Margaret Atwood is not impressed with artificial intelligence, and she made that clear during an appearance at the Babell Literary and Cultural Festival in Porto, Portugal. The author of The Handmaid’s Tale and The Blind Assassin described her single experience with a large language model as underwhelming, dishonest by design, and a cautionary example of why the “garbage in, garbage out” principle remains as relevant as ever.
One Conversation With Claude Was Enough
Atwood said she used an AI chatbot exactly once, turning to Anthropic’s Claude for information about the British detective series Father Brown. The result was not what she expected. Claude delivered a confident answer that was simply wrong. Atwood characterized the response as a lie, though she quickly clarified that the model had no intent behind the error.
“Claude gave me the wrong answer, or it lied,” Atwood said. “Of course, it didn’t know it was lying because it’s not a human being; it’s a large language model.” She explained that the model had scraped and sampled television reviews, but because critics rarely reveal endings, the model was misled by the incomplete patterns it had ingested.
This is a textbook example of a well-documented weakness in current large language models. They do not retrieve verified facts. They generate statistically plausible sequences of tokens based on their training data. When the training data contains systematic gaps, such as reviews that omit spoilers, the model has no mechanism to recognize that its output is inaccurate. It does not know what it does not know.
Garbage In, Garbage Out Is Not a Slogan — It Is the Architecture
Atwood did not stop at the anecdote. She framed the issue as a structural one, reviving the decades-old computing maxim “garbage in, garbage out” and applying it directly to modern AI systems. The point is not that LLMs occasionally fail. It is that they cannot do otherwise. A model trained on scraped, previously published, and potentially outdated information has no internal concept of truth. It has statistical correlation.
“Human beings are not robots, but they are opportunists,” Atwood said, “so if there’s an easy way to cheat and it’s hard to detect, people will do it.” She extended this critique beyond the technology itself to the humans who deploy it without scrutiny. She called AI-reliant individuals “opportunists” looking for shortcuts, and warned that even business users who depend on AI outputs must verify everything because the system is fundamentally fallible by construction.
This argument lands with particular force in a publishing and editorial context. If a model trained on reviews cannot correctly answer a factual question about a television series, what confidence should professionals place in its ability to generate accurate code, legal analysis, or medical summaries? The answer, as Atwood implies, is none without independent verification.
Why This Critique Matters Beyond the Literary World
Atwood’s comments arrive at a moment when enterprise adoption of generative AI is accelerating across every sector. Companies are embedding LLMs into customer service workflows, document generation pipelines, and decision-support systems. The assumption underlying much of this deployment is that the models are “good enough” for the task. Atwood’s argument challenges that assumption at its root.
The “garbage in, garbage out” principle is not a metaphor. It is a description of the statistical foundation on which these systems operate. Every LLM reflects the biases, gaps, and errors present in its training corpus. If the data contains systemic omissions, the model will reproduce them as confidently as it reproduces verified facts. The model has no way to distinguish between the two.
For professionals who rely on AI tools in their daily work, the implication is straightforward: treat every AI-generated output as a draft that requires human review. This is not a temporary limitation that will be solved by the next model release. It is a feature of the architecture. As long as LLMs are trained on static, scraped corpora, they will continue to produce plausible falsehoods with the same confidence as correct answers.
What This Means for Professionals Using AI Today
The practical takeaway from Atwood’s critique is not to abandon AI tools entirely. They offer genuine productivity gains when used appropriately. The takeaway is to understand exactly what these systems are and are not. A large language model is a pattern-matching engine trained on a fixed dataset. It is not a knowledge base. It is not a reasoning engine. It is not a reliable source of truth.
Anyone using an AI assistant for research, content generation, or analysis should build a verification step into their workflow. Check the model’s outputs against primary sources. Confirm factual claims independently. Never assume that a confident, grammatically correct answer is a correct answer. The technology will continue to improve, but the fundamental limitation that Atwood identified is structural, not temporary. The machine does not know it is lying. The human using it must.
