AI-generated text, images, and video are now good enough that spotting them reliably has become genuinely hard. No detection tool is accurate enough to trust on its own, and the tells that worked a year ago keep disappearing. What still works is a set of habits — knowing what to look for, and knowing when it matters. This guide covers both.

Start with the honest limitation

Before the techniques, the caveat: you cannot reliably detect AI content from the content alone. Automated "AI detectors" produce both false positives and false negatives at rates that make them unsafe for consequential decisions — they have wrongly flagged human writing, particularly from non-native English speakers, and they miss lightly edited AI text.

Treat every signal below as weak evidence that shifts probability, never as proof. The goal is calibrated suspicion, not certainty.

Signals in written text

AI writing has recognisable habits, though they fade as models improve and people edit output:

  • Fluent but hollow. Every sentence is grammatical, well structured, and says surprisingly little. Real expertise usually includes specifics — numbers, names, dates, edge cases.
  • Suspiciously even rhythm. Human writing varies: long sentences, then short ones. Unedited AI text often maintains a steady, uniform cadence.
  • Balanced to a fault. Constant "on one hand, on the other hand" framing without ever committing to a position.
  • Missing lived detail. An article about a place or product with no concrete sensory or first-hand specifics.
  • Confident wrong facts. Invented citations, non-existent studies, or fabricated quotes stated with total assurance. This is the most useful signal, because it is checkable.

Signals in images

Image generators have fixed many classic flaws, but errors cluster in predictable places:

  • Text. Signs, labels, book spines, and logos often contain garbled or nonsense lettering.
  • Repeating background details. Crowds where faces or clothing patterns duplicate.
  • Physics that does not resolve. Shadows falling in inconsistent directions, reflections that do not match the scene, jewellery merging into skin.
  • Impossible connections. Where objects meet — hands gripping items, chair legs meeting floors, eyeglass arms meeting ears — is where generators still struggle.
  • Too-perfect surfaces. Skin without pores, materials without wear, lighting with no imperfection.

The technique that actually works: check the source

Verification beats inspection. Instead of squinting at pixels, ask:

  • Who published this, and do they have a reputation to lose?
  • Can I find the claim reported independently elsewhere? A dramatic photo of a real event usually has multiple photographers.
  • Does a reverse image search show an older original?
  • Do the named sources exist? Fabricated studies and quotes collapse in seconds under a search.

This is the same discipline that underpins good research generally, and it is why we recommend verifying anything consequential that comes out of a model — a point we make in our guide to writing better AI prompts.

Provenance metadata: the real long-term answer

The industry's serious response is not detection but provenance — cryptographic labels attached at the moment of creation. Standards such as C2PA content credentials, and invisible watermarks embedded by some image generators, are being built into major tools. Google's image models, for example, ship with content credentials and watermarking by default.

The limitation is that metadata can be stripped, and not every tool participates. Provenance proves what was labelled; its absence proves nothing. Still, this approach is more sustainable than an endless arms race between generators and detectors.

When it actually matters

Be pragmatic. Whether a marketing email was drafted by AI is usually irrelevant. It matters when:

  • The content makes factual claims you plan to act on.
  • It depicts a real person doing or saying something.
  • It is evidence in a dispute, a purchase, or a news event.
  • It asks you for money, credentials, or urgent action.

That last category is where AI content does real damage — not through fake articles, but through convincing scams. Basic security hygiene matters more here than detection skill, which is why phishing-resistant sign-in protects you even when a message fools you completely.

Key takeaways

  • AI detectors are unreliable in both directions and should not drive consequential decisions.
  • In text, watch for fluent-but-hollow writing, uniform rhythm, missing specifics, and confidently fabricated facts.
  • In images, check text, hands and joins, shadows, reflections, and repeating background patterns.
  • Verifying the source beats inspecting the content — reverse image search and independent corroboration are strongest.
  • Provenance standards like C2PA are the long-term answer, but missing metadata proves nothing.

The bottom line

Stop trying to tell whether something was made by AI and start asking whether it is true and who stands behind it. That question has always been the right one — AI has simply made it unavoidable.