Almost everyone approaches synthetic media with the wrong question. "Does this look AI-generated?" puts you in a contest against a system explicitly optimized to beat that judgment, and it's a contest you will lose more often each year. The question that keeps working is duller and far more powerful: where did this come from? Provenance doesn't get harder as generators improve. A fabricated video still has to appear from somewhere, attributed to someone, at some moment — and that trail is where fakes reliably fall apart.
The Four Moves: SIFT
Professional fact-checkers use a method that fits in four letters, and it takes about ninety seconds in practice.
Before anything else, notice your reaction. Strong emotion is the delivery mechanism for nearly all viral falsehood — outrage and vindication both. If you feel an urge to share immediately, that urge is the intended output, and it's the moment to slow down rather than speed up.
Not the content — the account. When did it join? Does it post about anything else? Is the outlet's name a near-copy of a real one? Fabricated stories almost always ride on accounts with no history, or on domains built last month to look like local news. Thirty seconds on the profile answers more than an hour staring at pixels.
Search the claim itself and see who else has it. Genuinely significant events get covered by outlets with opposing incentives. If a dramatic claim exists only within one ideological cluster, that's not a scoop — that's the shape of a fabrication.
Follow the media back to its first appearance. Reverse image search a frame; search a distinctive phrase from a quote in quotation marks. Recycled real footage relabeled as a new event is more common than fully synthetic footage — and tracing catches both.
Content Credentials: Real, and Limited
There's now a technical layer worth understanding. C2PA — consumer-facing name Content Credentials — attaches tamper-evident provenance data to a file: who or what signed it, which tool created or edited it, and whether AI was involved. It's no longer a pilot: the coalition passed 6,000 members and affiliates by January 2026, with OpenAI, ElevenLabs, Nvidia and others joining through May [1].
More consequentially, on May 19, 2026 OpenAI and Google announced a dual-layer approach combining C2PA metadata with SynthID watermarking, with verification rolling into Google Search and Chrome [2]. That means provenance checks are moving from specialist tools into the places ordinary people already look.
C2PA is not a truth machine. It proves a provenance record exists, is cryptographically signed, and still matches the file. It does not prove the photo depicts reality — a genuine camera can photograph a staged scene, and a real image can carry a caption that lies [1].
Absence proves nothing. Most images circulating today have no credentials, and metadata is routinely stripped by screenshots and re-uploads. "No credential" means unknown, not fake — which is exactly the misreading to guard against as these badges appear in more places.
Why "AI Detectors" Aren't the Answer
The tempting shortcut is a detector that scores whether something is synthetic. Treat those outputs as weak evidence at best. Edinburgh research found the AI fingerprints most detection systems depend on are bypassable, which makes detection structurally reactive — always trailing the newest generator [3]. We've also covered the equivalent problem in text: AI-writing detectors produce false positives, and they fall hardest on non-native English speakers, as we documented in the students guide.
And the failure to internalize above all: do not ask a chatbot whether media is real. During the March 2026 circulation of synthetic strike footage, users asked a platform chatbot to verify the videos; it declared them authentic and produced fabricated Reuters and CNN citations to support the verdict [4]. A language model returns fluent, confident text whether or not it has any capacity to know — and fabricated citations are indistinguishable from real ones at a glance.
Stop asking whether it looks fake. Ask where it came from. That question doesn't get harder as the fakes get better.
— The one habit worth buildingThe Practical Toolkit
- Reverse image search — Google Lens, TinEye, or Bing Visual Search. Best single move for any photo. Catches recycled and relabeled images instantly.
- Quote search in quotation marks — if a public figure said it, the exact phrase appears in transcripts and coverage. Zero results for a supposedly explosive quote is the answer.
- Content Credentials inspection — via Google Search/Chrome as it rolls out, or a C2PA viewer. Present-and-valid is informative; absent is neutral.
- Wire services — AP, Reuters, AFP. If it happened, they carry it. Their absence on a huge claim is a red flag.
- Established fact-checkers — Snopes, PolitiFact, AFP Fact Check, Full Fact. Viral fakes get debunked fast; someone has usually done the work already.
- Video frame extraction — pause, screenshot a distinctive frame, reverse-search that. Works when searching the video itself doesn't.
Synthetic Text Is the Bigger Volume Problem
Images get attention, but the larger share of synthetic content is written: fabricated news articles, invented quotes, fake product reviews, and entire sites built to rank in search. Signals worth checking: no named author, or an author with no other work anywhere; no dateline or a date that doesn't match the events described; citations to studies that either don't exist or don't say what's claimed; and a curiously even, hedge-free tone across every paragraph.
The strongest single test remains checking one specific claim. Pick the most checkable fact in the piece — a number, a date, a named study — and verify it independently. Fabricated articles rarely survive one honest spot-check, because the details were generated to sound plausible rather than to be true.
Everything in this guide reduces to one shift: stop trying to perceive fakery and start tracing origin. Perception is a race you'll lose a little more each year. Origin-tracing gets easier as provenance standards spread — and it already works today, using nothing more than a reverse image search and a bit of patience.
Ninety seconds. Source, corroboration, original. That's the whole discipline, and it's the difference between being part of the correction and part of the spread.