“Is this image real or AI?” sounds like a clear question. It is usually too narrow. A photograph can be genuine but years old, accurately dated but wrongly located, lightly edited, fully synthetic, or assembled from both captured and generated material. The caption can be false even when every pixel came from a camera.
Technical provenance is becoming more useful. Content Credentials can carry signed information about a file's origin and editing history, while invisible watermarks such as SynthID can signal supported AI generation. Current verification tools from Google and OpenAI can inspect some of those signals. Their coverage is not universal, and a missing signal is not proof that an image is camera-made.
The practical habit is to separate three questions: what does the file say about its history, where did the image first appear, and does the claimed event match independent evidence? This guide turns those questions into a short workflow for an ordinary sharing decision.
Write down the claim attached to the pixels
Before opening a detector, rewrite the post as a checkable claim. Include who or what is shown, the claimed place, the claimed time, the event, and the action the post asks readers to take. “Flood photo” is not enough. “This photo shows flooding on River Road beside Northside School at 7:30 a.m. today, so families should avoid the western entrance” exposes several facts that can be checked separately.
This step prevents a common category error. Image-generation detection addresses how media may have been produced. It does not establish that a road is closed, that a person is correctly identified, or that the image was captured today. OpenAI's current provenance guidance says its verification result does not confirm accuracy, legal ownership, editing status, or correct context. The C2PA explainer makes the same boundary explicit: provenance records history; it does not decide whether the depicted claim is true.
- File claim: what origin or editing history does the asset carry?
- Caption claim: what does the post say happened?
- Context claim: where and when is the post placing the image?
- Action claim: what does the post want the reader to do?
A storm image in the school group
At 7:40 a.m., Maya, a volunteer school communications coordinator, receives a forwarded image of water across a road. The message says the road beside the school is closed and asks parents to use another entrance. The image has no photographer credit, and the forward hides the original account.
Maya's decision is not to classify the image for a benchmark. It is to decide whether the school should repeat a safety claim. Speed matters, but an incorrect warning could redirect families toward another hazard or amplify a rumor. Her minimum evidence is therefore stronger than “the picture looks plausible.”
She saves the message URL and time, asks the sender for the original post or file, and checks the school's official channel and the relevant road or emergency authority. These actions can resolve the operational question even if no provenance signal survives in the forwarded image.
Preserve the strongest version you can get
Start with the original file when possible. A screenshot, social-media download, or messaging-app copy may have been resized, recompressed, stripped of metadata, or cropped away from useful clues. OpenAI notes that platforms, editing tools, file conversions, and sharing can remove C2PA metadata; watermark signals can also be degraded by compression, cropping, noise, or repeated edits.
Do not treat this as a reason to give up. Preserve what you have and record its path: sender, post URL, account name, timestamp, caption, and any edits you made for analysis. If you take a screenshot for reverse search, keep the received file as well. AFP's published verification method recommends obtaining original files when possible and archiving both the questionable claim and the evidence used to evaluate it.
The chain does not need courtroom formality for a school notice. It needs enough detail that another person can repeat the check and understand why one result may differ from another.
Read provenance as a receipt, not a verdict
A valid Content Credential can show that provenance information is cryptographically associated with a particular asset and has not been tampered with. Depending on what the issuer recorded, it may describe the creating tool, time, edits, ingredients, or AI use. The credential becomes more useful when you recognize and trust the signer and when the history covers the version you are examining.
A detected watermark also has a limited meaning. Google's Gemini verification help says a detected SynthID signal indicates that all or part of supported media was created or edited by Google AI. It also warns that Gemini currently recognizes Google AI content for that check; no detected Google signal does not rule out other AI systems. OpenAI's verifier similarly looks for supported OpenAI provenance signals and lists several reasons a generated file may return no signal.
Record the literal result. “Valid credential from X with these edits” is evidence. “No supported signal found” is an absence. Neither should be silently rewritten as “true photo” or “fake image.”
Ask the assistant to map evidence, not pronounce a verdict
A weak prompt is: “Is this image fake? Give me a confidence score.” It encourages a single label from visual style, even when the tool cannot inspect the original file, search the web, or query the relevant watermark system. A numerical confidence can make those missing capabilities harder to notice.
A stronger request is: “Help me verify the attached image and its caption. Do not decide from visual appearance alone. Separate your work into (1) available provenance metadata or supported watermark results, (2) reverse-search evidence about the earliest source, and (3) evidence for the claimed place, time, and event. State which checks you can actually perform, link every external result, distinguish confirmed, contradicted, and unknown, and do not convert ‘no signal found’ into ‘not AI-generated.’ End with a share, hold, or reject recommendation and the evidence that would change it.”
If the assistant cannot access a verifier or the open web, use it only to organize manual work. Ask it to produce search terms, a claim table, or a list of visible clues. Do not let it narrate tool results it did not obtain.
Run three verification lanes
The first lane is technical provenance. Inspect the original file with a Content Credentials-aware viewer and, when relevant, a provider's supported watermark verifier. Note the issuer, validation status, recorded actions, and coverage limits. Do not upload sensitive or private media without checking the verifier's data practices.
The second lane is source tracing. Run a reverse image search, inspect exact matches, and look for the earliest credible appearance. Search with crop variants when a post added borders or text. AFP notes that reverse search is not exhaustive: an image may be new, unindexed, flipped, or only loosely matched. A missing result is therefore not a clean bill of health.
The third lane is real-world context. Compare landmarks, signs, road layout, weather, shadows, language, uniforms, and other testable details with maps, official notices, reliable local reporting, and other images from the event. Contact the creator or named organization when the stakes justify it. For Maya, the official road notice matters more than whether the image contains an odd reflection.
- Provenance lane: What signed or watermarked history is detectable in this file?
- Source lane: Where and with what caption did this image appear earlier?
- Context lane: Does independent evidence support this place, time, event, and requested action?
Five results people overstate
Verification errors often come from saying more than the evidence permits. The correction is usually a narrower sentence, not another detector.
- No credential found becomes “camera photo.” Metadata may be absent, removed, unsupported, or attached to another version.
- A valid credential becomes “the scene is true.” A credential can authenticate recorded history without proving the real-world caption.
- An AI watermark becomes “every pixel is synthetic.” The result may cover all or only part of supported edited media.
- A strange hand or shadow becomes “definitely AI.” Visual artifacts are clues for follow-up, not reliable provenance by themselves.
- A reverse-search match becomes “original source.” Search results can be incomplete, later reposts can rank first, and similar images can be mistaken for exact matches.
Close with a share, hold, or reject note
Maya writes a four-line decision record: the exact claim, the strongest supporting evidence, the strongest contradiction or gap, and the action. She shares only when the operational claim is corroborated by an authoritative current source and the image evidence does not conflict. She holds when the road status remains unknown. She rejects the post when an earlier source, mismatched location, or official update contradicts it.
The note should keep provenance and truth separate. A useful outcome might read: “Hold. No supported provenance signal was found in the forwarded screenshot, which is inconclusive. Reverse search found no exact indexed match. The council road page does not list this closure as of 7:52 a.m.; waiting for school confirmation.” That is less dramatic than “AI fake,” but it is more actionable and defensible.
The durable lesson is simple: verify the claim attached to the image, not only the image's visual style. Provenance signals can strengthen the check. Source tracing and real-world corroboration finish the job.
Continue learning on JoyfulGrid
Frequently asked questions
Does no Content Credential mean an image is real?
No. The file may predate adoption, come from an unsupported tool or export path, or have metadata removed during sharing, editing, conversion, or screenshotting.
Does a valid Content Credential prove the caption is true?
No. It can provide tamper-evident provenance about the asset and its recorded history. You must still verify the claimed person, place, time, event, and context.
Can an AI chatbot identify a generated image by looking at it?
It may notice clues, but visual judgment alone is not a dependable verdict. Use supported provenance checks, reverse search, original-source tracing, and independent context evidence.
What should I do when the evidence stays inconclusive?
Do not force a binary label. Mark the claim unknown, avoid amplifying it, and state what additional evidence—such as the original file, creator confirmation, or an official notice—would resolve the decision.
Sources
- Verify AI-generated images, videos, and audioGoogle Gemini Apps Help
Checked on August 7, 2026 for current Gemini verification access, SynthID and Content Credentials behavior, file limits, negative-result caveats, and recommended complementary checks.
- Provenance signals (Content Credentials, SynthID) in OpenAI-generated contentOpenAI Help Center
Checked on August 7, 2026 for current OpenAI image provenance coverage, verifier behavior, context limits, and reasons a supported signal may be missing.
- C2PA and Content Credentials Explainer, version 2.3Coalition for Content Provenance and Authenticity
Used for the distinction between cryptographically verifiable provenance, signer trust, asset association, value judgments, and factual truth.
- About Content CredentialsContent Credentials
Used for the public explanation of embedded provenance, invisible watermarking, digital fingerprinting, and recorded creation and editing history.
- How we workAFP Fact Check
Used for established verification practice: preserve original files, reverse-search images, compare contextual clues, contact sources, archive evidence, and avoid relying on an AI detector alone.