Pixels Don’t Lie (Usually)#

Week 4 — image forensics#

Two long histories collided#

• old: retouched, staged and misattributed photographs — reverse image search, EXIF, shadow-and-light consistency
• new: diffusion models hallucinate an image from nothing, leaving a different kind of trace entirely

The cheap checks first#

• reverse image search — has this exact image appeared before, with a different caption or date?
• metadata (EXIF) — camera, timestamp, edit history, when it survives re-upload
• these catch a shocking amount, because most misinformation is a real photo misdated, not a synthetic one

What a diffusion model leaves behind#

• irregularities in frequency space invisible to the eye
• inconsistent reflections and shadows across a scene
• anatomical tells — hands, teeth, text-within-the-image — that are getting rarer every model generation

The landmark case: Balenciaga Pope#

• March 2023 — a Midjourney image of Pope Francis in a white puffer jacket, made as a joke by a Chicago construction worker
• fooled a huge share of the internet — including Poynter’s own media literacy team, and nearly half of 240,000 people polled by a science communicator
• what gave it away: a smudged ear, glasses melting into shadow, a coffee cup warped in a hand that doesn’t quite hold together

Reading#

Deepfake Media Forensics: Status and Future Challenges (2025) organises the field into detection, attribution, passive authentication, and real-world-scenario detection — worth skimming for how fast each category is moving.