TTG Plus > FAQs > More still on FAQ #200

Why not label “fake” images instead of “real” images?

  • 1. A “This is fake” label would depend for its success on compliance from the persons who are most inclined to deceive.

    (It is hard to picture a punishment that is both enforceable and strong enough to deter anyone bent on deception.)

    Furthermore, millions of image creators who do not mean to deceive would also have no interest in using a “This is fake” label.

    History suggests that those creators would likely resist telling viewers when an image that looks like an undoctored photo is anything but that.

    TTG never uses the term “fake” except in quotation marks

  • 2. There would be no legal entity that could enforce the use of a “This is fake” label except within a small subsection of the world.

    Images have no national borders, and citizens of most countries see many images from other countries.

    (As explained in #3 below, a disclosure label would be pointless unless it was universally used.)

    TTG never uses the term “fake” except in quotation marks

  • 3. If a “This is fake” label is not used by every image-creator worldwide who creates what the public calls “fake” images...

    . . . then any selective use of such a label would benefit those who create “fake” images but who don’t apply that label.

    The more people who used such a label, the more that non-users of the label would benefit.

    TTG never uses the term “fake” except in quotation marks

  • 4. If everyone relied on a “This is fake” label to alert them to possible deception, viewers wouldn’t know what to think when they saw an unlabeled photo.

    (This is also covered in FAQ #102)

    Consider the use of AI-generated content:

    Suppose a viewer lives in the U.S. — where the AI industry has pledged to watermark AI-generated images — and the viewer comes across an image that they strongly suspect contains AI-GC but it has no “This is AI” watermark.

    The absence of a “This is AI” watermark might mean the image is not AI-generated.

    But the absence of a “This is AI” watermark also could mean that the image is AI-generated but

    • it came from a source outside the U.S.; or

    • the watermark had been removed; or

    • the addition of AI-GC wasn’t extensive enough (e.g., it was only used for upsizing the image) to trigger a “This is AI” watermark, as explained in #6 below; or

    • the creator of the image was not inclined to attach a watermark.

    TTG never uses the term “fake” except in quotation marks

  • 5. Even before AI-GC came along, it would have been very difficult to reach a broad consensus on exactly which images warranted a “This is fake” label.

    Most advanced photographers could quickly point out multiple concerns with the French and Norwegian legislative initiatives referenced in the page linked above (the U.K. proposal has not yet been formally drafted).

    The problem is that all digital photos — including all news photos — undergo some kinds of “manipulations” (including some inside the device or camera, before the photographer reviews the image on the screen).

    It would be extremely difficult to identify which specific manipulations contribute to the social problems the laws intend to address.

    TTG never uses the term “fake” except in quotation marks

  • 6. AI-GC is used in too many different ways to be covered by a one-size-fits-all “This is AI” label...

    The myriad ways AI-GC is already being employed by millions of amateur and hobbyist photographers is perhaps the most misunderstood aspect of AI-GC among those who are clamoring for AI-GC-disclosure labels.

    Pledges by AI companies to add a watermark on plaigis will not address the capabilities involved in aigmenting photographs.

    A. Any AI-GC-disclosure label that covers full plaigis — i.e., images that are wholly created on a computer, without a camera — would likely not be adopted by millions of photographers who use AI-GC to aigment photos in what they regard as “minor” ways (for example, to fill in tiny threads in bird feathers when upsizing photographs).

    B. Requiring AI-GC-disclosure labels only on 100%- computer-generated plaigis would not produce the desired outcome of alerting the public to potential deception. It is easy to insert a small photographically recorded item into a plaigi, thereby exempting the image from any “full-plaigis-only” labeling requirement.

  • 7. . . . but creating different levels of a “This image contains AI” label would be forbiddingly complex.

    Formulating multiple varieties of disclosure labels to cover different levels of AI-GC would result in widespread confusion, for example:

    A. A label for plaigis

    B. A label for plaigis with an inserted photo-based element (see “B” in #6)

    C. A label for photo-based images that have been aigmented

    Any such variations would likely be ignored on a large scale.

  • 8. The implication behind the question at the top of this page is no longer accurate.

    Now that AI-generated images fill roles that photographs used to fill, few people still assume that photo-like images are “Real unless labeled fake.”

    See also #919 and 920.

    TTG never uses the term “fake” except in quotation marks

 

“Apart from regularly applied trust tests, there is no other viable solution for combatting the millions of deceptive AI-generated images that are coming.”