AI for Normal People Part 4: How to Spot AI Content

By now you've used a chatbot (we covered that in Part 2), you've worked through the honest "is this thing dangerous" question (that was Part 3), and you've got a working mental model of what AI actually is (that was Part 1). Good. Now the practical question: when something lands in front of you online — an article, a photo, a video, an email from a stranger — can you tell whether a human or a machine made it?

The honest answer: sometimes, and less reliably than the marketing for "AI detector" tools wants you to believe. The category is genuinely useful to understand, but treat anything that claims 99% accuracy with deep suspicion.

This post is the field guide. We'll cover what to look for in text, in images, in video, and in audio — and then we'll talk honestly about why the "detector tool" market is mostly snake oil. By the end you'll have a working set of mental tools, plus a healthy dose of "it's complicated."

Three years ago, AI-generated text gave itself away constantly. Six fingers on a hand. The word "delve" used three times in a paragraph. Articles that talked confidently about non-existent books. Easy mode.

Things have moved on. Modern models (GPT-5, Claude Sonnet 4.6 and equivalents) write fluently enough that a careful human edit makes the output essentially indistinguishable from competent human writing. Image models (Midjourney v7, Google Imagen 4, Stable Diffusion 3.5) now produce hands correctly most of the time, render text legibly, and handle lighting consistently. The easy tells of 2023 have closed, and the new tells are subtler.

It is also true that the most-asked question about "AI content" — "was this written by AI?" — is often the wrong question. A lot of online content in 2026 is a human draft polished by AI, or AI text edited by a human, or an AI summary of a real source. The binary "human or machine" framing doesn't actually capture what is going on. The more useful question is usually "is this content accurate, original and worth reading," which is closer to old-fashioned media literacy than to a technical detection problem.

If you are looking at a piece of writing and wondering whether AI was involved, the most reliable indicators are pragmatic rather than technical. The patterns that recur in AI-generated text without a careful editing pass:

Bland fluency. Sentences that are grammatically perfect but say nothing distinctive. Lots of subordinate clauses, balanced structures, a vaguely TED-talk-ish rhythm. The classic AI fingerprint is sentences that sound polished but where you would struggle to summarise the actual claim if asked.

Lists of three. Long-form AI text loves a list of three things in a sentence. "Strategic, scalable and sustainable." "Fast, friendly and affordable." "Quality, innovation and customer success." Three is a magic number for AI writers; competent humans use it deliberately and rarely.

Em-dashes — used like this — to signal sophistication. AI models love em-dashes. Many overuse them. The presence of em-dashes alone is not a tell (this post uses several), but a piece that has them in nearly every paragraph and never uses a comma or a parenthesis instead is suggestive.

Hedging without commitment. "Many experts agree that…" "Studies have shown that…" "It is widely believed that…" When you can't name the experts or the studies, the writing was either lazy or AI-generated. Real informed writing names sources.

Lack of specificity. Real human writers about a subject have weird, specific knowledge. They remember a particular conversation, a specific date, a brand name, a smell, a price they paid in 2019. AI writing tends to deal in plausible generalities because the model is averaging across thousands of similar pieces. If the text is fluent but somehow weightless, that's the signal.

Confident wrongness about details. AI models will state facts with full confidence and the facts will sometimes be wrong. Made-up book titles, invented quotes, statistics that don't match any real source. Spot-checking three claims against Google takes thirty seconds and catches a lot.

The combined principle: AI text is fluent but lacks the kind of specific, unrepeatable evidence that comes from a real person actually doing or knowing something. Look for the things only a human-in-the-loop would have.

AI-generated images have improved fast, but the technology still leaves traces if you know where to look. The easy 2023 tells (six fingers, melted faces) are mostly gone. The 2026 tells are subtler:

Background detail that falls apart on close inspection. A photo of someone in a café — the café is the giveaway. Look at the menu boards behind the subject. The cake labels in the display case. The poster on the wall. AI image models are good at the main subject and weaker on background text and detail. Words on background signs are still often gibberish or near-gibberish in AI images.

Hands and small details. Hands are better than they were but still imperfect. Watches, rings, pens, glasses, dog leads — small handheld objects often look slightly wrong on close inspection. Earrings that don't match. Wedding rings on the wrong finger. Buttons that don't quite line up.

Light direction inconsistencies. AI models occasionally get light direction wrong across an image. A face lit from the left while the table in front of it is lit from the right. Most humans don't notice this consciously, but the image feels "off" in a way that is hard to articulate.

Plastic-perfect skin and lighting. AI portrait images often look slightly airbrushed. Pores too smooth, lighting too even, no shadows under chins or eyes. Real photography almost always has some imperfection somewhere.

Suspiciously perfect composition. Real candid photos are rarely perfectly composed. AI photos almost always are. If a "casual" photo has the subject in textbook rule-of-thirds with a beautifully blurred background, that's a possible signal.

The "too good to be a stock photo" test. If an image is exactly what you would design if you were generating it specifically to illustrate the article, it might be exactly that. Real photography has imperfection because reality is imperfect.

Video and audio are where AI generation has improved most dramatically in 2024-2026 and where the easy tells have closed fastest. A few remaining markers:

Video: blink rates, micro-expressions and hand-mouth coordination. AI-generated talking-head video has historically had subtly wrong blink rates (too regular, too infrequent, or both eyes blinking at exactly the same instant). 2026-generation models have largely fixed this on the obvious examples, but micro-expressions — the small involuntary face movements during emotional content — are still imperfect. People speaking in AI video often look slightly under-expressive, as if reading the lines.

Audio: breath pauses and lip-smack noises. AI-generated speech in 2026 is genuinely difficult to tell apart from human speech on first listen. The remaining tells are subtle: missing breath pauses where a human would naturally inhale; an absence of the small "lip-smack" mouth-noises that real speakers make between words; uniform pacing that doesn't speed up or slow down with emotional content.

Cloned voices and "vishing." A specific threat worth knowing: voice-cloning tools can now produce a believable 30-second clone of someone's voice from a 10-second sample. Scam calls ("hi mum, it's me, I'm in trouble, please send money") have got considerably more convincing. The standard defensive practice — agree a family safe word, call back on a known number, do not act on urgency-pressure — applies more than ever.

The general principle on video and audio: assume that a 30-second clip from a stranger that demands urgent action is potentially synthetic, regardless of how authentic it sounds. The technology now makes a convincing fake faster than humans can pattern-match a real one.

A small industry has sprung up around "AI detector" tools — GPTZero, Originality.AI, Turnitin's AI checker, Copyleaks, ZeroGPT, and a dozen others. Many of them advertise 99%+ accuracy. Treat that claim with extreme scepticism.

The core problem is structural. Detector tools work by analysing statistical properties of text — sentence length distribution, perplexity (how "surprising" each word choice is given the surrounding context), word-choice patterns. The trouble is that careful editors, non-native English writers, professional copywriters and any number of legitimate human writing styles can produce text with the same statistical signatures the detectors flag.

The well-documented failure modes:

• High false-positive rates on non-native English writing. Multiple studies have shown that text written by non-native English speakers is disproportionately flagged as AI-generated, even when entirely human-written, because non-native writing tends toward more measured, structured prose that the detectors confuse with AI patterns.
• High false-positive rates on plain, careful writing styles. A clear, factual explanation written by a competent human will often score as "likely AI" because clarity and simplicity look statistically similar to AI output.
• Easily defeated by light editing. A human can take an AI-generated draft, rewrite a quarter of the sentences, and most detectors lose the ability to identify it.
• No principled way to verify accuracy claims. Most detector marketing cites internal benchmarks against curated test sets. Real-world accuracy on mixed human+AI content is substantially worse than the marketing claims.

The practical advice: if you must use a detector tool, treat its output as one weak signal among several, never as a definitive answer. A high "AI probability" score on a piece of writing should prompt closer investigation, never an automatic conclusion.

The structural solution to "is this AI-generated" is to ask the AI itself to flag the content at the moment of generation. Two technical approaches are being rolled out:

Cryptographic watermarking. Some AI models embed a subtle, statistically-detectable pattern in their outputs that does not change the content meaningfully but can be detected by a verifier tool. Google's SynthID is the most prominent example — it watermarks images, audio and text from several Google models. The marketing claim is high reliability against the marked model's outputs. The limitation is obvious: it only works on content generated by participating models, and a determined adversary can usually disrupt the watermark with minor editing.

Content provenance metadata (C2PA). The Coalition for Content Provenance and Authenticity — a consortium including Adobe, Microsoft, the BBC, the New York Times and most major camera manufacturers — has developed an open standard for "content credentials" embedded in images, video and audio at the point of creation. The credentials note whether the content was captured by a camera, generated by AI, or edited subsequently, and they are cryptographically signed. By 2026, most professional cameras and several phone cameras stamp C2PA credentials by default; the standard is starting to appear in browser interfaces ("this image's content credentials say…").

Neither solution is universal yet. Watermarking only catches content from cooperating producers; C2PA only catches content from cooperating cameras. But the trajectory is clear — the long-term answer to "is this real" is going to look more like the lock icon next to a website URL (cryptographic provenance) than like a heuristic guess (statistical detection). For now, both are useful weak signals to add to your toolkit when they are available.

The compressed practical advice, after all of the above:

Trust the source, not the artefact. A claim is more credible if it comes from a publication, individual, or platform that has a track record of being right. The provenance is more reliable than any detection trick applied to the content itself.

Spot-check three facts. When something looks too good or too neat — an article confidently citing studies, a photo perfectly illustrating an unlikely event — pick three concrete claims and check them against Google. AI confabulation usually shows up within thirty seconds.

Notice when something has no human texture. Real writing carries the weight of a real person knowing real things. Real photography has imperfections. Real video has natural pacing. If a piece of content is perfectly fluent and perfectly composed and perfectly polished and somehow feels weightless, the question of "who made this and why" is more useful than "is this AI."

Be especially careful of urgent calls to action. Most AI-driven scams in 2026 — whether voice-clone phone calls, deepfake videos demanding urgent transfers, or text messages from "trusted contacts" — work by combining synthesised content with time pressure. The defensive answer is the same as before AI: slow down, verify through a known channel, call back on a number you already had.

Don't rely on detector tools alone. They will have you wrongly accusing your neighbour's GCSE-resitting daughter of AI-cheating on her coursework while the careful AI-assisted scam slips through. Use them as one weak signal in a bigger pattern, never as proof.

And, honestly: don't lose sleep over it. Most online content has been not-quite-trustworthy since long before AI got involved. Good old-fashioned scepticism — "who is telling me this, and why" — remains the most powerful tool you have.