What evidence did fact-checkers use to determine the Mark Carney–Epstein images were AI-generated?
Executive summary
Fact-checkers concluded the Mark Carney–Epstein/Maxwell/Hanks images were AI-generated based on a cluster of technical signs — visible watermarks linking the images to X’s Grok model, outputs from multiple AI-detection tools, anatomical and visual artefacts such as warped limbs and missing fingernails, and the absence of antecedent photographic matches in reverse-image searches — while noting that authentic, unrelated photos of Carney with Maxwell from 2013 do exist and have been misused in the narrative [1] [2] [3] [4].
1. Watermarks and provenance: a visible Grok signature
Several fact-checkers flagged that higher-quality versions of the beach and pool photos carried a Grok watermark — the signature of X’s (formerly Twitter) AI model — and that some social-media reposts had cropped the image above the ankles in ways consistent with attempts to remove that watermark, a strong clue the images originated from an AI generator rather than a camera [2] [5] [6].
2. Machine detectors: multiple tools converged on ‘likely synthetic’
Independent outlets and investigators ran the images through a battery of AI-detection tools — including AFP’s Paired Model, LeadStories’ array (Hive Moderation/DeepFake‑O‑Meter), SightEngine and others — and reported high probabilities the pictures were synthetic (for example, Hive rated one image 96.8% likely to be AI/deepfake; other tools returned 99%+ likelihoods), producing convergent technical verdicts across platforms [1] [7] [8].
3. Visual forensic clues: warped anatomy and inconsistent details
Human-led visual inspection complemented the algorithmic flags: fact-checkers noted clear artefacts such as a warped hand draped over Carney’s shoulder with a missing thumbnail, distorted arms in the background, and other physical inconsistencies that commonly appear in generative-image outputs — evidence that AI synthesis, not genuine photography, produced those anomalies [1] [2].
4. Absence of provenance in time: no prior photographic matches
Reverse image searches failed to find earlier versions of the alleged pool and beach photos predating January 2025; the only matches often led back to the same social posts circulating after Carney entered the leadership race, suggesting these images lacked photographic provenance and were newly created rather than republished from an archive or press agency [1] [7].
5. Social spread and intent: timing, promoters, and political context
Fact-checking reports documented that the AI images surfaced as Carney entered Canadian political prominence and were amplified by accounts and outlets with adversarial or conspiratorial interests; researchers and outlets traced early posts to accounts that frequently share AI-generated content and noted that the timing and distribution fit patterns of politically motivated misinformation [1] [3] [4].
6. Counter-evidence and limits: real photos miscontextualized
Reporting was careful to separate the AI fakes from a genuinely photographed, widely circulated image of Carney next to Ghislaine Maxwell at a 2013 public event — an authentic, earlier encounter that has been used as raw material for disinformation — and fact-checkers emphasized that the existence of that real photo does not validate the newer Epstein-island claims tied to the AI images [4] [9].
7. Final appraisal: technical and journalistic consensus
Across Reuters, AFP, LeadStories, Full Fact, Snopes and others, the verdict combined algorithmic detection, human visual forensics, failure of provenance checks, and the presence of AI watermarks; together those independent lines of evidence formed a robust consensus that the viral images were manufactured by generative models rather than genuine photographs, even as outlets noted the political motives behind their circulation [2] [1] [7] [5].