As tools like ChatGPT and Midjourney get better, telling real from synthetic gets harder. Here are the practical tells that still hold — and an honest warning about "detectors."
Spotting AI-written text
No single sign is proof, but these patterns cluster in AI writing:
- Generic and hedged — lots of "it's important to note," "in today's world," "in conclusion."
- Even, tireless tone — no rough edges, no strong opinions, no personal specifics.
- Repetitive structure — every paragraph the same shape and length.
- Confident vagueness — says a lot without concrete detail, names or numbers.
- Slightly-off facts — plausible but subtly wrong (a hallucination tell).
Spotting AI-generated images
- Hands, teeth and ears — still the classic giveaways (extra fingers, odd merges).
- Text in the image — often garbled (unless made with a text-specialist tool).
- Background logic — melting objects, impossible reflections, warped patterns.
- Too perfect — flawless skin, unnaturally symmetrical, "plastic" lighting.
- Metadata & context — check where it came from; reverse-image-search it.
Spotting AI video & deepfakes
- Lip-sync and blinking that feel slightly off.
- Edges and hair flickering around the face.
- Emotion mismatch — voice tone not matching expression.
- Verify the source — for anything important, confirm through a second, trusted channel.
Why "AI detectors" aren't reliable
Detector tools give a confidence score, not proof — and they produce false positives (flagging human writing, especially from non-native speakers) and false negatives (missing lightly-edited AI). Never make a high-stakes decision (grading, hiring, accusations) on a detector alone.
The honest takeaway
The tells are real but shrinking. The durable defense isn't spotting pixels — it's verifying sources and context. Treat unsourced, urgent or too-perfect content with healthy skepticism, and confirm anything that matters.