deepfake
Enable an AI agent to recognise suspected or established deepfake media, assess what its apparent authenticity conceals, and determine justified handling actions.
Research draft, second pass
A second pass drafted this model: the structure a model of this thing needs, and what is known about it in the world. The line under this one says how the second half was obtained - researched against sources, or recalled without web access, in which case nothing here was read anywhere and every claim is a lead to verify. Unreviewed either way.
Researched by: Codex + Grok
Purpose and description
Enable an AI agent to recognise suspected or established deepfake media, assess what its apparent authenticity conceals, and determine justified handling actions.
A deepfake is a synthetic audiovisual or still-image artefact in which a generative model has replaced, transplanted, or invented a person's appearance or voice so that the result is presented, or readily taken, as a recording of a real event.
It can be Mark suspected synthetic regions and preserve the evidence underlying that assessment.; Compare a represented utterance, action or identity with authenticated reference material.; Trace an encountered copy to known source material and related versions without assuming missing lineage proves fabrication.; Assess whether recorded consent and applicable external rules cover a proposed use.; Attach or repair accessible synthetic-media disclosure on an authorised derivative.; Route the artifact for review, restrict a proposed use or preserve it as evidence according to recorded authority and uncertainty..
Distinguishing features
Identify evidence of AI generation or alteration affecting the represented identity, speech, action or event; a misleading caption or ordinary splice alone does not establish a deepfake.
State exactly what a viewer or listener might mistake for an authentic recording, distinguishing that representation from clearly stylised or nonrepresentational generation.
Separate fabrication of the depicted content from misdescription of authentic content: a genuine recording with a false date is a neighbouring case unless the media itself also meets the model's criteria.
Test whether a suspected voice or face substitution changes apparent attribution, rather than treating every AI-assisted restoration, translation or cosmetic adjustment as a deepfake.
Record disclosure and consent independently of technical classification: an authorised, labelled impersonation can still fall within this model.
Scope
+ Suspected and established deepfake artifacts, including identifiable segments within mixed authentic and synthetic media
+ The identities, voices, likenesses, actions and events the artifact appears to represent
+ Known or inferred generation and alteration operations, input materials and derivative versions
+ Evidence for authenticity, manipulation and attribution, with uncertainty attached to each conclusion
+ Subject consent, disclosure, dissemination context and artifact-specific handling decisions
- General-purpose generative models, training datasets and model-development processes
- Complete identity records for depicted people or organisations
- Disinformation campaigns, influence networks and coordinated distribution operations
- Non-AI deceptive editing, misleading captions and authentic recordings presented out of context
- Synthetic media that makes no apparently authentic representation of a person, utterance, action or event
- General legal rules, platform policies and adjudication of disputes
Characteristics
- Deepfake classification
- suspected | supported | contested | ruled-out | unresolved, with evidence and assessment date Prevents suspicion or an automated flag from becoming an established authenticity judgment.
- Affected modality
- image | audio | video | multimodal Determines which parts require examination and which disclosure methods are usable.
- Alteration operation
- face replacement | facial reenactment | voice synthesis or conversion | lip synchronisation | scene synthesis | mixed | unknown Distinguishes the fabricated representation from unaffected material.
- Synthetic extent
- Affected time intervals in seconds, image regions in pixels or normalised coordinates, audio channels, or whole-artifact designation; unknown permitted Supports precise review when only part of a recording is altered.
- Represented subject
- Links to claimed or assessed people, personas or events, with attribution basis and uncertainty Separates apparent identity from verified identity and supports subject-specific decisions.
- Authenticity proposition
- Explicit statement of the identity, utterance, action or event presented as authentic Makes clear what the evidence supports or contradicts.
- Evidence strength
- unassessed | limited | corroborated | conflicting, assessed separately for each proposition Avoids a single confidence label concealing different uncertainties about synthesis, identity and meaning.
- Consent coverage
- unknown | documented-within-scope | documented-outside-scope | disputed | explicitly-refused Connects recorded permission to the actual identity use, audience and distribution context.
- Disclosure visibility
- absent | metadata-only | perceptible-in-media | accompanying-context-only | mixed | unknown Shows whether recipients of the specific copy can encounter the disclosure.
- Artifact lineage
- Links to input recordings, generated masters, edits and circulated copies, with verified or asserted relationship Helps determine which findings survive cropping, recompression and other transformations.
Also called
Where this came from
wikidata · CC0 1.0
Drafted structure
Bundle to layer to finding to question, as the second pass will find it: 6 bundles · 11 layers · 11 findings · 21 questions.
Fabricated representation What the media appears to show or sound like, and what part of that representation may be synthetic.
An agent needs a precise authenticity proposition before it can distinguish a deepfake from unrelated editing or misleading context.
Apparent reality
The identity, utterance, action or event that the artifact invites a recipient to accept.
Authenticity target
Record the specific representation at issue without adopting it as fact.
- Who or what does the artifact appear to depict, and which words, actions or circumstances are attributed to that subject? definition
- What makes this appear to be an authentic representation rather than an overtly fictional or stylised depiction? boundary
Altered content
The location and semantic effect of suspected or established synthesis.
Synthetic contribution
Separate the suspected AI contribution from authentic source material and ordinary edits.
- Which frames, image regions, audio intervals or channels are believed to be generated or altered, and how precisely are they located? measurement
- Does the alteration fabricate identity, speech, action or circumstances, or merely improve presentation without changing the authenticity proposition? boundary
Synthesis and lineage How the artifact was produced and how the examined copy relates to inputs and other versions.
Deepfake assessment depends on separating evidenced production history from guesses about a generator or editing technique.
Production operation
Known or asserted methods used to create the synthetic representation.
Generation basis
Record the operation and its evidential basis, allowing the tool and operator to remain unknown.
- What evidence identifies face replacement, reenactment, voice cloning, lip synchronisation, scene synthesis or another operation? provenance
- Which production details come from examined records or reproducible evidence, and which are creator assertions or analytical inferences? provenance
Input and copy history
Relationships between source recordings, synthetic outputs and encountered derivatives.
Traceable transformations
Identify supported lineage links and transformations that may affect authenticity assessment.
- Which source recordings or likeness and voice references can be linked to this artifact, and what supports each link? provenance
- Has this copy been cropped, dubbed, recompressed, screen-recorded or stripped of disclosure or provenance information? provenance
Authenticity assessment Evidence, limitations and competing explanations for the artifact's authenticity.
An agent must distinguish a detector result from a justified conclusion and avoid confusing synthetic production with falsity of every depicted claim.
Examination evidence
Observations from media analysis, provenance checks and comparison material.
Bounded test results
Tie each assessment result to the tested copy, method and proposition it can address.
- Which copy was examined, using which method and version, and what result was returned on the method's documented scale? measurement
- What known limitations, input-quality problems or domain mismatches constrain interpretation of that result? boundary
Competing explanations
Alternative accounts and unresolved distinctions between synthesis, attribution and event truth.
Claim-specific conclusion
Keep conclusions about AI alteration, subject identity and depicted events separately supported.
- Could compression, conventional editing, authorised dubbing or mistaken identity explain the observations, and what evidence discriminates between these accounts? boundary
- What additional reference material or review would resolve the uncertainty that matters for the proposed action? action
Identity use and disclosure Whose likeness or voice is used, what permission covers that use, and how synthesis is communicated.
Technical classification alone cannot establish whether impersonation is authorised or whether recipients understand the fabrication.
Subject authorisation
Evidence of permission for the specific identity use and intended distribution.
Permission scope
Record the basis and limits of asserted consent without inferring it from public availability of source media.
- What evidence supports permission to synthesise this person's likeness or voice, and who supplied or verified it? provenance
- Does that permission cover the depicted content, proposed purpose, audience, duration and onward distribution? boundary
Recipient disclosure
Whether the encountered artifact communicates its synthetic nature in its actual presentation context.
Disclosure survival
Assess disclosure on the relevant copy and modality, including whether ordinary reuse removes it.
- What disclosure is visible or audible to a recipient of this copy, and what information exists only in metadata or surrounding text? measurement
- What authorised change would keep disclosure understandable when the artifact is clipped, embedded or consumed in only one modality? action
Use context and response The artifact's current use, the consequences of mistaken authenticity and the available handling decisions.
The same synthetic representation may require different treatment depending on audience, claimed authority, permission and evidential uncertainty.
Encounter context
How this copy is presented and what a recipient may be induced to believe or do.
Reliance and exposure
Record observed presentation and plausible consequences without inventing intent, reach or harm.
- Is this copy presented as entertainment, reconstruction, evidence, an identity check, a personal message or another use, and what establishes that context? provenance
- What decision could a recipient make by treating the represented identity, speech or event as authentic, and which exposure or consequences are actually observed? boundary
Handling decision
A justified response tied to the specific artifact, use and authority.
Authorised response
Record what may be done now and which unresolved issues require review before further use.
- Which applicable permission, policy or external determination authorises allowing, labelling, restricting, preserving or escalating this artifact? action
- Which action is justified by the current evidence, who may execute it, and what new evidence would trigger reconsideration? action
Evidence and external alignment What the world already says about this thing, gathered so the model can be checked against it.
A model that cannot be lined up against existing standards, identifiers and practice cannot be adopted by anyone who already uses them.
Reported evidence
Findings from the breadth pass, kept separate from the structural claims.
Kinds and varieties
Reported by the breadth pass; each item needs checking against its source before it becomes normative.
- face-swap video
- voice cloning / audio deepfake
- full-body or puppeted video
- text-to-image or text-to-video likeness synthesis
- lip-sync / talking-head reenactment
- image deepfake (still photograph)
- real-time interactive deepfake
- partial-manipulation deepfake (expression, gaze, or attribute edit)
- Which of these kinds and varieties hold for the sense of deepfake this model covers, and on what evidence? provenance
Sources
- Research in progress - Will be replaced after lookups
What the second pass must settle
- With no registry definition recorded, should this entry include fully synthetic people and events, or require imitation of an identifiable real subject?
- Where should the operational boundary fall for AI dubbing, lip synchronisation, beauty filters and restoration that alter appearance or speech without intending deception?
- How should the PHY / PHY.MAT placement distinguish the media artifact from its physical carrier, represented subject and transient streamed instances?
- Which evidence combinations and evaluation conditions justify moving from suspected to supported classification for each modality?
- How should consent withdrawal, disputed authority and jurisdiction-dependent handling requirements be represented through links to neighbouring models?