Margaret Atwood Tried Claude Once, Got Wrong Answers, and Called It “Garbage In”

At the Babell Literary and Cultural Festival in Porto, Portugal, Margaret Atwood, author of The Handmaid's Tale and The Blind Assassin, shared her one and only experience with an AI chatbot. She tried Anthropic's Claude, asked it about the British detective series Father Brown, and got the wrong answer. Her verdict: the model did not know it was wrong because it is a large language model, not a human. She described the root problem as "garbage in, garbage out," a pointed reference to training data quality.

What happened

Margaret Atwood was interviewed at the Babell Literary and Cultural Festival in Porto, Portugal. The conversation turned to AI, and Atwood was direct about her experience. According to Deadline’s recap of the event, she has used an AI chatbot exactly once: Anthropic’s Claude.

Her test case was simple. She asked Claude about Father Brown, a British detective series. Claude gave her incorrect information. Her read on why: “It didn’t know it was lying because it’s not a human being; it’s a large language model.” The model had, in her telling, skimmed sources without actually understanding them.

She wrapped up her critique with the phrase “garbage in, garbage out,” pointing at the quality of the data AI models are trained on as the source of the problem.

Why it matters

Atwood is not a tech journalist or an AI researcher, which is exactly what makes her comments worth noting. She is a careful, skeptical user who tried a tool once, hit a real failure, and articulated what went wrong more clearly than most press releases do.

The “garbage in, garbage out” framing is not new, but it cuts to something the AI industry tends to gloss over. Large language models produce confident-sounding text regardless of whether the underlying training data was accurate, complete, or representative. A model does not flag its own uncertainty unless it is specifically designed and prompted to do so.

For business owners using AI tools for research, content, or customer-facing copy, the Father Brown example is a useful reminder. The failure mode is not dramatic. The model did not crash or refuse to answer. It answered fluently and incorrectly, which is the harder problem to catch.

Our take

Atwood’s critique is fair and specific, which is more than you can say for most AI commentary from public figures. She did not say AI is evil or that it will replace writers. She said she asked a question, got a wrong answer, and identified a structural reason why that happens.

From where we sit, the “garbage in, garbage out” problem shows up constantly in client work. AI tools are genuinely useful for drafting, summarising, and generating options. They are unreliable for factual lookups, especially on niche topics like a specific TV series, a local business, or a technical specification that may not be well-represented in training data.

The practical split looks roughly like this:

• Use AI for: drafting copy, brainstorming structures, rewriting for tone, summarising documents you supply directly.
• Do not rely on AI for: verifying facts, sourcing quotes, researching specific titles or names, or any claim you plan to publish without checking.

Atwood’s single test was a factual lookup on a niche subject. That is close to the worst case for current LLMs. It does not mean the tools are useless. It means the tools have a clear boundary, and users need to know where it is.

What to do about it

If you or your team use AI chatbots for research, build a simple rule: anything the model tells you as a fact gets checked against a primary source before it goes anywhere public. That includes product names, episode counts, dates, quotes, and statistics. The model will not tell you when it is guessing. That check is your job.

Source: The Verge · AI