What Polymarket’s Fake Wins Reveal About Manufactured Consensus in the AI Era

In June, the Wall Street Journal reported that Polymarket had been paying creators, mostly college-aged, to film themselves placing bets on websites built to look exactly like the real platform, with some of the videos showing wins that never actually happened.

Polymarket then hired a network of “clippers” to repost the footage until it looked like an organic wave of people getting rich off the platform. Across a set of 118 videos, creators celebrated close to $900,000 in winnings that, had the bets actually been placed, would have lost more than $166,000, and the full campaign covered 1,105 videos, none of which reflected an actual trade.

Most of the reaction has focused on the idea of false advertising, and that’s a fair description, but it stops short of the more useful questions of how Polymarket got caught, and if that kind of campaign will go undetected moving forward. The company ultimately was caught because this kind of campaign relied on humans, and humans leave a trail. But AI is changing that. The cost of manufacturing a convincing online narrative is falling fast, and our ability to catch it is falling right along with it.

A model that ran on people, up against one that doesn’t need them

Polymarket’s campaign worked because it looked organic. Creators posted from their own accounts as though they were recording their own experience, and the clips spread until it felt like a genuine wave of enthusiasm rather than a paid operation. Paid creators only account for half of what made this work, though, and the other half, less visible but just as essential, was fake engagement.

It is not a coincidence that these videos pulled in 140 million views on TikTok, YouTube, and Instagram from creators who otherwise averaged a few hundred views per post. Reach like that usually requires pumping a video with artificial likes, comments, and views, typically through bots or coordinated accounts, until the platform’s recommendation engine mistakes it for something people genuinely want to watch. Paid creators alone rarely produce that kind of lift, which is why Polymarket almost certainly ran both plays at once.

What made this expensive, and ultimately traceable, is that it ran through people at every step. Fake sites had to be built and hosted, creators had to be paid, and a firm had to be hired to manage the repost network, and each of those steps left a record that investigative reporters were able to follow back to the source.

AI removes that constraint, since synthetic personas, generated video, and cloned voices can produce the same illusion of organic enthusiasm without a payroll, a contract, or a paper trail. The mechanics stay the same. What disappears is the need for anyone to sign anything or take a wire transfer to make it happen.

That shift also changes the scale of the problem. Online deepfake volume went from roughly 500,000 in 2023 to about 8 million in 2025, and once fabricated content grows that fast, it stops being an occasional intrusion into a feed and starts becoming a meaningful share of what makes up the feed.

The line between staged and real doesn’t hold

Most executives still think about online narratives in binary terms, where it’s either staged or genuine. But Polymarket’s pitch, that anyone can win big trading on their platform, started off manufactured and still gained traction with real users once enough people saw it, believed it, and acted on it. At that point, the origin of the narrative stops being the relevant fact, and a story that begins as fabrication can still produce entirely real consequences once it takes hold. Recent research found that humans can distinguish AI-generated content from authentic content only 51% of the time, close to a coin flip, which means for most people, seeing something on their feed is still enough to believe it.

Polymarket turned engineered confusion into trading fees. Point the same mechanics at an election or any contested public issue, and AI makes that kind of manipulation easy to run at scale, with the stakes moving from lost revenue to who the public trusts and how they vote.

Why the usual monitoring tools miss this

Most reputation monitoring was built to track a text-first internet, watching for mentions, keywords, and hashtags, but perception now forms inside short-form video and inside the recommendation systems that decide what gets seen. Tools built to scan captions and comment threads were never built to read tone, visual context, or the credibility of the person on screen, and none of that is optional information anymore. It’s the story itself narrated within video.

AI video generation has also gotten good enough, fast enough, that producing a convincing clip of almost anything no longer takes a studio or a production budget. Anyone running one of these campaigns already knows better than to tag the brand or use searchable language, which means the signals that actually carry the narrative, tone, creator credibility, community reaction, stay invisible to a keyword-based system by design.

The old approach of investigating after the fact, following the paper trail the way the Wall Street Journal did with Polymarket, won’t scale either, because that kind of reconstruction depends on humans leaving a trail in the first place. Take the humans out of the loop, and there’s no trail left to follow, which is why detecting this requires watching for the pattern as it’s forming rather than reconstructing the fraud after it’s already spread.

Polymarket got caught because it still needed people

Nobody should be shocked that a company manufactured its own hype. That’s a standard disinformation playbook that’s been in use for years, and Polymarket got caught specifically because its version of the playbook still ran on humans, leaving a trail a reporter could follow.

The next version of this campaign won’t need any of that. Once the cost of manufacturing consensus drops close to zero, the ability to tell real from manufactured stops being a nice-to-have capability and becomes the thing every downstream decision, about who to trust, what to believe, and how to respond, depends on.