detectai.media
Plain-English explainers on how AI / deepfake detection works, how accurate each method really is, and where it breaks down.
Liveness detection catches the playback event, not the synthesis, so it works on noisy replayed deepfakes and degrades sharply on clean or injected ones.
Research measures high bypass rates against verification and detection systems, but every measured bypass leaves its own physical evidence in the file.
Four method families separate a live speaker from a replayed recording, and each one keys on a different part of the playback chain the replay cannot avoid.
Yes, at measured rates well above the marketing. One vendor advertises under one percent false positives; an independent benchmark measured 4.7 percent on human music. What a wrong flag costs a real musician, and why the advertised rate is not the field rate.
Mostly no. A cloned vocal over a real instrumental sits in the blind spot of both detector families: artifact detectors flag a synthetic region but not whose voice, and fingerprint matchers identify the song, never the singer.
One reads the vocal stem for a cloned or synthesized singer; the other reads the full mix for generator artifacts. They answer different questions, fail differently, and a pass on one says nothing about the other.
Depends on the edit. A two-semitone pitch shift or a 64 kbps re-encode can push a music detector near coin-flip, while time-stretching, EQ, and reverb barely move it. The split follows where the evidence lives in the signal.
Yes, and mostly by accident. A small pitch shift or a low-bitrate re-encode can flip a music detector to real, a generator it has never seen can slip past untouched, and deliberate attacks are barely studied. What that fragility means for trusting a score.
In the lab, the best 2026 systems report F1 scores above 0.99. On unseen generators, edited files, and broadcast audio, the same families fall to the sixties or to chance. Both numbers are real; the conditions decide which one you get.
Liveness detection asks whether a voice is live or a recording played back, reading the physical tells a loudspeaker and a second room leave behind.
Attribution tools can sometimes name the generator, but only inside a closed set of models they were trained on. On an unseen model accuracy collapses toward chance, and no tool names the human behind the prompt.
Two voice detectors can read the same clip and split, and usually neither is broken. They read different traces, trained on different generators, at different thresholds.
Music detectors are trained on a benchmark's generators and clean audio, and deployed on everything else. Off that ground, equal error rates near one percent climb past forty.
Two music detectors can read the same song and split, and usually neither is broken. They read different traces, trained on different generators, at different thresholds.
Image detectors are trained on curated benchmarks and deployed on everything else. Off that ground, accuracy near ninety percent falls toward a coin toss on unseen generators and compressed files.
A flag means a track resembled a detector's generated training examples. It often says more about the codec and the backing mix than about whether the vocal was cloned.
A flag means an image carried a trace that resembled a detector's generated training examples. It often says more about how the file was compressed and resized than about whether a machine made it.
A music-detector score is a similarity reading at an operating point you cannot see, reported as equal error rate on a benchmark, not the chance the track is AI.
An image-detector score is a similarity reading at a threshold you are not shown, calibrated on a benchmark rather than your photo, not the chance the picture is AI.
A SynthID video check finds Google's invisible Veo watermark, not AI-ness. Sora uses C2PA and Meta uses Video Seal, so a SynthID check returns not detected on them.
A SynthID check on a song finds one invisible Google watermark, not AI-ness. Suno and Udio carry no SynthID, so a check on them returns not detected.
An AI-music detector works well only on the generators it was trained on. On a known Suno or Udio version it is near-perfect; on an unseen model, or after a little processing, it falls to near chance.
A voice detector does not hear a fake. It measures faint traces the synthesis pipeline leaves in the waveform, then maps them to a probability at a chosen threshold.
A music detector does not hear a fake. It reads a fingerprint the generator leaves at fixed frequencies, then maps it to a probability. That makes it strong on known models and weak on everything else.
Often, yes. AI-music detectors read a fingerprint at fixed frequencies, and MP3 and other lossy codecs cut exactly there. A negative on a compressed track carries little information.
Yes, sharply. MP3 and phone codecs strip the fine high-frequency detail a voice detector reads, and it fails both ways: fakes slip through and real clips flag as AI.
Voice detectors are trained on curated benchmarks and deployed on everything else. Off the training distribution, equal error rates near 1 percent climb into the tens of percent.
A flag means a clip resembled a detector's synthetic training examples, and the research keeps finding the deciding feature is often the silence, codec, or noise around the voice rather than the voice itself.
A voice-detector score is a probability at an operating point you usually cannot see, calibrated on a benchmark rather than your clip, not the chance the voice is fake.
A detector is a useful first-pass screen against a cloned voice, not proof, and on the short compressed phone clip a scam gives you it is at its least reliable.
A SynthID check finds one invisible watermark from Google or OpenAI, not AI-ness. What a hit, a miss, and an OpenAI Verify result actually mean.
A detector gives you a probability with a known error rate, not a verdict, and in real conditions that error rate is far higher than the number on the vendor's page.
A detector can move you toward an answer but cannot settle it. Here is the checklist, where the score is one input and a confident number can still be wrong.
Real-time detection runs the same classifier as offline detection but under a latency budget, and for live calls it pairs a synthesis check with a liveness check.
Streaming detection works as a fast screen, but a latency budget forces short-window decisions that sit below offline benchmark numbers, and real call conditions lower the ceiling again.
No. A missing watermark or Content Credential is not evidence a human made an image. The marks are opt-in, easy to strip, and possible to forge.
Can AI images be detected after compression? Often not: JPEG and social-media resizing collapse the fingerprint detectors rely on, and it fails both ways.
Are AI image detectors reliable? They work on the generators and conditions they were trained on, and degrade sharply on an arbitrary image in the wild.
Real photos get flagged as AI-generated, especially after ordinary editing and compression. Why these false positives happen, and which tools produce the most.
Post-hoc AI detectors are unreliable, biased against identifiable innocent students, and using them on minors inverts due process. The vendors and major institutions already concede the score is not proof.
Self-checking your own writing before you submit it is rational, but the flag tracks predictability, not authorship, so the only way to clear it is to write worse.
Watermarking is the strongest AI-text detection there is, and still the wrong basis for an accusation: opt-in, cheaply scrubbed, and forgeable onto the innocent.
Sometimes, but only for cooperatively watermarked text. Captured media leaves an acquisition fingerprint; text leaves none, so post-hoc detection only guesses.
An AI detector score is a perplexity reading at an operating point you are not shown. It says the writing is statistically plain, not that a machine wrote it.
The false positive is structural, not a tuning bug: the only signal a post-hoc detector has is low perplexity, which whole classes of innocent writers share.
A calm, evidence-led guide to contesting a false AI-writing accusation: make non-use the more plausible account and put the burden back on the accuser.
The reliability ruling for AI text detectors: dependable where you own the watermark key or generation log, not reliable enough to carry any consequential decision.
Two AI detectors can read the same image and disagree. That is rarely a bug: they read different signals, trained on different models, at different thresholds.
AI image detectors do not see that a picture is fake. They read a statistical trace the generator left in the pixels, then compare it to what they learned.
It depends on what you are detecting and under what conditions. On a known generator, accuracy is very high; on an arbitrary real-world file it falls toward a coin toss.
No articles match — try another filter.