What is a deepfake? Separate generated media, altered identity and false context

Start with a description more precise than “fake”

NIST uses synthetic content as a broad category for material generated or substantially changed by algorithms, including AI. Its transparency report treats provenance tracking, watermarking and detection as different approaches. This guide uses a practical classification to help you choose the next check; it is not a legal definition or a claim that every edited image is a deepfake.

For identity substitution, ask whether an existing person’s face or voice was inserted into a performance. For a generated person or scene, ask whether the depiction came from a model rather than the claimed recording event. For an ordinary edit, identify the crop, retouch or composite. For false context, compare the caption’s place, date and event with the recording’s source. More than one category can apply to the same post.

Sources: National Institute of Standards and Technology: Reducing Risks Posed by Synthetic Content (NIST AI 100-4)

Use four fictional examples to select a check

Imagine a fictional mayor appearing in a clip. In version A, an actor performs and AI replaces the face: examine the production record and the claimed speaker’s original statement. In version B, an invented character reads a disclosed fictional script: keep the creator’s disclosure with the clip. Neither version establishes that a real official made the statement.

In version C, a genuine recording is cropped before the speaker says “that claim is incorrect.” Find the complete sentence. In version D, last year’s genuine footage is labeled as today’s event. Search an identifiable frame and compare the earlier publication. Google’s image-context feature can surface earlier uses or similar images, which are useful leads for that last question.

Sources: Google Search Help: Learn more about an image

Build a small evidence record

Use the following original worksheet for one claim and one exact file. Give each observation a source rather than turning several weak clues into a stronger-sounding verdict. NIST emphasizes that no single transparency method resolves every context; the worksheet keeps each check tied to the question it can answer.

1. Write the claim: who is supposedly doing what, where and when?
2. Keep the source URL and original file where available; distinguish a repost or screenshot from the original.
3. Name the suspected change: identity, generated scene, audio, ordinary edit or caption/context.
4. Choose evidence for that change: production statement, earlier recording, matching frame, saved image record or a supported signed credential.
5. Record what is established, what is contradicted and what remains untested. Do not promote an unavailable check to a negative result.

Sources: National Institute of Standards and Technology: Reducing Risks Posed by Synthetic Content (NIST AI 100-4)

Keep origin separate from truth

C2PA’s explainer distinguishes a file’s provenance from whether its depiction is factually true. A recorded editing action helps explain production; it does not resolve every statement in a caption. Likewise, a missing credential does not establish that a file was generated.

Suppose an authentic photograph shows a film set, and a repost calls it a real emergency. Your conclusion can identify the mismatched context without claiming the photograph is AI-generated. Conversely, a creator’s disclosed fictional AI scene does not need to be described as an undisclosed real event. Use wording that preserves the evidence you actually have.

Sources: Coalition for Content Provenance and Authenticity: C2PA and Content Credentials Explainer, version 2.2

Questions about this case