Industry

Deepfake videos work even when everyone knows they are fake

NewsGuard debunked two deepfaked Ukrainian MPs after 130,000 views. Six checks to run on a video before you share it.

Multi Chats Team
September 29, 2026 · 7 min read

On 13 August 2026 two short videos began circulating on pro-Kremlin social media. In one, the Ukrainian member of parliament Andriy Levus argues for opening negotiations early, on the grounds that the terms will only get worse. In the other, his colleague Ivan Yunakov says "We are losing our lands", a sentence he never said.

Both clips are AI fabrications. NewsGuard's Alice Lee published the analysis on 26 August 2026, and by then, in NewsGuard's count, the videos had gained 130,000 views in two weeks.

Two weeks is the number worth sitting with. The verdict was clear and it arrived after the clips had been watched. The Decoder, writing the findings up the same day, put the consequence in one line: "these videos always serve their purpose, whether they get debunked or not". The Decoder wrote that sentence itself rather than quoting NewsGuard, and it is why a checking habit is worth three minutes even though you will never outrun a fabricated clip.

What NewsGuard found in the two clips

The evidence in the report is unglamorous, and that is the useful part. NewsGuard counted blinks. Levus appears to blink once in 27 seconds. Yunakov blinks 13 times in 12 seconds. Both men, in the report's words, "speak with a monotonous voice and little facial expression".

Then came an ordinary comparison against real footage. In the video Levus has dark hair and looks younger than he does in footage of him taking the oath of office in Ukraine's parliament in June 2026, where his hair is greyer. NewsGuard also recorded what was missing: no credible reports of either man calling for a peace deal or saying Ukraine is losing the war. Both were emailed for comment and neither replied. The report states plainly that NewsGuard could not confirm who produced the videos.

One step in the report needs a caveat attached. NewsGuard also ran both clips through an AI-detection tool called Hive, which put them at 97.6 percent and 99.9 percent likely to be AI-generated. Those figures sit beside the blink counts, the hair and the missing coverage rather than in place of them, and we would not tell you to hand your judgement to a detector score.

A debunk does not travel the way the clip did. The people who watched a 27-second forward on 15 August were not, as a group, reading a fact-checking newsletter on 26 August. Assume every correction reaches a fraction of the original audience, later, and in a calmer mood.

What the EU labelling rule reaches, and what it leaves alone

From 2 August 2026 the European Commission's AI Office and the national authorities working with it began enforcing the AI Act, and the new transparency rules started to apply on the same date. The Commission's own summary of them is short. Chatbots and other interactive AI systems have to tell users they are dealing with AI. AI-generated or altered content has to carry machine-readable marks so it can be detected more easily. And, in the Commission's words, "Deepfakes (images, videos, or audio that have been edited or generated using AI) will have to be labelled." A duty only reaches whoever is willing to answer for it. Whoever made and posted these two clips was not going to declare them, and no rule changes that, so the labelling requirement does its work on the material that was already going to be honest about itself. Treat a video with no label as telling you nothing either way.

Six checks to run before you share a video

None of these need software you do not already have, and together they take about three minutes, which is three minutes longer than forwarding takes.

  1. Look at the account before you look at the video. Open the profile that posted it. When was it created, what else has it posted, and does anything on it exist outside one political theme? A clip that arrives with no origin, on a channel with no history, is the ordinary case here rather than a rare one.

  2. Find the earliest copy you can reach. Search the claim in your own words and push for older results. If everything you can find was posted in the last day, the clip has no traceable origin yet. The Ukrainian videos ran for a fortnight before any analysis existed, so "nobody has debunked it" carries no weight during the window when nobody has looked.

  3. Screenshot one clear frame and search the image. Pause where the face is well lit, take a screenshot, then upload it to an image search that accepts uploads, such as Google Images, TinEye or Yandex. The question here is narrow: where else have this face, this room and this framing appeared? An earlier copy sometimes surfaces the genuine footage a fabrication was built from, and dates the material.

  4. Ask who else would be carrying this story. A sitting MP arguing for surrender terms would be covered within hours. Check a national outlet, a wire service, the person's own account. Part of NewsGuard's finding on these clips was an absence: there were "no credible reports" of either man saying any of it. Silence where coverage should be is itself a finding.

  5. Watch it twice, once with the sound off and once with your eyes closed. Muted, watch the blinking and the face: too few blinks, a sudden burst of them, an expression that barely shifts while the speaker says something they should feel. With your eyes shut, listen for a flat, even delivery with no breath in it. Then compare the person against recent real footage, which is how greying hair gave the Levus clip away.

  6. Treat a label as a label and nothing more. A platform tag saying AI involvement was declared or detected tells you that someone or something flagged the file. The absence of a tag tells you only that nobody flagged it. Neither says anything about whether the events in the video happened.

Nothing on that list can certify a video as genuine. Each step raises or lowers your confidence, and the honest result is often that you still do not know. That is a usable answer, because it settles what to do next. You do not forward it.

Can you just ask an AI to check it?

You can hand a model a frame. 47 of the 60 models in MultiChats accept an image upload, so you can screenshot a still, attach it and ask what is in the picture, which genuinely helps with a badge or a shopfront you cannot place. What comes back is a description, and a description carries no information about how the frame was made. A model will narrate a fabricated still as fluently as a real one, for the same reason it will invent a citation. Use the answer to generate leads you then check yourself.

If you cannot tell, do not pass it on

The lesson from these two clips is that the fabrication only had to be good enough for a forward, and that bought it a fortnight. You are not going to outpace that on your own. What you can do is decline to be the hop it needs, a smaller contribution than a debunk and a more reliable one.

If something looks important and you cannot place it, send it to a fact-checker before sharing it more widely. NewsGuard publishes its findings in the Reality Check newsletter, and most European countries have a fact-checking desk that accepts reader submissions. Sending a clip to someone whose job is checking it costs nothing, and it is the one step here that does not depend on you being right.

Sources