🏛️ Issues and Advocacy Training
Courses › AI Image, Video and Voice Generation

Lesson 4 of 9 · about 16 minutes

AI Images in News and Advocacy: Illustration versus Evidence

You'll be able to

  • Sort an image use into illustration or evidence and know the rule for each
  • Understand why a label in the pixels survives sharing when a caption does not
  • Explain what C2PA Content Credentials and invisible watermarks like SynthID record and where they fail
  • Check an incoming image for provenance signals before sharing or relying on it

People forgive a plain picture. They do not forgive being fooled. Every rule in this lesson comes from that, and it holds for newsrooms, campaigns, advocacy groups and agencies alike.

Illustration or evidence

Every image attached to a public argument is doing one of two jobs. Evidence says: this happened, this is what it looked like. Illustration says: here is an idea, drawn. AI images can do the second job. They can never do the first.

UseJobAI acceptable?
Diagram of how a bond is repaid over 20 yearsIllustrationYes, labeled
Header art for an opinion column on library hoursIllustrationYes, labeled
"What the proposed warehouse could look like from Route 9"Illustration of a scenarioOnly with a heavy label, and better built from the developer's actual plans
The flooded intersection last TuesdayEvidenceNo. A real photo or no photo.
The mayor at the meetingEvidenceNo. A real photo, or a plain name card.
A real person doing anythingEvidence by implicationNo, ever

The NPPA Code of Ethics for photojournalists says not to manipulate images in ways that mislead viewers. The SPJ Code says to label illustrations and reenactments. Neither was written with diffusion models in mind. Both apply squarely.

Never depict a real, identifiable person doing or saying something they did not do. Not for satire, not as a "what if," not with a label. A label does not stop a screenshot.

Why real people are off limits

The problems stack up. It is false, which breaks the basic promise of honest public argument. It can be defamatory, which is covered in Fairness and defamation. And in an election, Idaho's synthetic media law, Idaho Code 67-6628A, lets a candidate sue over deceptive synthetic media; it covers audio and video recordings, not still images, so a fake photo falls back on defamation law and platform rules (lesson 7 covers the statute). Platforms ban much of it too: TikTok bars synthetic likenesses of private people outright.

Advocates are tempted more than reporters, because a vivid picture of an opponent behaving badly feels persuasive. It is persuasive right up to the moment someone shows it is fake. Then the fake becomes the topic, and the actual argument (the water rates really did go up 38 percent) gets buried under it.

Public officials can still be shown: in real photos the user has rights to, or in a neutral treatment: a name and title on a plain card, a silhouette clearly marked as generic, or a drawing of an empty chair with a nameplate.

Labels that travel

An image escapes its original page within minutes. Somebody screenshots it, crops it, posts it with their own caption. Labels that survive that trip sit in more than one place.

  1. In the pixels. Small text in a corner: "AI illustration." Readable at phone size. On a 1080-pixel-wide image, around 28 to 32 pixels tall is readable without shouting. Inside the safe zone, a casual crop will not cut it.
  2. In the caption or credit. "Illustration created with AI (Google Gemini) by Dana Ruiz from a prompt describing a patched county road. It does not depict a real location."
  3. In the post text when shared on social media, and in the platform's own AI label if it has one (lesson 7).

Weak and strong: a caption

Weak: Image: AI

Strong: "AI-generated illustration (ImageStudio) showing a generic two-lane road. Not a photograph of any real road in Harlan Valley."

The strong version names the method, says what it shows, and says what it does not show. That third part is what stops misuse.

Content Credentials: the nutrition label

The Coalition for Content Provenance and Authenticity (C2PA) publishes an open standard called Content Credentials. Its steering committee includes Adobe, Amazon, the BBC, Google, Meta, Microsoft, OpenAI, Sony, TikTok and Truepic. C2PA describes the credentials as a nutrition label for digital content.

How it works, from the C2PA specification explainer:

image pixels manifest (Content Credential) assertions: made with AI tool, edits, date hard binding: hash of the pixels signature: signer's X.509 certificate Change the pixels and the hash no longer matches. Strip the manifest and it is simply gone.
A manifest is tamper-evident, not tamper-proof. It can be removed whole.
  • A manifest holds assertions: statements about origin, edits and AI use.
  • It is digitally signed by the software or device that produced it.
  • A hash ties it to the exact pixels. Edit the pixels without updating the manifest and the mismatch shows.
  • The manifest can be stripped from the file. Then it is gone, unless a "soft binding" such as an invisible watermark or a fingerprint lets a service look it up again.

Adobe Firefly attaches Content Credentials to certain exports. OpenAI says images from ChatGPT and its API carry C2PA metadata and a SynthID watermark, and it warns the metadata can be removed by platforms, editors or file conversions.

Note a practical clash: ImageStudio resizes, crops and strips all metadata, which protects privacy and also removes any C2PA data a provider embedded. So with ImageStudio output, the visible label is the only label.

Invisible watermarks: SynthID

SynthID is Google DeepMind's invisible watermark. It goes into images, video, audio and text from Google's models, and Google says it is designed to survive cropping, filters, frame-rate changes and lossy compression. Google offers a SynthID Detector portal, and Gemini will answer whether Google AI made or altered an image. Google says the detector also covers content from partners including OpenAI and NVIDIA.

Limits, plainly:

  • A detector only finds its own watermark. Images from a tool that does not use SynthID come back clean.
  • OpenAI's own help page says compression, cropping or format conversion can degrade a watermark.
  • A watermark tells which system made something. It says nothing about whether the content is true.

Common mix-ups

  • Claim: No Content Credentials means it is a real photo. What the evidence shows: Most images carry none, and a screenshot removes them. Absence proves nothing.
  • Claim: AI detectors can reliably spot any fake. What the evidence shows: Watermark detectors only find their own marks. General "AI or not" classifiers make errors both ways. They are one clue, never the verdict.
  • Claim: A stylized image needs no label. What the evidence shows: Style helps, but a label costs nothing and some styles look photographic on a small screen.

Reporting angle

When a striking image arrives from a reader, a campaign or a social post, before you run it:

  • Ask for the original file and how it was made. Real photographers can usually send the camera original.
  • Check it with the free Content Credentials inspector at contentcredentials.org and with Google's SynthID tools. Record what you found, including "nothing found."
  • Run a reverse image search to see whether it appeared earlier, elsewhere, in a different context.
  • Look for the classic tells (garbled signs, odd hands, shadows pointing different ways), knowing the best 2026 models have fewer of them.
  • If a local campaign circulates an image, ask the campaign directly whether AI was used. Their answer, or refusal, is reportable.

Research angle

Check yourself

1. Which is an acceptable use of an AI image in material about a public issue?

2. OpenAI's help center says its images carry C2PA metadata. What does it also say about that metadata?

3. An image has no Content Credentials. What does that show?

4. ImageStudio strips all metadata from finished files. What does that mean for labeling?

5. Which disclosure placement survives when an AI illustration is posted to social media on its own?

6. What does SynthID detection tell you?

Terms in this lesson

Photo illustration
An image built or altered to express an idea, labeled so viewers do not take it as a documentary photo.
C2PA
The Coalition for Content Provenance and Authenticity, which publishes the open standard for signed provenance data called Content Credentials.
Content Credentials
Signed, tamper-evident metadata attached to a file that records how it was made and edited, including AI use.
SynthID
Google DeepMind's invisible watermark for AI-generated images, video, audio and text, checkable with Google's detector.
Provenance
Where a piece of media came from and what has been done to it since.

Full glossary

Sources