Margaret Atwood Tried Claude Once, Got a Wrong Answer, and Has Thoughts About AI

Margaret Atwood — author of The Handmaid's Tale and The Blind Assassin — used an AI chatbot exactly once. One try was enough.

Atwood was speaking at the Babell Literary and Cultural Festival in Porto, Portugal on June 27, 2026. According to Deadline's recap of the event, she described asking Anthropic's Claude about the British detective series Father Brown and receiving wrong information.

Her explanation for why it happened was more precise than most tech critics manage:

"Claude gave me the wrong answer, or it lied. Of course, it didn't know it was lying because it's not a human being; it's a large language model... It had skimmed and sampled a lot of television reviews, but they never give away the ending in online criticism, so it was misled by the things it had read about the show."

That's a technically accurate description of a well-documented failure mode. Large language models generate responses by predicting likely text based on training data. If the training data about a subject is systematically incomplete — critics don't spoil endings, so the model never learned the endings — the model fills the gap with plausible-sounding fabrication. The AI research community calls this "hallucination." Atwood called it lying. The practical result is the same.

The Garbage-In Problem

Atwood's core critique wasn't emotional. It was structural. "The thing about AI is that it's garbage in, garbage out," she said. "Even people who use it for business reasons have to check it because it makes mistakes."

This is not a fringe position. It's the reason every major AI company — including Anthropic, OpenAI, and Google — includes disclaimers telling users to verify outputs. The disclaimers exist because the systems hallucinate, sometimes confidently and convincingly.

The strongest counterargument to Atwood's take is that one failed query on a niche trivia question isn't a fair benchmark for a technology that handles millions of tasks daily, many of them accurately. Coding assistance, document summarization, translation, and research synthesis are areas where LLMs demonstrably save time and produce usable results — when users know enough to catch mistakes. That last condition is the catch, and it's the one Atwood is flagging.

She didn't dispute that AI has uses. She disputed the habit of trusting it without verification.

On the People, Not Just the Machine

Atwood reserved separate criticism for users who treat AI as a shortcut with no oversight. She called them "opportunists" and pointed to the incentive structure directly: "Human beings are not robots, but they are opportunists, so if there's an easy way to cheat and it's hard to detect, people will do it."

This cuts across the AI debate in a way that doesn't map neatly onto left or right. The concern about AI-generated work flooding publishing, academia, and media is genuine and bipartisan. Publishers are struggling to detect AI-written submissions. Schools have banned and then unbanned AI tools with no coherent policy. Businesses are deploying AI-generated customer communications without telling recipients. The "hard to detect" part of Atwood's observation is, as of June 2026, still largely true.

What Atwood Didn't Address

Atwood's critique focused entirely on accuracy and opportunism. She didn't engage with questions about copyright, including the ongoing legal disputes over whether AI companies trained on authors' work without permission, a fight that directly affects writers like her. Several lawsuits from authors' groups against major AI developers are still working through the courts. Atwood has spoken on those issues in other contexts, but Porto wasn't that conversation.

The Verge covered the remarks straightforwardly, framing Atwood's critique through the specific technical explanation she gave about training data. Deadline's original recap, which The Verge cited, was the primary source.

Where This Lands

Atwood's "garbage in, garbage out" argument doesn't require her to be an AI expert to be valid. The principle predates machine learning by decades. It's a foundational concept in computing. Applied to LLMs, it means a model trained on incomplete, biased, or outdated text will produce incomplete, biased, or outdated responses.

The open question isn't whether hallucination is a problem. Every serious AI researcher agrees it is. The open question is whether the verification burden Atwood describes — checking AI outputs because they make mistakes — will narrow as the technology improves, or whether it's a permanent feature of systems that generate plausible text rather than retrieve verified facts. Anthropic, OpenAI, and Google have all promised improvement on that front. The track record so far is mixed.