AI Content Detection
How detectors for AI-generated text, images, and video actually work — what they can tell you, what they cannot, and which tools are worth trying.
AI Text Detection
Text detectors estimate how predictable a passage is to a language model (perplexity) and how uniform that predictability is across sentences (burstiness). Human writing tends to vary more.
Limitations: Accuracy drops sharply on short passages, edited AI text, and writing by non-native English speakers. Treat any score as a signal, never as proof.
- GPTZero
- Originality.ai
- Copyleaks
AI Image Detection
Image detectors look for generator artefacts — inconsistent lighting and reflections, malformed hands and text, frequency-domain patterns — and increasingly for C2PA content credentials or invisible watermarks such as SynthID.
Limitations: Cropping, re-compression, and screenshots strip metadata and weaken watermarks, so a negative result does not mean an image is authentic.
- Hive AI Detector
- Illuminarty
- Content Credentials (C2PA) verify
AI Video & Deepfake Detection
Video detectors analyse temporal consistency: blink rate, lip-sync error against the audio track, face-boundary blending across frames, and heart-rate signals in skin tone.
Limitations: New generators appear faster than detectors are retrained, so detection quality varies wildly by model and by video quality.
- Deepware Scanner
- Sensity AI
- Reality Defender
A practical workflow
- Check provenance first — look for C2PA content credentials or original camera metadata before running any detector.
- Run at least two independent detectors; agreement between them is more informative than a single score.
- Use long samples. Under roughly 300 words, text detection is close to guesswork.
- Read the content critically: fabricated citations, invented statistics, and generic structure are stronger tells than any score.
- Never take action against a person on a detector score alone — ask for drafts, sources, or version history.