Margaret Atwood says AI’s “garbage in, garbage out” fails readers with wrong answers

Margaret Atwood does not sound impressed by AI, and her reasoning is painfully concrete. In a Deadline recap of her interview at the Babell Literary and Cultural Festival in Porto, Portugal, the storied author of The Handmaid's Tale and The Blind Assassin said she used an AI chatbot exactly once, Anthropic's Claude, and came away unimpressed because it delivered the wrong answer. She was looking for information about the British detective series Father Brown, and, as she put it, “Claude gave me the wrong answer, or it lied.”

That is the headline stake in plain English: Atwood’s issue was not “AI is imperfect.” It was that the system can output information that sounds confident but is simply incorrect, and it may not even have the human concept of honesty behind the words. She added that “it didn't know it was lying because it's not a human being; it's a large language model.” In other words, she frames the failure mode as structural, not personal.

For executives and board members, that matters because the product risk is not just bad facts. The risk is trust. Large language models are designed to generate text that reads like answers, and when those answers are used in high-stakes contexts, the harm can compound quickly. Atwood’s example is small scale and literary. But the underlying question she raises is the same one regulators, buyers, and risk teams keep circling: how do you prevent confident nonsense from being treated like knowledge?

The phrase Atwood used, “garbage in, garbage out,” is often repeated in AI governance discussions because it is a clean way to describe a messy reality. Models can be trained on large datasets, but the world you are trying to query does not always map neatly onto what the system has seen, what it has stored, or what it can verify on demand. Even when a model is not “lying” in the human moral sense, it can still produce an output that is wrong. The second-order problem is that many workflows do not force a verification step. If a tool returns an answer instantly, teams may adapt their process to treat the output as a first draft of truth instead of a prompt for confirmation.

This is where the literary and cultural framing in Atwood's appearance becomes relevant for decision-makers. Her complaint surfaced in a festival setting, but it lands directly on the deployment question business leaders face: what is the acceptable level of error for the use case? For customer support, a wrong detail might create a refund ticket. For compliance, it can create audit findings. For healthcare or legal contexts, it can create real-world consequences. Atwood’s interview does not provide a new regulatory rule, but it reinforces a pressure point that already exists in AI policy: transparency about limitations, and controls that require humans and systems to catch mistakes.

Regulatory attention to AI has been accelerating across jurisdictions largely because of exactly these trust gaps: outputs that appear authoritative, even when they are not reliable. When a board asks for AI risk frameworks, the core is usually the same: identify where models can fail, quantify acceptable error, and define guardrails. Atwood’s “wrong answer” for Father Brown is a micro-incident that illustrates the category of failure, not the exact policy mechanism. Still, her comments are a useful reminder that governance is not only about preventing unsafe behavior. It is also about preventing incorrect information from being presented as if it is correct.

There is also an incentive dynamic that boards should keep in mind. AI product teams often optimize for helpfulness, speed, and conversational fluency. Those metrics can unintentionally reward confident phrasing. If the system is evaluated on perceived helpfulness rather than verified accuracy, it can become better at sounding right while remaining wrong. Atwood’s blunt observation that Claude “gave me the wrong answer, or it lied” points at the gap between language that persuades and language that is true. Even if the system is generating plausible text from patterns, the end user experiences it as an assertion.

For peers in similar roles, the strategic stake is simple: credibility is an asset, and misinformation is corrosive. The fastest way to lose customer and stakeholder trust is not one dramatic failure, it is a pattern of small errors that users only notice after they have already acted. Atwood’s example is a reminder that “garbage in, garbage out” is not an abstract slogan. It is a description of how an AI output can fail the human who just asked a normal question, and why that failure should trigger guardrails, verification, and accountability, not hand-waving.