Lecture 04 — Pixels don't lie (usually)

An introduction to image forensics, from reverse search to GAN fingerprints

15 March 2027

Image forensics has two long histories that recently collided. One is old: photographs have been retouched, staged and misattributed since the daguerreotype, and reverse image search, EXIF inspection and shadow-and-light consistency checks were built for that world. The other is new: diffusion models don’t retouch a photograph, they hallucinate one from nothing, and they leave a different kind of trace — statistical regularities in frequency space, inconsistent reflections, hands that almost work.

March 2023 gave the new history its landmark case: an image of Pope Francis in a white puffer jacket, generated on Midjourney by a Chicago construction worker as a joke, convinced a large share of the internet — including, Poynter’s media literacy team admitted, its own staff. Nearly half of 240,000 people polled by a science communicator believed it was real. What gave it away, once people looked closely, was exactly what this week teaches: a smudged ear, glasses melting into a shadow, and a coffee cup warped in a hand that doesn’t quite hold together.

The old tools still catch a shocking amount, because most fabricated images aren’t generated from scratch — they’re a real photo, misdated and recirculated with a new caption. The new tools catch a different, harder category, and they are locked in a race they are currently losing. You will run both kinds this week, on the same image, and see where they disagree.

Outline#

• reverse image search and metadata as the first, cheapest checks
• what a diffusion model leaves behind: artefacts, frequency signatures, anatomical tells
• why “the hands look wrong” was never a durable method

Further reading#

• TIME — How to Spot an AI-Generated Image Like the “Balenciaga Pope” — the case this lecture opens with, and a practical checklist written in its immediate aftermath.
• Deepfake Media Forensics: Status and Future Challenges (2025) is a recent survey of the field, organised around the FF4ALL research project — useful for seeing how detection, attribution and authentication are treated as genuinely separate technical problems, not one undifferentiated “spot the fake” task.