Skip to content
LikelyAILikelyAI
Menu
Sign inGet API key

One API to estimate whether content was made by AI.

Calibrated probabilities, a separate confidence score, and versioned results you can store, explain, and re-evaluate a year from now.

Start buildingRead the docs

500 free credits. No card required. Your content never trains our models.

A real request, running now

campaign-04.png, 2400 × 1600, 312 ms

AI probability 87%, classified as likely ai.

Likely AI
Confidence
0.74, moderate
Signals
3 evaluated
Model
image-2026-08-1
Warnings
None
Request
curl https://api.likelyai.dev/v1/detect \
  -H "Authorization: Bearer $LIKELYAI_KEY" \
  -H "Idempotency-Key: req_01J8..." \
  -F "[email protected]"

Response

200 OK
id
det_01JQ8F3ZK4
ai_probability
0.87
confidence
0.74
classification
likely_ai
signals[0].model_ensemble
0.91
signals[1].content_credentials
not_found
warnings
[]

Three modalities, one response shape.

The same fields come back whether you send a paragraph, a photograph, or a voice note. The audio example below deliberately shows low confidence: that is the product working, not failing.

  • Text

    “The keyboard has a reassuring weight and the switches feel crisp, though the companion app still needs work.”

    18%Likely human

    Confidence high

  • Image

    campaign-04.png · 2400 × 1600 · 3.8 MB

    87%Likely AI

    Confidence moderate

  • Audio

    support-call.wav · 42 s · 48 kHz · heavy compression

    61%Uncertain

    Confidence low

A probability is not a verdict.

LikelyAI separates the estimate from how well supported it is. That distinction is what lets your product respond sensibly to weak, transformed, or unfamiliar input, instead of acting on a confident looking number that is not.

ai_probability
How likely the submitted content is to be AI-generated or materially modified.
confidence
How reliable that estimate is for this specific sample and its quality. Never the same number.
signals[]
Each piece of evidence, scored on its own. A missing credential is reported as absent, never as evidence of a human author.

response.json

{
  "id": "det_01JQ8F3ZK4",
  "status": "completed",
  "modality": "image",
  "ai_probability": 0.87,
  "confidence": 0.74,
  "classification": "likely_ai",
  "signals": [
    { "name": "model_ensemble", "score": 0.91 },
    { "name": "content_credentials", "status": "not_found" }
  ],
  "model_version": "image-2026-08-1",
  "warnings": []
}

What we will and will not claim.

There is no accuracy number on this page. A single accuracy figure without a dataset behind it is not a claim, it is a decoration.

  1. Probabilistic, not definitive

    We report likelihood and our own uncertainty. We never assert that authorship is proven.

  2. Versioned model output

    Detector version, calibration version, and timestamp travel with every result.

  3. Explicit warnings for weak input

    Short, corrupted, or heavily transformed samples lower confidence rather than inventing precision.

  4. Provenance checked separately

    Content Credentials appear as their own signal. Missing metadata is reported as missing, nothing more.

  5. No permanent media retention

    Uploads are deleted on a short, configurable window. Your content never trains our models.

LikelyAI

Build on evidence, not assumptions.

Get your API key and make your first detection in minutes.

Get your API key