No: an AI image detector's verdict alone does not prove how a particular portrait was made. Treat it as a result to investigate alongside the exact file, the tool's documented limitations and evidence of creation history. An unresolved result should remain unresolved. It is not a sound basis for publicly accusing the person pictured of deception.
This matters when a profile photo looks unusually polished, a colleague challenges a headshot, or you want to label your own work accurately. The useful question is not simply whether a website displays an AI badge. It is which claim you need to check, what evidence would support it, and whether you have that evidence.
1. Define what you mean by an AI photo
A completely generated person, a face swap, an AI-edited background and a camera portrait with an ordinary crop are different cases. A detector trained for one task may not document support for the others. Before interpreting its answer, write down your actual question: was the whole image generated, was a particular region replaced, or is the image being used with a misleading claim?
Keep origin, representation and context separate. A labelled synthetic illustration can be honest about its purpose. A camera photo can carry an inaccurate caption. A detector cannot establish that someone attended the event shown in a background, holds a qualification, or currently looks like a heavily altered headshot.
If you already know a portrait was generated and need to decide whether it represents you fairly, use our AI headshot likeness checklist. That is a different decision from investigating an unknown file's origin.
2. What detector research can actually tell you
NIST's 2024 report on synthetic content transparency describes image detection as dependent on the tool, evaluation data and image processing. Testing on unfamiliar generators is harder than testing on a generator represented in training; compression and resizing can affect results. The report also distinguishes evaluation metrics and the consequences of false positives and missed synthetic content. It is an overview, not a current accuracy certificate for a particular website.
The 2025 NeurIPS AIGIBench study evaluated 11 research detectors across different generators and testing conditions, including face swaps and personalized generation. Performance varied across sources and processing conditions. Its controlled retraining used selected object categories, and its tests were broader than everyday portraits. Its results therefore do not supply the probability that a current consumer service is right about your headshot.
A separate 2024 preprint, Fake or JPEG?, examined the GenImage dataset. Compression and image-size differences between real and generated classes could become shortcuts learned by the tested classifiers. Reducing those biases changed measured performance. This is a warning about evaluation design, not evidence that every detector merely recognizes JPEG files.
The practical consequence is to ask for evidence that matches your file and question. A strong benchmark headline is useful background, but it does not replace validation on relevant portraits and editing workflows. Equally, finding a limitation in a paper does not make every detector result worthless.
3. Read the label before reading the percentage
A false positive is a camera-origin image labelled synthetic under the test's definitions. A false negative is a synthetic image labelled real. Ask whether the documentation reports both, what kinds of images were tested, when the evaluation occurred, and how the decision threshold was chosen. If those details are absent, record the gap instead of filling it with assumptions.
Do not interchange a displayed score, advertised overall accuracy and the probability that one particular verdict is correct. Check how the service defines its number. A page that does not explain whether its percentage is a validated probability leaves that interpretation unsupported. Likewise, a real label is not a certificate of an unedited camera original.
Save the service name, date, model version if supplied, exact output and file used. Describe the result narrowly: this service labelled this file as AI-generated on this date. Avoid upgrading that statement to a conclusion about the subject's honesty, identity or character.
4. Provenance answers another part of the question
C2PA Content Credentials can provide signed, tamper-evident assertions about creation and editing. Check validation status, the signer and the actions actually recorded; a badge alone is insufficient. Participation is optional, history can be incomplete, and embedded metadata can be removed. Some implementations support rediscovery through durable credentials. Absence does not identify origin, while valid provenance alone does not establish that the depicted scene or its caption is true.
For example, a recorded editing step may help answer whether a background was generated. It does not automatically explain every pixel or establish the date of the scene. Read the available assertions as specific pieces of evidence, and keep unsupported parts of the history open.
5. A result-to-action table for portraits
| What you have | Reasonable interpretation | Next action |
|---|---|---|
| A detector says AI | A classification result requiring context; an error remains possible. | Keep the exact file and check the task, documentation and creation evidence. |
| A detector says real | No positive synthetic label from that test; other editing or origin questions remain. | Check the claim you actually need to establish before describing the portrait as unedited. |
| A validated credential records an AI action | Evidence of that recorded action within the available signed history. | Identify its scope and avoid assuming it describes every stage. |
| No credential is found | There is no available provenance record from this check. | Seek source information; leave origin unverified if evidence remains missing. |
| An earlier matching publication is found | A lead about prior use, not automatic proof of creation or a truthful caption. | Read the original context and check whether the files and claims actually match. |
6. Use this seven-step verification checklist
- Write the claim. Specify whether you are checking full generation, a face replacement, an edit or a caption. Decide what you would need to know before using the portrait.
- Preserve the evidence. Keep the original available file, page URL, caption and retrieval date. Do analysis on a copy. Do not confuse a messaging-app export with a file supplied directly by its creator.
- Check permission before uploading. For another person's portrait, consider whether you are authorized to send it to an external service. Review that service's handling terms. A verification task does not require publishing the person's face or your accusation.
- Check the tool's scope. Look for its supported tasks, evaluation date, relevant test images and error reporting. If it does not document the kind of image you have, state that limitation.
- Inspect independent history. Examine any verifiable credentials and available source material. A reverse-image search may identify earlier uses; follow the source rather than treating a thumbnail match as a verdict. Ask the creator for relevant context when appropriate.
- Record results without inflating them. Keep outputs and observations separate from conclusions. Conflicting outputs mean you have a disagreement to resolve, not permission to select whichever answer supports your suspicion.
- Choose a proportionate next step. Request a clearer source file, use another clearly documented image, label your own synthetic work appropriately, or leave origin unverified. For consequential allegations, seek qualified verification instead of making a public claim from an automated label.
Suppose your own camera portrait receives an AI label after you export it for a profile. Keep the original capture and editing history, and compare what each file actually represents. Do not announce that the tool has proved the portrait synthetic. Conversely, a real label on an unfamiliar headshot does not establish the pictured person's identity or the history claimed in its caption. These are hypothetical examples of decision-making, not measured error rates.
7. Keep verification separate from appearance feedback
For your own photos, reviewing known edits can be more useful than guessing origin from appearance. Our beauty filter photo audit helps you compare versions deliberately. If you need to share evidence containing other people, our photo privacy checklist explains why hiding a face may still leave identifying context.
FaceScore offers photo-based appearance feedback; it does not certify a portrait's origin or authenticity. Use your own image, or one you have permission to submit, for that stated purpose. Neither an appearance score nor a synthetic-image label measures a person's value. The useful outcome of verification is a defensible account of what you know, what supports it and what remains uncertain.