Humans cannot tell if news was written by AI or a person — even experts struggle

Researchers showed 1,054 people news articles written by humans and by large language models, then asked them to guess which was which. People failed. They couldn't reliably distinguish machine-written text from human-written text, even when they tried hard, and the failure held across six different AI models ranging from small to large. This means you cannot rely on reading something carefully and deciding whether it's AI or human — the structural difference is not visible to human judgment.

For years, the implicit defense against AI-generated misinformation was that humans could spot it if they paid attention. This paper shows that defense does not exist. The finding is clean: domain expertise helps a little (people who know a field do slightly better), but even experts fail most of the time, and cognitive fatigue makes detection harder the longer you read. The implication is direct — if detection at the reader level doesn't work, the only viable countermeasure is system-level: cryptographic proof of source, watermarking, or server-side verification of where text came from. Detection happens nowhere else.