How to Tell If a News Article Is AI-Generated

The Challenge of AI Content Detection

As large language models have become capable of producing fluent, plausible-sounding journalism, the ability to identify AI-generated content has become an important skill for media professionals and engaged readers. The challenge is substantial: frontier models like GPT-4o and Claude 3.5 Sonnet produce text indistinguishable from competent human writing for most readers in most contexts. No detection tool achieves near-perfect accuracy, and both false positives (flagging human writing as AI) and false negatives (missing AI-generated content) carry significant consequences in journalistic contexts.

Linguistic Patterns Associated with AI-Generated News

While not individually definitive, several linguistic patterns are systematically more common in AI-generated news text than in professional human journalism: excessive hedging without specificity ("sources suggest," "experts indicate," "many believe" without named attribution); structural uniformity (near-perfect five-paragraph essay structure regardless of topic complexity); uncommon phrase combinations that occur frequently in training data ("it's worth noting that," "delve into," "tapestry of"); generic specificity (statistics and facts that sound precise but lack direct source attribution); and absence of genuine editorial voice — AI text tends toward a neutral, encyclopaedic register even when a human journalist with a byline would naturally have a distinctive voice.

Technical Detection Tools

Several AI content detection tools are available, though all carry significant false positive and false negative rates. GPTZero provides probabilistic assessments of AI authorship using perplexity and burstiness metrics (human writing has more variable sentence complexity than AI text). Originality.ai combines AI detection with plagiarism checking and is popular among editorial teams. Turnitin's AI detector is widely used in academic contexts but also applicable to journalistic content. Copyleaks offers sentence-level AI authorship detection.

All detection tools should be used as screening instruments rather than determinative verdicts. A detection score of "70% likely AI-generated" warrants additional scrutiny but is not proof of AI authorship — particularly for non-native English speakers or writers with unusually clean prose styles.

Editorial Disclosure Policies

In response to growing audience concern about AI-generated content, major news organisations including the BBC, AP, Reuters, The Guardian, and the New York Times have published AI use policies that require disclosure when AI tools have been significantly used in the production of published content. The AP's policy specifies that AI-generated text may not be published as news without significant human editing and editorial verification. The BBC requires that any AI-generated content be explicitly labelled and that human journalists remain accountable for the accuracy of all published material regardless of how it was produced.