How to detect AI-generated text, images, and video.
Single detectors are wrong constantly — GPTZero flags Shakespeare, Hive misses obvious deepfakes. The only reliable signal is agreement across multiple detectors. Here's what to look for, which tools to trust, and how to triangulate a verdict.
What gives AI content away
Text — too smooth, too even
AI text has uniform sentence length, low burstiness, and avoids contractions. Run through GPTZero, Originality, and Copyleaks — agreement matters more than any single score.
Images — fingers, ears, jewellery, text
AI still botches high-frequency detail. Check fingers, teeth, ear shape, reflections, and any text inside the image. Hive and Sensity catch most diffusion outputs.
Video — micro-expressions and edges
Look for blink rate, mouth-shape mismatches, and warped edges around hair and glasses. Intel FakeCatcher and Reality Defender lead for video provenance.
Provenance metadata (C2PA, SynthID)
Major models now embed invisible provenance markers. Check Content Credentials before trusting any detector — provenance is more reliable than any heuristic.
One detector is a coin flip
Stanford found GPTZero misclassifies non-native English writing as AI 60%+ of the time. Never accuse on a single tool.
Triangulate the verdict
Run the same sample through 3+ detectors. If they disagree, treat the result as inconclusive — that's the honest answer.
How to detect responsibly
- 01
Check provenance first
Look for C2PA Content Credentials and SynthID watermarks. These are designed to be reliable; heuristics aren't.
- 02
Run multiple detectors
Use at least three independent detectors. Agreement is the signal — single tools are too noisy to act on.
- 03
Get a synthesised verdict
Triangulate runs your sample through multiple detectors and a judge AI synthesises the result with a confidence score.
AI detection FAQs
How accurate are AI text detectors?›
Individually, not very. GPTZero, Originality.ai, and Copyleaks all post 90%+ accuracy in their marketing but score 60–70% on independent benchmarks — and worse on non-native English writers. Never make a serious accusation on a single tool.
How do I detect AI-generated images?›
Combine three signals: provenance metadata (C2PA / Content Credentials), detector tools (Hive, Sensity, AI or Not), and manual checks on fingers, text, ears, and reflections. Agreement across all three is your verdict.
How do I detect AI-generated video?›
Look for unnatural blink rate, mouth-shape mismatches, warped edges around hair, and inconsistent lighting between frames. Intel FakeCatcher and Reality Defender are the leading dedicated tools.
Can ChatGPT detect AI-written text?›
Poorly. OpenAI shut down its own detector in 2023 because of low accuracy. Asking ChatGPT 'did you write this?' is unreliable — it confabulates either way.
Is there a free AI detector?›
GPTZero and ZeroGPT have free tiers for text; AI or Not is free for images. Triangulate combines multiple free detectors and synthesises one verdict so you don't have to copy-paste between sites.
Get the AI fusion playbook
Weekly: real prompts, real verdicts, real workflows. No fluff.
We never sell your email. Unsubscribe anytime.
Stop guessing. Triangulate it.
One sample. Multiple detectors. One honest verdict — with a confidence score you can actually defend.