How to Spot AI-Generated Text and Images

As AI-generated content has gotten better, the confident tricks people relied on a couple of years ago (perfect grammar, oddly generic phrasing) have gotten much less reliable. Here's what's actually still a useful signal, for both text and images.

For text: the signals that still hold up reasonably well

• Repetitive sentence structure — a subtle rhythm where sentences tend toward similar lengths and openings across a whole piece, lacking the natural variation of human writing.
• Overly balanced, hedge-everything phrasing — a tendency to present "on one hand, on the other hand" even where a human writer with a real opinion would just state it.
• Generic specificity — examples and details that sound concrete but are actually vague and interchangeable ("a busy professional named Sarah" appearing in one AI-generated example after another).
• An absence of genuine, specific personal detail — human writing tends to include small, oddly specific details that wouldn't occur to someone generating generic content.

None of these are reliable on their own — skilled editing of AI output, or a human writer with a very consistent style, can trip these signals too. Treat them as raised eyebrows, not proof.

For images: what to check first

• Hands, teeth and ears — still the most common source of small anatomical errors, especially at unusual angles.
• Background text — signs, labels and text on clothing in the background of an AI image frequently come out as garbled, nonsensical characters.
• Repeating patterns — fabric weaves, tiled floors, brick patterns often show subtle inconsistencies or unnatural repetition under close inspection.
• Lighting and shadow consistency — check whether shadows across the whole image agree on where the light source actually is.

Tools that claim to detect AI content — treat cautiously

AI-detection tools exist for both text and images, but independent testing has repeatedly found meaningful false-positive and false-negative rates, particularly on text that's been lightly edited after AI generation. Don't treat a detector's verdict as definitive, especially in any situation with real consequences, like an academic integrity case — corroborate with other evidence rather than relying on a single tool's score.

Why this is only going to get harder

Every signal listed above is a moving target — as detection methods improve, generation quality tends to close the gap not long after, particularly for high-effort, human-edited AI content. The most durable approach isn't chasing a foolproof detection trick; it's applying the same healthy skepticism to unusually convincing content that you'd apply to any unverified claim, especially when the content is emotionally charged or trying to sell you something.

Frequently Asked Questions

Are AI content detectors accurate?

Not reliably enough to be the sole basis for an important decision — independent studies consistently show meaningful error rates, especially on text that's been edited after being AI-generated. Use them as one weak signal among several, not a verdict.

Continue Reading

How to Free Up Storage on Your Phone Without Deleting Your Photos

Your storage is full but you don't want to delete memories. Here are the genuinely…