Can you reliably tell whether an online article was written by AI?

Why text alone cannot give a verdict

People often point to repeated phrases, a smooth but generic tone, awkward analogies, certain punctuation, headings, or a suspiciously broad summary. Any of those can be a useful prompt to investigate the article's claims. None is a reliable authorship test. Human writers use templates, editors standardize style, translation changes phrasing, and an author can revise generated text until no distinctive pattern remains. The reverse is also true: a person writing in a formal, predictable style can resemble the pattern a detector expects.

Detectors classify text using patterns learned from examples. They do not inspect a hidden record of keystrokes, model calls, or a writer's intent. Their performance can change with the model, language, passage length, prompt, domain, editing, and the rate at which the tool is allowed to falsely flag human writing. A score such as “82% AI” is therefore a tool-specific estimate under unstated assumptions, not a measurement of authorship.

There are concrete reasons to be cautious about false accusations. OpenAI retired its own text classifier in 2023 because of low accuracy; in its published English challenge set it identified 26% of AI-written text as likely AI-written and incorrectly labelled 9% of human-written text that way. OpenAI classifier notice That result describes one retired classifier, not every detector available today. It does show why a detection score should not be treated as proof.

The risk is also unevenly distributed. In a 2023 study of several widely used detectors, all tested detectors together flagged 19.8% of human-written TOEFL essays as AI-authored, and at least one detector flagged 97.8%. Liang et al. in Patterns This is a specific set of tools and essays, not a rate for all online articles or every current product. It is strong evidence that linguistic style can produce serious false positives, especially for writers who are not native English speakers.

Read the evidence before guessing authorship

Authorship and reliability are different questions. A human writer can misquote a report, copy an error, omit a conflict of interest, or make a misleading inference. An AI-assisted article can still point to a genuine source. The immediate reading task is to determine whether the claim holds up, regardless of how the sentences were produced.

Use this short routine when an article matters to a decision:

1. Identify one checkable claim. Prefer a number, a quotation, a policy change, a scientific result, or a claim that a person or organization did something.

2. Follow the strongest cited source. Open the report, official record, study, filing, or full interview rather than relying on the article's paraphrase. Check its date, jurisdiction, sample, definitions, and limitations.

3. Look for missing support. An article that cites nothing, names only vague “experts,” links to irrelevant pages, or gives quotations that cannot be found deserves less confidence. A real link still needs to support the sentence next to it.

4. Cross-check consequential claims. Compare an important fact with an independent primary or authoritative source. For fast-changing information, such as a product feature or a public policy, check the current official page.

5. Record uncertainty honestly. If the source is absent, inaccessible, or contradictory, pause before repeating the claim. “I could not verify this” is more accurate than deciding it must be AI-written.

Hypothetical example: An article says that a city has “eliminated” a household fee next month. Its prose is unusually repetitive, so a reader is tempted to call it AI-written. The reader instead opens the city notice, finds that the fee is waived only for a defined group and only after an application, then checks the effective date. The article is materially misleading whether it was written by a person, an AI system, or both. Verifying the policy resolves the decision the reader actually faces.

What can support an authorship claim

Reliable authorship evidence is usually external to the prose. It documents a relationship between a person or organization, a work process, and the published article. Even strong evidence may establish only part of the story because writing can be collaborative and AI use can range from spellchecking to a generated first draft.

Content Credentials are a useful example of provenance, not an AI detector. The current C2PA specification describes signed assertions about an asset's creation and edits that can be validated as associated with that asset and free from tampering. It explicitly says the system does not make a value judgment that provenance data is “good” or “bad.” A valid credential can therefore increase confidence in the origin record it contains, while leaving the reader to assess the signer, missing history, and the article's truthfulness.

For an ordinary web article, the best available evidence may be mundane: a publisher's corrections page, previous reporting by the byline, a transparent methodology, interviews with identifiable sources, and an editor willing to answer a narrow question. Several independent pieces of evidence are more persuasive than one detector result. Avoid demanding private drafts from an author unless you have a legitimate reason and appropriate authority to do so.

Respond to a suspicion without making a false claim

If an article seems synthetic or unreliable, describe the observable problem. Say that a citation does not support the stated conclusion, a quotation cannot be located, an image is miscaptioned, a claim has no date, or an author has not answered a specific sourcing question. Those statements invite correction and can be checked by others. Saying “this was written by AI” claims knowledge you probably do not have.

When the article is published by a site with an editorial contact, send a short, factual query. For example: “The article says X, but the linked report says Y. Can you identify the source for X or correct the passage?” If the publisher requires disclosure of AI assistance, ask how that policy applies to the article. Keep the question about a policy, source, or documented process rather than treating a detector output as evidence of misconduct.

If the article affects a purchase, health choice, public claim, workplace decision, or other consequential action, do not wait for authorship to be settled. Use the source material or a qualified official route to make the decision. A generic detector may still be useful to a researcher as one clue when combined with independently verified records, but it is a poor basis for public allegations, discipline, or a conclusion that an article is false.

The National Institute of Standards and Technology advises organizations evaluating content-provenance techniques to measure both false positives and false negatives and to document their limitations. NIST Generative AI Profile That standard of caution is appropriate for everyday reading too: look for evidence, keep the detector's limitations in view, and separate uncertainty about authorship from evidence about the claim.