A task type is one detection model. Every job runs exactly one task type, chosen with taskTypeId when you create the job. This page explains how to read results, which is the same for every task type, and lists all task types. Each task type also has its own page with its inputs, limits, result fields and artifacts.

Every scored detection result carries one score, fakeConfidencePercent, and a tag derived from it using the task type's boundaries. Use fakeConfidencePercent or tag for decisions; never threshold on confidencePercent.

Result fields

Task types that don't produce a single score, such as face-intelligence and the utilities, return their own fields instead. Each task type's page says whether its results are tagged.

Tags

Boundaries are inclusive at the upper edge: exactly 50.0 is likely_real. The defaults are 30 / 50 / 75, and a task type may use its own. Read the boundaries from tagBoundaries on each result rather than hard-coding them.

Example

"result": {
  "isFake": false,
  "confidencePercent": 89.6,
  "fakeConfidencePercent": 10.4,
  "realConfidencePercent": 89.6,
  "tag": "real",
  "tagBoundaries": { "real": 30, "likelyReal": 50, "likelyFake": 75 },
  "modelVersion": "v2.1.0"
}

Things to know

• Scores are specific to a model and its version. A score of 60 from one task type is not comparable with 60 from another.
• Re-validate custom cutoffs after a model update. When modelVersion changes, score distributions can shift. We give notice before model changes.
• Past results never change. The tag and boundaries are stored with each job when it runs, so retuning boundaries later doesn't alter earlier results. Jobs that ran before these fields were introduced don't have tag, tagBoundaries or modelVersion, so treat all three as optional.
• A score near 50 is not a coin flip. It means the model is uncertain. Treat the middle tags as review cases.
• Expect new fields. Ignore fields you don't recognize; new ones can be added without notice.

All 12 task types are available on the cloud API; your plan decides which you can use. Pass the ID as taskTypeId when creating a job, and filter job lists by slug with taskTypeSlug.

Every task type has a single input slot named original. On-prem deployments may expose a subset and use different IDs, so keep the slug as the identifier in your code and look up the ID per environment.

Choosing a task type

Each model is tuned for one kind of manipulation, and a real result only covers what that model checks. To screen for more than one kind, run one job per model, for example ai-image-detection and image-forgery-detection on the same photo.