AI image detector: review an image’s possible AI origin

An image-origin review is stronger when it distinguishes recorded history, model inference and visual clues. These are evidence categories to evaluate, not a promise that this page supplies a C2PA validator, classifier or forensic-map viewer.

• Provenance: Content Credentials can record signed information about creation and edits. Validation concerns integrity and the signer; it does not prove that the depicted event is true. Missing credentials are not proof of AI generation.
• Pixel-based inference: a trained classifier can look for patterns associated with its training examples. Its output depends on the model and input, and unfamiliar generators or edits can affect the assessment.
• Forensic context: compression, noise and camera metadata can inform a specialist review. An unusual pattern or missing EXIF field alone does not establish that an image was generated.

An image might be associated with many tools and workflows. Do not interpret this list as a tested compatibility matrix or a promise to identify a specific generator. Save a creator’s disclosure and any available original file.

• Midjourney imagery: seek the original source or creator disclosure rather than judging by a familiar visual style.
• DALL·E imagery: a missing label in a copied or converted file does not establish a camera origin.
• Stable Diffusion output: custom models and post-processing can change the clues available to a reviewer.
• Flux output: do not assume a classifier has been evaluated against every model release or derivative.
• Gemini or Imagen imagery: preserve any origin disclosure and compare it with the actual file being reviewed.
• Ideogram output: readable lettering is not, by itself, a reliable way to rule AI generation in or out.
• Adobe Firefly imagery: look for recorded workflow information where available, while separating provenance from scene truth.
• ComfyUI workflows: a workflow can mix models, edits and source photos, making a simple generator label incomplete.
• GAN-generated faces: a synthetic-looking portrait does not establish a real person’s identity or prove impersonation.

The right follow-up depends on what the image asks you to believe or do. A detector score cannot replace source verification.

• Profile pictures and avatars: review account context and other verification signals; a face score is not identity verification.
• Social posts: find the earliest credible source before reposting an image or repeating its caption as fact.
• News images: verify where, when and by whom the photograph was supplied. A real capture can still have a misleading caption.
• Marketplace photos: request additional product views or independent evidence when a listing seems inconsistent.
• Receipts and payment screenshots: confirm the transaction in the actual payment record, rather than approving a transfer from image appearance.
• IDs and scanned documents: use an appropriate verification process. An image-origin assessment does not authenticate a document or its holder.
• Art and illustration: ask for a disclosed workflow or source material. Appearance alone cannot reliably establish authorship.
• Edited photographs: distinguish an original capture, a small AI-assisted edit and a fully synthesized scene; one label may not describe the whole history.
• Forwarded screenshots: ask for the original file and source context, since repeated captures remove useful information.

Use analysis to prioritize further review, with a record of uncertainty and a way to obtain independent evidence.

• Trust and safety reviewers: apply a consistent review process and allow evidence beyond a single automated assessment.
• Fraud and claims reviewers: corroborate a receipt or claim with underlying records before a payment decision.
• Onboarding teams: keep image-origin screening separate from a validated identity-verification process.
• Marketplace teams: compare listing imagery with seller history, product evidence and delivery records.
• Journalists and fact-checkers: keep the source chain, contact the supplier and document what remains unverified.
• Teachers and reviewers: do not make an accusation or academic penalty from a detector output alone; review the work and disclosed process.
• Designers and art directors: obtain appropriate workflow and licensing information from the supplier.
• Legal and compliance reviewers: preserve originals and use qualified evidence procedures; this page does not certify admissibility or authenticity.
• Everyday readers: pause before sharing a dramatic image and check an independent source for the claimed event.

A careful review should explain its evidence and its limits instead of treating a number as a verdict.

• Understand file handling: choosing a photo here creates a local preview. That alone does not establish that every possible analysis runs locally.
• Ask what supports a conclusion: distinguish credentials, model output and a visual observation. This page does not claim to display forensic maps or a validated metadata report.
• Allow uncertainty: a compressed, cropped or unfamiliar image may not support a useful conclusion. An unavailable assessment must not become a guessed score.
• Keep detection separate from editing: choosing an image needs no registration, and image-origin review does not consume photo-editing credits. This is not a free paid-generation trial.

• A local preview is not a complete on-device classifier. This page does not promise a model download, local-only analysis or offline operation.
• An assessment is a supporting signal, not proof of truth, authorship, fraud or deception. Do not accuse someone or make a consequential decision from a score alone.
• Small crops, repeated screenshots and strong compression can remove clues or introduce misleading artifacts.
• AI-assisted cleanup, heavy retouching and full synthesis are different processes. A single score cannot reliably describe every edit.
• The file picker accepts supported still images, not video, audio or PDFs. Export a relevant still image separately if appropriate.
• This is not face recognition, reverse image search or a means of establishing someone’s identity or an image’s complete source history.