Checking AI Profile Photos Without Treating Them as Identity Proof

xiang li — Thu, 08 Oct 2026 02:45:01 +0000

I maintain Check AI Free. This revision corrects the earlier article's privacy and forensic-verification claims.

An AI-generated profile photo and a fraudulent account are different questions. A person can knowingly use an illustration or generated portrait. A camera photo can be stolen. A detection score cannot establish who controls an account, whether a candidate is qualified, or whether their story is true.

Start with the actual file

Use the original image when it is available and you have permission. A social profile screenshot is a new image with a different processing history. Keep its source page and surrounding context separately; do not treat the screenshot as a complete record of the original file.

The profile-photo guide helps organize image review. The tool observes editable source fields and can request a general pixel-model estimate when needed and available. It does not perform a dedicated iris, ear or identity check. It does not identify the exact model or verify C2PA signatures.

Keep uncertainty visible

Camera metadata can be copied or edited; absent fields can result from routine export. A completed model score can be wrong. A service failure is not evidence that a photo is real. Our current evaluation does not justify a validated overall accuracy claim.

In a hiring context, use a consistent, proportionate process. Ask for clarification when appropriate, compare independent information and give the person an opportunity to explain. The LinkedIn review checklist treats the image as one piece of context. Do not reject or accuse someone solely because a detector flags a portrait.

Understand image handling

The website checks embedded source clues locally first. It may then send the first 64 KiB of the file and, for a model check, a smaller redrawn metadata-free JPEG to the backend and its provider. A direct image URL is retrieved by our server. The general detector therefore does not keep every image byte on the device. See the current privacy explanation before selecting a sensitive image.

What our screenshot experiment actually shows

In an exploratory batch of eight licensed photographs, camera fields were present in the originals and absent from our specified metadata-free JPEG export and browser screenshots. The study gives the exact settings, sources and limits. It is a local field-retention experiment, not a study of social platforms, identity fraud or detector accuracy.

For a disputed image, seek the original and independent context. Read the method and limitations alongside any result.

Disclosure: I maintain the linked tool. This correction was prepared by an AI agent from the project's current implementation and experiment records.

AI Image Checks: Source Clues, Optional Models and Their Limits

xiang li — Mon, 05 Oct 2026 13:17:44 +0000

I maintain Check AI Free. This article corrects earlier descriptions of its detection and privacy boundaries. The tool does not authenticate an image, a person or an exact generator. It reads editable source clues and, on the website when needed and available, requests a model estimate.

What a source clue can tell you

An original image can contain camera fields, generation settings or source text. Those fields help explain the file's claimed history. They can also be removed, edited or copied. Observing a C2PA-related string is not the same as checking a signature, asset binding and certificate trust. The current tool does not perform those cryptographic checks or decode SynthID.

The initial image check inspects up to the first 64 KiB. A field beyond that slice can be missed. Missing camera fields do not establish AI generation, and camera fields do not prove an unaltered scene or a real account owner.

For saved creation settings, the existing Stable Diffusion guide also has a separate local PNG text reader. It reads bounded tEXt, iTXt and zTXt fields, including compressed text and text after image data. It can copy or download fields that exist. It cannot reconstruct a removed prompt or verify a workflow, and it never executes one.

What leaves the device

The website reads source clues in the browser first. An unresolved check can send the original file's first 64 KiB to the backend; that fragment can include metadata and pixel bytes. When a pixel-model check is needed and available, a redrawn, metadata-free JPEG with a longest edge of at most 1536 pixels is sent to the backend and its configured detection provider. Direct image links are fetched by our server.

This means a blanket claim that every check stays on the device would be inaccurate. The separate PNG text input stays local; the general detector has the network paths described above. The privacy page explains current processing, local extension history and provider boundaries.

Why an inconclusive result is useful

A completed model score is an estimate, not a calibrated probability or a certificate. Our available evaluation material does not support a validated overall accuracy percentage. Source-only results do not display an invented model percentage. A missing or unavailable model remains distinct from a completed low score.

Keep the original file, inspect the available evidence and seek context or an independent source when a decision matters. A tool can help organize that review; an uncertain result should stay uncertain.

A small field-retention experiment

We compared eight licensed photographic originals with two specified local transformations: a deliberately metadata-free JPEG export and an 800 × 600 browser screenshot. The independent reference reader found camera fields in all eight originals and none in the transformed versions. That is eight independent images and 24 condition files.

The full experiment reports settings, counts, source attribution, hashes and limitations. It uses one photographer and no generated-image sample. It measures field retention under those settings; it does not measure Instagram processing or AI-detection accuracy.

See the method and limitations. Controlled site demos illustrate metadata behavior and are not independent accuracy evidence.