For anyone
Judge an advertisement yourself
Six checks, in the order that costs you least effort. You do not need to know anything about AI to use them — the first two catch most of what matters, and none of them requires a tool.
Draft. These checks follow the module’s teaching and will be revised by the academic team as the archive grows.
Check 1
Find out who paid for it
What to look for
Political advertising in the EU must say who is behind it and who paid. Look for a “paid for by” line, a transparency note, or an entry in the platform’s ad library. If none of that exists, that absence is your first finding.
Why it works on you
An ad with no named sponsor is asking you to treat a paid message as if it were a neighbour’s opinion. Regulation (EU) 2024/900 exists precisely because undisclosed political advertising works better than disclosed political advertising.
Documented elsewhereTikTok · Romania · 2024-11 A presidential campaign that TikTok never labelled as political
In the weeks before Romania's annulled first round, content promoting Călin Georgescu spread through accounts that carried no political-advertising label. Reporting and the subsequently declassified material describe influencers paid off-platform to post campaign hashtags, and TikTok itself later removed clusters of inauthentic accounts. Whatever else was true of the campaign, the first thing a viewer could not find was who had paid for what they were seeing.
Read the reporting · VSquare ↗Check 2
Look at hands, teeth, ears and lettering
What to look for
Generated images still struggle with things that repeat: fingers, rows of teeth, the curl of an ear, the crowd at the back, and above all written text. Signs, badges and captions in a synthetic image tend to be almost-words.
Why it works on you
You read a political image in under a second, and in that second you check the face, not the hands. The parts of an image nobody inspects are the parts generators get away with.
Documented elsewhereMeta ads · Germany · 2024-05 Paid ads whose images appear generated, carrying no AI label
Ekō examined 23 ads paid for by AfD regional groups and elected representatives, together over 470,000 impressions. Several — an injured woman, a group of hooded men — appear to the researchers to contain AI-generated imagery while carrying no disclosure, which Ekō argues would breach Meta's own policy on AI-manipulated ads. These are bought, sponsored, ad-library-visible advertisements: the sponsor is known, and the synthesis is what is undeclared.
Read the reporting · Ekō ↗Check 3
Listen to the voice, not the words
What to look for
Play it with your eyes closed. Synthetic speech tends to breathe in the wrong places or not at all, keeps an unnaturally even pace, and sits in no room — no echo, no traffic, no chair creak.
Why it works on you
A familiar voice saying something outrageous is far more persuasive than text making the same claim, and it spreads further. Closing your eyes removes the face that was doing the convincing.
Documented elsewhereTikTok · Germany · 2025-02 Real march, real crowd, synthetic chant
Videos claiming to show hundreds of thousands marching for the AfD used genuine footage — including a January 2024 Hamburg demonstration that was in fact against the AfD — recaptioned, with the original sound replaced by an AI-generated crowd chant. Correctiv first revealed the manipulation and Euronews confirmed it with audio detection. Nothing in the picture was fake. The only fabricated element was what the crowd could be heard shouting.
Read the reporting · Euronews ↗Check 4
Ask whether this moment would have been recorded
What to look for
Think about the camera. Who was standing there? Would a private remark have been filmed from that angle, in that light, with that framing? Then look for a second source: any real event is covered more than once.
Why it works on you
Fabrications are usually of moments nobody could have filmed — a private admission, a closed-door remark. The scarcity is the point: something no one else has feels like a scoop rather than a warning sign.
Documented elsewhereTelegram, then TikTok · Moldova · 2025-09 A president rapping in Russian, made with a video generator
Weeks before Moldova's September 2025 parliamentary election, a video showing President Maia Sandu performing a mocking rap in Russian circulated on Kremlin-aligned Telegram channels and was amplified on TikTok and X. Researchers at Antibot4Navalny traced it to the Matryoshka / Operation Overload network and identified the Luma AI video generator. No camera crew was ever in the room, because there was no room.
Read the reporting · France 24 / AFP ↗Check 5
Notice the feeling before the claim
What to look for
Name the emotion the first three seconds produced in you — fear, disgust, vindication, contempt. Then ask what the ad actually claimed. Often there is no checkable claim at all, only an association.
Why it works on you
An ad that makes you feel something does not need to convince you of anything: you will supply the argument yourself, and you will remember the feeling long after you have forgotten where it came from.
Documented elsewhereTelegram, then Facebook · Slovakia · 2023-09 An audio recording nobody was allowed to correct
Two days before Slovakia's 2023 election, audio surfaced appearing to catch Progressive Slovakia's Michal Šimečka and journalist Monika Tódová arranging to buy Roma votes. AFP's fact-checkers found signs of AI manipulation; it was never authenticated either way. It was posted inside the 48-hour pre-election moratorium, when media and politicians are required to stay silent, and passed 100,000 views. The clip made no claim anyone could check — it offered the feeling of catching someone red-handed.
Read the reporting · VSquare ↗Check 6
Go and find the original
What to look for
Search the platform’s ad library for the sponsor. Reverse-image-search a still. Look for the same clip with a different caption. Check whether the account posting it existed six months ago.
Why it works on you
Most manipulation is not a fabricated video: it is real footage from another country, another year, or another context, with a new caption. Finding the original is usually easier than detecting a fake, and it settles the question.
Documented elsewhereTelegram and social media · Romania · 2025-05 A newsroom's format and logo, on a report it never made
After Romania's May 2025 runoff, a video built to look like a Euronews social clip — same format, borrowed branding — claimed Romania's foreign ministry had sent France a note of protest over interference in the vote. Romania's MFA refuted it, France's foreign ministry confirmed no such letter existed, and Ukraine's Centre for Countering Disinformation traced it to Operation Matryoshka. It passed 85,000 views. Opening Euronews' own channel settled it in seconds.
Read the reporting · Euronews ↗
Not an entry in this archive. Someone else reported this one, and it is linked here so the check has a worked example while the archive fills.
When you are not sure
Most of the time you will not reach certainty, and that is the honest outcome. Detection tools are unreliable and the people who make these ads know it. “I cannot tell whether this is real” is a reasonable place to stop — and a good reason not to share it. This archive records the same uncertainty: entries marked Suspected are ones we could not confirm.