The short answer

There is no single best AI image detector for every file. Pixel scores estimate patterns, metadata records declarations, C2PA binds signed provenance claims to a file, and SynthID checks an embedded watermark from supporting tools. Start with the claim and the original file, then choose the checks that can answer your question. No single signal establishes whether a scene or caption is true.

This guide compares methods and when to use them; it is not a tested ranking of detector brands. For false positives and benchmark interpretation, read Are AI image detectors accurate?. For Image Evidence's own model status and launch criteria, see our evaluation method.

Four signals and the questions they answer

Choose a method by the evidence you need. These checks can complement one another, but agreement does not make them independent: several tools may be reading the same metadata field or using related models.

  • Pixel-based detector: does this file resemble the AI outputs represented in a model's training and evaluation? Useful for triage when provenance is missing; it cannot reconstruct the file's history.
  • EXIF, IPTC and XMP metadata: what does the file declare about its software, source or editing process? Inspectable clues, but unsigned fields can be edited or omitted.
  • C2PA Content Credentials: which provenance claims are signed, and do they validate against this file? Validation and trust in the signer are separate questions.
  • SynthID watermark: does an official detector find a watermark from a supported AI tool? Coverage is limited to participating tools and detectable signals.

1. Pixel-based AI detectors

A pixel model can flag an image even when it carries no metadata or credentials. That makes it useful for prioritizing human review of a large collection. It estimates resemblance to learned patterns; it does not observe who created the image.

Generalization needs testing. Ojha, Li and Lee found that classifiers trained on GAN images could miss images from newer diffusion and autoregressive models. Compression, crops and partial edits add further uncertainty. Ask what generators, real-photo categories and transformed files were evaluated before applying a score to your own collection.

Image Evidence's own on-device pixel model remains disabled after failing its evaluation gate. We make no claim of validated detection accuracy. The optional Sightengine score is a separate third-party service, requested by a signed-in account, and uploads the image for analysis. As the homepage explains, the first 3 accounts each day get 3 free, and Pro includes 100 a month. Daily free allowances reset at midnight UTC; availability also depends on network and site limits.

2. Metadata and generator parameters

Metadata can name software, declare an AI source type or contain generation parameters. Some Stable Diffusion interfaces write prompts and settings into PNG files. These are declarations to inspect, not a signature guaranteeing that the fields are accurate.

The AI image metadata checker guide explains what Image Evidence reads locally: EXIF, IPTC and XMP sections, declared AI source types and recognizable PNG parameter structures. The free evidence check runs in your browser without uploading the image. A missing field does not establish camera origin; an editor or sharing service may have removed it.

3. C2PA Content Credentials

C2PA defines signed provenance records that can describe capture, generation and edits. Successful validation checks the signed claims and their binding to the asset. It does not establish that a photographed event occurred as captioned or that every part of its history is available.

Use the C2PA checker to inspect an original file in your browser. Image Evidence reports a C2PA validation state but does not evaluate signer trust. Keep unverified or invalid records labeled that way. If trust matters, follow up with a validator that explains its certificate and trust-list decisions, such as the Content Authenticity Initiative's Verify tool.

4. SynthID and invisible watermarks

SynthID embeds a watermark in generated content rather than ordinary EXIF fields. It is designed to remain detectable through some edits, including cropping, filters and compression, but robustness has limits. Only participating tools embed it; no detection result covers every generator.

Google's public SynthID Detector at synthid.com checks supported media from Google and partners that have adopted SynthID. A found watermark supports AI-tool involvement, not a conclusion about the truth of a caption. A negative result cannot rule out AI generation.

Image Evidence has no SynthID detector integration and does not send files to Google. Our SynthID guide explains the official upload routes and their limits. Review the receiving service's current terms and coverage before sharing a file.

For the Google-specific workflow, see our Gemini AI image detector guide.

What screenshots, compression and re-exports change

Ask for the original export before comparing results across tools. A screenshot is a new file, while a social download may be a resized or recompressed copy. Record which copy you checked rather than assuming all copies carry the same evidence.

  • Metadata: screenshots and re-exports can omit the original fields. The new file may carry its own metadata, which does not describe the original creation process.
  • C2PA: an embedded manifest can be lost or its binding invalidated. Some systems can rediscover credentials through soft bindings or remote records; Image Evidence's local check does not retrieve those records.
  • SynthID: some transformations leave a detectable watermark, while heavier changes can weaken it. Do not assume a screenshot always removes the watermark or always preserves it.
  • Pixel scores: changed pixels can change the result in either direction. Do not interpret a score difference as a measured amount of AI editing.

Decision sequence: which check first?

Start with what you need to establish. A news caption, a disclosed illustration and a marketplace listing raise different questions. The following sequence helps you decide what to inspect and when to stop with an unresolved result.

  • 1. Record the claim and source. If you have only a screenshot or preview, request the original and note that limitation; you can still inspect the copy you have.
  • 2. Read metadata and C2PA locally. If you find a declaration or credential, record its exact meaning and validation state. If nothing is found, continue without calling the file authentic.
  • 3. If a supported watermarking tool may be involved, use an official verifier. Record its coverage and result, including any inconclusive response.
  • 4. Use a pixel score when it adds useful triage information and uploading is appropriate. Record the provider and tested file; a score cannot settle the case alone.
  • 5. Compare context and process evidence: reverse image search, the source's explanation, drafts or working files. Record contradictions and unknowns, and involve a person before a consequential decision.

A practical Image Evidence workflow

Open the free checker with the best available copy. The current picker accepts JPG, PNG, WebP, HEIC or AVIF up to 25 MB. It reads metadata, PNG parameters and Content Credentials in your browser, and can add one free AI likelihood score a day from the pixels. That score uploads the image once; it is not stored or used for training, and you can choose the label check alone to avoid any upload.

Read the report's validation states and limitations before requesting an optional detector score. That separate action sends the image to our server and on to Sightengine for analysis, then discards it. HEIC, AVIF and files over 10 MB are converted to JPEG in the browser for scoring, which can affect the score.

Supported formats and limits can change. If a file cannot be read, keep the original and try a PNG or JPEG re-export from the source app. Re-exporting may change or remove provenance, so identify the derivative in your notes and keep the original result separate.

For volume, Pro plans add batch checks and reports you can save as PDF or download as CSV and JSON. Exporting a report preserves the observations and limits you reviewed; it does not turn them into an authenticity certificate.

For generator-specific evidence and limits, see the Meta AI image detector and Grok AI image detector guides.

How to choose a tool without a misleading ranking

A useful comparison names the method, covered generators, file handling and evaluation evidence. A headline accuracy percentage cannot tell you whether the tool answers your question. Use our accuracy guide for benchmark details and the step-by-step image guide for visual checks and reverse image search.

  • For privacy: can you inspect metadata and credentials locally, and which optional steps upload the file?
  • For provenance: does the tool distinguish finding a manifest, validating it and trusting the signer?
  • For watermark checks: which tools and creation dates are covered, and what does a negative result exclude?
  • For pixel scoring: are false-positive rates, sample counts and generator coverage published for images like yours?
  • For review: can you record the result, limitations and unresolved questions in a form another person can assess?

What is the difference between C2PA and SynthID?

C2PA records signed provenance claims; SynthID embeds an invisible watermark in content from supporting AI tools. One describes a declared history and its integrity, while the other detects a supported watermark. A file can carry both, either or neither.

Does a missing C2PA or SynthID signal mean the image is real?

No. The creating tool may never have written that signal, or a later transformation may have removed or weakened it. A local C2PA check and an official SynthID check also have different coverage. Record what was not found without declaring the file camera-original.

Can I compare AI image evidence without uploading?

Image Evidence's free checker reads metadata and embedded C2PA records in your browser without uploading. It does not run SynthID or a pixel detector as part of that check. Requesting a Sightengine score uploads the image; using Google's SynthID Detector involves that service's separate upload flow.

Should teachers or hiring teams rely on one detector score?

A score can prompt further review, but it should not be the sole basis for an accusation, grade or hiring decision. Follow the relevant institution's policy, invite the creator's process evidence and document unresolved questions. A file report is evidence to interpret, not a verdict about a person.