How to Spot AI-Generated Text, Images and News
Last updated June 2026
AI-generated content has become good enough that "does this look real?" is no longer a question you can reliably answer by eye. The better question — the one that actually protects you — is "is this true, and where is the evidence?"
A few years ago you could often catch machine-made content at a glance: stilted prose, mangled hands in photographs, captions full of nonsense letters. Those giveaways still exist, but they are fading fast. Each new generation of models writes more naturally, renders more convincingly, and makes fewer of the obvious mistakes people learned to look for. Treating any single "tell" as proof is increasingly risky.
So this guide does two things. First, it runs through the signs that something might be AI-generated, with honest caveats about how weak those signs are becoming. Second — and more importantly — it explains why none of that detection should be the basis of your decision. Whether a human or a machine produced a claim tells you almost nothing about whether the claim is true. What matters is the evidence behind it.
Signs that text may be AI-generated
Language models are trained to produce fluent, plausible-sounding text, which is exactly what makes them hard to catch. Still, a few patterns recur often enough to be worth noticing — as prompts to dig deeper, never as a verdict.
- Generic, frictionless phrasing. AI prose often glides along without ever committing to a sharp, specific point. It hedges, summarises and balances, but rarely says anything that only a particular person with particular knowledge could say.
- Confident vagueness. Watch for text that sounds authoritative while staying suspiciously non-specific — sweeping statements with no named source, date, place or person attached.
- Fabricated citations and "hallucinations". This is one of the more useful signals. Language models can invent studies, quotations, statistics, court cases and book titles that simply do not exist, presented with total confidence. If a reference cannot be found anywhere when you search for it, that is a serious red flag regardless of who wrote the piece.
- No real reporting. Genuine journalism contains things a model cannot invent: a named source who was actually spoken to, a document that was actually obtained, an event the writer actually attended. Content that summarises and rephrases without ever adding first-hand observation is consistent with automated generation — though plenty of low-effort human writing looks the same.
Notice that the most valuable item on that list — fabricated citations — is not really an "AI detector" at all. It is just fact-checking. You are not asking "did a machine write this?"; you are asking "does this source exist?"
Signs that an image may be AI-generated
Image generators improved dramatically in a short time, and the classic flaws are disappearing. The well-known artefacts are still worth a glance, but each one is a weaker signal than it was even a year ago.
- Hands, teeth and ears. Extra or merged fingers, too many teeth, or oddly shaped ears used to be reliable giveaways. Newer models get these right far more often, so their absence proves nothing.
- Garbled text. Lettering on signs, labels, packaging and documents in the background is often warped or meaningless. This remains one of the more durable tells, though it too is improving.
- Impossible light and physics. Shadows falling in inconsistent directions, reflections that do not match, jewellery or patterns that blend into skin, or backgrounds that dissolve into mush on close inspection.
- Too-perfect surfaces. Skin, hair and textures can look uncannily smooth or plasticky, with a glossy, over-polished quality.
The honest caveat is large here: these checks catch sloppy or older fakes, but a carefully made image from a current model may show none of them. And the reverse trap matters too — real photographs are sometimes wrongly dismissed as fake because someone "spotted" an artefact that was just ordinary compression, motion blur or an awkward camera angle.
Video and audio
Synthetic video and cloned voices are now realistic enough to fool casual viewers, and the tells are subtle and shrinking. In video, look for unnatural blinking or none at all, lips that drift slightly out of sync, hair or earrings that flicker between frames, and edges that warp where a face meets the background. In audio, listen for flat emotional delivery, odd pacing, missing breaths, or a strange lack of room ambience. But be realistic: a short, low-resolution clip shared on social media gives you very little to judge, and these are exactly the conditions under which fakes spread most easily. For anything important, identifying the original source of the footage will tell you far more than scrutinising the pixels.
Why AI-detector tools are not the answer
Given all this, it is tempting to reach for an automated "AI detector" and let it decide. Resist that temptation. These tools are unreliable in both directions, and the consequences of trusting them can be serious.
- False positives. Detectors regularly flag genuine human writing as machine-made — particularly clear, well-structured prose, and text written by people who learned English as a second language. People have been wrongly accused on the strength of these scores.
- False negatives. Lightly edited or paraphrased AI text often slips straight past detectors, so a "human" result is no guarantee of anything.
- A moving target. Detection is locked in a race it tends to lose: every improvement in generation erodes the patterns detectors rely on. A tool that worked last year may be near-useless against this year's models.
A confident-looking percentage score gives a false sense of certainty. Treating "98% AI" as proof is a mistake — and so is treating "100% human" as reassurance.
The reliable approach: verify the claims, not the authorship
Here is the shift that actually helps. Stop trying to determine who or what wrote something, and start checking whether what it says is true. Authorship is almost beside the point: a human can write a convincing lie, and an AI can summarise a true and well-sourced fact. The credibility of a claim lives in its evidence, not in its origin.
That means doing the same things a careful reader would do with any source. Can you trace each significant claim back to a primary source that genuinely exists and genuinely says what is claimed? Do named studies, quotes and figures hold up when you look them up? Is the same account reported independently by outlets that do their own reporting? These checks work identically whether the text came from a person, a model, or some blend of the two. For a step-by-step routine, see how to fact-check a news article yourself and lateral reading.
Provenance signals — and their limits
A more promising long-term answer is provenance: a record of where a piece of content came from and how it was made. Some AI tools now add visible or invisible watermarks to their output, and an industry standard called C2PA (often branded as "Content Credentials") aims to attach tamper-evident information about an asset's origin and edit history.
These are genuinely useful and worth taking seriously when present. But they have real limits today. Watermarks can be cropped, compressed or stripped out, and not every tool applies them. Content Credentials only help when the creator, the software and the platform all support the standard and preserve the data — and much of the internet still does not. Crucially, the absence of a provenance signal does not mean something is fake, and its presence does not by itself make a claim true. Provenance tells you about an asset's history, not about whether its underlying assertions are accurate.
What this means for news
For news specifically, the practical upshot is liberating. You do not need to become a forensic image analyst or trust a detector's score. The questions that have always separated trustworthy reporting from the rest still apply: Who is making this claim? What is their evidence? Is it sourced to something real? Is it corroborated elsewhere? A story can be entirely AI-assisted and still accurate, or entirely human-written and completely wrong. Judge the substance.
It is also worth keeping perspective. Cheap, fast generation makes it easier to flood the zone with plausible-looking material, which is a real problem. But it does not change what good evidence looks like. The defences that worked against misinformation before AI — checking sources, reading laterally, being wary of claims with no traceable origin — are the same defences that work now.
Fact or Fiction News doesn't try to guess whether a human or a machine wrote an article — that question rarely settles anything. Instead it assesses the claims: what the piece asserts, whether those assertions hold up, and what evidence supports them.