Chapter 10 — Fabrication

Two thousand people were shown a mixture of real and synthetic images and video and asked to sort them. Two of them got everything right. [ESTABLISHED] (iProov study, 2025, via Adaptive Security)

That is 0.1%. It is the number to open with because it disposes of the most common assumption in this subject — that a careful person can tell. A careful person cannot tell, and neither, mostly, can the software sold to tell for them.

This is the chapter where I have to report that verification is currently losing, and where the correct response turns out not to be a better detector.

The state of detection

Commercial deepfake detection baselines are reported around 55% accuracy, and industry analysis flags this openly as insufficient to function as a safeguard — at that rate roughly half of synthetic content passes. [ESTABLISHED] as reported; the figure comes from industry reporting rather than a standardised benchmark, so treat the precise number as indicative.

Three structural problems sit underneath, and they are structural rather than temporary:

The adversarial gap is architectural. Detectors are trained to recognise the artefacts left by known generator architectures — historically generative adversarial networks. Current diffusion models produce different artefacts, which GAN-trained detectors were never built to catch. Every advance in generation resets detection, and generation advances faster than detectors are retrained. This is not a gap that closes. [ESTABLISHED]

Detector reliability varies by who is in the image. Peer-reviewed evaluation found measurable performance disparities by age, ethnicity and gender. So a detector's error rate is not one number — it is a different number depending on the subject, in the same way, and for related reasons, as the facial recognition differentials in Chapter 6. [ESTABLISHED]

And the human-machine comparison is not what you would guess. Some studies report human perceptual detection now outperforming machine detection on video, while machines edge out humans on static images. There is no clean hierarchy to fall back on. [CONTESTED] (University of Florida, Feb 2026)

The alternative that is actually being built

If you cannot reliably analyse content to determine whether it is synthetic, the remaining option is to establish where it came from. That is the wager behind C2PA — the Coalition for Content Provenance and Authenticity — and its Content Credentials, at specification version 2.3 as of January 2026.

The adoption is real and larger than most people realise. Leica shipped the first camera signing C2PA credentials at capture in 2023; the Samsung Galaxy S25 and Google Pixel 10 now sign natively, with Nikon and Canon integrating at hardware level. OpenAI attaches C2PA metadata to generated media; Microsoft added credentials across Microsoft 365 content in February 2026; Google is rolling verification through Gemini, Search and Chrome; TikTok reports labelling over 1.3 billion videos with provenance data. The EU AI Act's Article 50 mandates machine-readable marking of AI content, which C2PA satisfies, and California's SB 942 took effect January 2026 with AB 853 following in 2027. [ESTABLISHED] (C2PA; adoption status 2026)

Now the two limitations, which matter more than the adoption figures:

Provenance is destroyed by ordinary handling. Re-encoding, screenshotting or platform recompression strips a C2PA manifest exactly as it strips EXIF. The chain is only as good as every step that preserved it, and most steps do not.

It cannot work backwards. C2PA certifies content signed at capture or creation. It says nothing about older media, unsigned media, or anything that passed through a tool that did not participate.

Put the two together and you get the honest position: C2PA can eventually establish that something is authentic, through an unbroken signing chain. It can never establish that something is fake — absence of a credential is the normal condition of nearly everything, not a signal.

That asymmetry is the whole design and it must be understood before the technology is relied on.

What an investigator does instead

Three things, in order of value.

Get upstream of the artefact. Do not analyse the image; find where it came from. The original upload, the account that posted it first, the device or the person who claims to have made it, the platform record. Chapter 8's geolocation is a provenance technique, not an analysis technique — it does not ask whether the image was manipulated, it asks whether the world it depicts exists where it claims to be. A fabricated scene must fabricate the whole scene consistently, and consistency across independent physical details is very hard to fake and very easy to check.

Corroborate across independent channels. A video is one channel. Its claims usually touch others: a flight that happened or did not, a building that exists or does not, a shadow at an angle that is possible or impossible on that date, a document filed somewhere. The MH17 work in Chapter 8 was never a judgement about whether individual photographs were genuine; it was a demonstration that dozens of independent items agreed with each other. Fabrication scales badly against that requirement.

Preserve the original bitstream. Chapter 9's discipline is doubly load-bearing here. Hash at acquisition. Preserve the file as delivered, not a re-encoded copy — because re-encoding destroys both the provenance manifest and any forensic artefact that might later be analysable by tools that do not yet exist.

The consequence nobody wants to say

The most important effect of synthetic media on investigation is not that fakes will be believed. It is that genuine evidence will be doubted, and that doubt is now always available, for free, to anyone against whom evidence exists.

A recording used to carry a presumption. It no longer does, and it will not again. The presumption does not transfer to detectors, because detectors sit near 55%. So the burden moves to provenance — to being able to say where a thing came from, who held it, and what happened to it in between.

Which returns you to Chapter 9, and to the observation that ties this part of the book together: the skill that turns out to be decisive in the synthetic era is not analytical cleverness. It is custody — the unglamorous, procedural, entirely learnable discipline of being able to account for where everything in your file came from.

The Move — Never ask "is this fake." Ask "where did this come from." When you receive a piece of media, do not begin with analysis or with a detection tool. Begin by tracing it upstream: earliest appearance, first uploading account, the physical claims it makes that can be checked against independent sources. Record the chain. If the chain cannot be established, that is your finding — not "fake," but unverifiable, which is a different and more honest word and is the one that survives challenge.

What would change my mind: a detector demonstrating high accuracy against a generator it was not trained on, in independent evaluation, would reopen analysis as a viable path. The architectural argument in this chapter predicts that will not happen, and I would treat any vendor claiming it as [CLAIMED] until an independent lab reproduces it against unseen generators.