October 1, 2026 | Hallucination v. Defamation: Actual Malice in the Age of Artificial Intelligence

The sudden arrival of Artificial Intelligence has introduced any number of questions that sound as though they require entirely new categories of law. Defamation would seem to be one of them. What happens when a machine falsely announces that someone is a murderer? Who, exactly, published the statement? And where does actual malice reside when no actual person wrote the offending sentence—or perhaps even knew it existed?

Last week, the Federalist Society convened Professors Eugene Volokh and Lyrissa Lidsky for a webinar aptly titled Large Libel Models? When Do AI Hallucinations Become Defamation? The discussion centered on two recent cases against Google. In Keene v. Google, an AI Overview allegedly reported that James Keene—who once served as a prison informant against a suspected serial killer—was himself serving life imprisonment for murdering three women. In Starbuck v. Google, the allegations were more numerous and similarly imaginative. Both cases survived Google's motions to dismiss, although both remain at preliminary stages.

Yet as the professors worked through the problem, the discussion kept returning to surprisingly familiar ground.

Volokh explained the general temptation to locate the legally relevant mind inside the machine. For actual-malice purposes, he said, “the real focus is on . . . the mental state of the company, not the program.” The program “may not have a mens” at all; the company, by contrast, “employs people who do.”

That becomes particularly important after notice. Keene alleged that he notified Google three times about false information concerning him; Starbuck alleged repeated notice efforts of his own. Lidsky therefore found these cases comparatively easy at one crucial point: AI hallucinations may sometimes be inevitable, but once a company receives specific notice of a particular defamatory falsehood, what it does next becomes an ordinary question of responsibility. Her practical advice to AI companies was correspondingly old-fashioned: first install a functioning notice-and-cure system. Then use it.

What constitutes a reasonable cure is harder. The panel discussed block lists, source retrieval, machine unlearning and other safeguards of varying cost and effectiveness. Volokh compared the problem to product design: declining a precaution that would add $50,000 to the price of a car may be reasonable; declining an effective five-dollar precaution may not be. Lidsky likewise emphasized that technical feasibility will evolve with the technology. More broadly, she warned against the notion that one can simply “wave your hands and say ‘technology’ and get a ‘get out of jail free’ card.”

The question, then, is not whether every hallucination can be eliminated, but which errors remain genuinely unavoidable after reasonable precautions—and what counts as reasonable at that particular moment in technological history.

Underneath all of this sits a more basic problem: Is the machine retrieving somebody else's authorship, distributing it, or generating its own?

Lidsky suggested that generative-AI providers might in some respects be treated like traditional distributors—a bookstore, for example—which historically acquire greater responsibility once put on notice of defamatory material. Volokh identified the complication: a bookstore did not write the books on its shelves. Generative-AI companies, by contrast, deploy systems that actually create the challenged output. In this light, authorship suddenly has consequences.

Even the machine's apparent account of its own intentions does not solve the problem. In one of the stranger aspects of his case, Starbuck alleged that Google's AI told him it had been deliberately producing lies about him. Volokh's response was wonderfully concise: “I don't think chat bots admit things.” Lidsky agreed that treating such output as reliable evidence was troubling. Human-sounding language invites human-sounding explanations. Defamation law still requires evidence.

By the end of the hour, there were audience questions but almost no time left. Volokh asked that they take one. Moderator Shlomo Klapper began reading a question that sounded oddly familiar to me. The reference to actual malice being “brought home” to those responsible for publication caught my attention; by the time he reached “a new theory of attribution,” I realized why. It was the question I had submitted the previous evening.

The “brought home” language comes from New York Times v. Sullivan: the constitutionally required state of mind must be brought home to the people within an institutional defendant who bear responsibility for the publication. My question was what happens to that old rule when an AI system generates and publishes the particular factual assertion without any human having composed, reviewed or perhaps even known about it. Does existing doctrine suffice, or does autonomous AI publication create an accountability gap requiring a new theory of attribution?

Neither professor took the futuristic exit.

For the cases discussed that day, Lidsky said, the answer becomes “pretty easy” after notice. The harder actual-malice problem arises before notice: generalized knowledge that an AI system has some statistical rate of false output does not establish knowledge of a particular defamatory falsehood. Volokh agreed entirely, distinguishing that constitutional problem from negligence-based claims, where reasonable design and precautions may matter considerably. 

I had asked whether autonomous AI might require a new theory of legal attribution. Their answer sent the inquiry somewhere older. The AI may be novel; but traditional legal notice remains perfectly operative.

That may be the more useful lesson. Novel technology does not necessarily require novel first principles. Before inventing entirely new categories of responsibility for machines, the law can still ask its familiar questions: Who published? What was false? What did the defendant know? When did it know it? And what could reasonably have been done?

Put more simply, AI-generated libel looks like a futuristic legal problem until the law begins asking its ordinary questions.

My previous Note considered what happens to authorship when AI increasingly answers in an author's place. Libel presents the question from the other direction: When that answer is false, where does responsibility go—and to whom can it be brought home?

The machine may be new; but the humans behind it are not.

Or, before inventing a new legal person, it may be worth finding the old ones. 

                                                                                    *

For further reading:

Volokh: Large Libel Models — collected writings

Lidsky & Daves: Inevitable Errors: Defamation by Hallucination in AI Reasoning Models

Starbuck v. Google LLC: July 24, 2026 opinion denying Google's motion to dismiss

Keene v. Google LLC: September 14, 2026 memorandum opinion and order

                                                                                    ***

September 17, 2026 | The Author Has Left the Building

Some months ago, I quietly removed most of the substantial writing from this website. Essays on constitutional law, birthright citizenship, presidential power, Skrmetti, Ramaswamy, Trump, and various other subjects disappeared. I didn't stop writing, and I haven't particularly changed my mind about the value of writing, or the positions I took. What changed was the internet.

EffingtonEastman.com was built under a bargain so ordinary that nobody needed to state it: I wrote something; Google indexed it, somebody searching for the subject might find the title, click the result, arrive here, and encounter the person who wrote it. No idealism necessary; it was simply how the internet worked. Google was hardly a charitable institution, and search was hardly pure. But that era-defining architecture preserved something important: the search engine pointed toward the source.

Google’s new AI search model changes that bargain. Google itself has described generative AI as a way to “do the searching for you.” Ask a sufficiently capable AI a question and it may no longer be necessary to visit the writer, the paper, or the peculiar little website where the idea originated. The new AI model simply synthesizes what it has learned and answers in its own remarkably competent voice. And here is the problem as I see it: Google used to introduce the questioner to the author, while now AI increasingly answers in the author's place.

This reconfiguration of search is extraordinarily convenient (I use it myself every day). It is also troubling. And it's why I took my essays down.

Google itself announced the change without ambiguity. In 2025, CEO Sundar Pichai called its new AI Mode a “total reimagining of Search”—an end-to-end AI search experience built not merely to locate information, but to reason across it and answer increasingly complicated questions. A year later, Google described its AI-powered Search box as the biggest upgrade to Search in more than twenty-five years.

The consequences aren't merely theoretical. Rather neatly supporting the point, a 2025 Pew Research Center study found that Google users presented with an AI summary clicked through to traditional search results only about half as often as users who weren't—and clicked the sources cited within the AI summary only 1 percent of the time.

Meaning, what we are talking about here represents nothing less than a generational paradigm shift. 

Whether the resulting arrangement also qualifies as another specimen of what Cory Doctorow has memorably called “enshittification,” I'll leave for another Note.

Ok, progress happens. Fantastic. And Google never promised to preserve its old model forever.  I—and countless others—nevertheless built our websites around it. Then Google changed its bargain. 

Fair enough. So I changed mine.

The difference matters particularly to people who create things. Under the old arrangement, somebody asking an obscure question might end up at an obscure website. Search was an introduction. A question about, say, §3 of the Naturalization Act of 1795 might lead to several legal writers offering different interpretations of the same language. The reader had to encounter the sources—and decide among them.

There was a good kind of friction in this system: you had to choose among sources. You might find the answer, but you might also find the person who said it. Sometimes, knowing the source is part of understanding the answer.

AI is vastly more efficient. Increasingly, you simply ask—and get the answer.

And it is that very efficiency that changes the economics of authorship. AI is driving the marginal cost of accessing synthesized linguistic knowledge toward zero, while increasingly separating an idea from the identity, context, traffic, meaning, and compensation of the person who originally developed it.

The problem, for me, is less theft than attribution. Suppose somebody spends several weeks researching an obscure legal question, discovers a connection that hasn't been adequately explored, and publishes 7,000 words explaining it.  Once that work enters the accessible digital corpus, a future reader may never find the writer, visit the website, or encounter the argument in its original form. Under the new model, the intellectual residue increasingly reappears through an AI interface as simply something the machine knows. The original insight becomes general information; its provenance fades into the corpus.

That is where attribution becomes something more than a footnote. Sometimes the source is part of the meaning, even when the answer can stand without it.

Much of what made the old model interesting was precisely this profusion of small, peculiar sources whose purpose or value might not be immediately obvious.  Sometimes their pointlessness was part of the point.  

Somebody somewhere cared enough about something to build a website around it. Search provided the introduction. The reader decided whether to stay.

This is where my complaint becomes complicated, because I love using AI. My objection is therefore not to AI learning from human knowledge. Nor do I particularly yearn for some restoration of the internet circa 2004. Something has been gained here; enormously so. But something has also disappeared.

I don't know that the answer is to stop AI.  I trust AI with ideas, yet I don't necessarily trust the present information economy with my authorship.

Whatever copyright may ultimately mean as courts, legislatures, publishers and technology companies sort through these questions, AI has already destabilized the practical bargain between public authorship, attribution, readership and compensation. I don't need to resolve that argument for everybody else. I only need to decide what I am going to publish here.

So I did. And the long essays came down.

They may continue to be written. They may appear elsewhere, in other forms, or under circumstances where the relationship between author and work makes more sense to me. As every intentionalist knows, withdrawal is itself an exercise of authorship. 

The old arrangement produced the essays. The new one has produced something else. As of September 2026, this page is called Notes.

I haven't stopped writing because of AI; I've stopped assuming that publishing substantial original work to the open internet remains the same bargain it used to be. It isn’t.

Here, the thinking continues, and the writing continues. They simply no longer come with the assumption that the internet is entitled to them.