a ruling — falls on the person doing the asking.
Some fraction of any population will accept information that confirms what they already believe, regardless of source, quality, or provenance. I won't put a number on that fraction. Any figure I named would arrive with the same air of authority as if an AI had generated it — which is, more or less, the problem in miniature.
Geordie