As AI-generated content has improved, the obvious tells from a couple of years ago have mostly disappeared. What's left are softer, statistical patterns — useful as signals, not proof, and worth treating with appropriate humility.
Text: patterns worth noticing
- Overly balanced structure — AI writing tends toward a very even, "on one hand, on the other hand" structure, with lists and even paragraph lengths that feel slightly too tidy.
- Vague specificity — confident-sounding claims that, on closer look, don't cite anything concrete or checkable.
- Repetitive phrasing patterns — certain transitional phrases and sentence structures appear disproportionately often in AI-generated text compared to typical human writing.
- A flattened voice — missing the small inconsistencies, tangents, and personal quirks that show up in most genuine personal writing.
None of these are reliable proof on their own — human writers can produce tidy, balanced prose too, and AI detection tools built to flag text automatically have a real false-positive rate. Treat these as signals to look closer, not a verdict.
Images: what still gives it away
- Hands, teeth, and fine anatomical detail — still an occasional tell, though far less reliable than a couple of years ago as models have improved.
- Text within the image — garbled or nonsensical lettering on signs, labels, or clothing remains one of the more consistent tells.
- Physically implausible detail — reflections that don't match their source, shadows falling in inconsistent directions, background elements that don't quite make spatial sense.
- An unnaturally smooth or "too perfect" quality — especially in skin texture and lighting that doesn't behave the way a real camera and lens would produce it.
Metadata and reverse image search
For images, a reverse image search can sometimes reveal whether the same image (or a very similar one) already exists elsewhere, or trace it to an AI generation tool's gallery. Some platforms are also starting to attach visible or embedded labels to AI-generated content, though this isn't yet universal or reliable enough to depend on alone.
The honest limitation
Dedicated "AI detector" tools for text exist but have a meaningful error rate in both directions — flagging genuine human writing as AI-generated, and missing AI text that's been lightly edited by a person. Treat any single detector's verdict as one data point, not a conclusion.
Quick signals to check
- Text: overly even structure, vague specificity, a flattened personal voice.
- Images: garbled text, inconsistent reflections and shadows, unnaturally smooth detail.
- Use reverse image search to trace an image's origin where possible.
- Treat AI detection tools as one signal among several, not a verdict.