On 20 September 2026, the French news site Atlantico ran a cross-interview on one question: which method will give the clearest view of the 2027 presidential election? The options were telephone polls, online panels, prediction markets, or samples of voters simulated by AI. On one side was Émile Servan-Schreiber, co-founder and Chief Scientific Advisor of The Forecasting Machine and CEO of Hypermind. On the other was Laure Salvaing, directrice générale of the polling firm Verian, formerly Kantar Public. Servan-Schreiber shared the piece on LinkedIn on 22 September.
The starting point: an American case for the phone poll
Atlantico built its questions around Nate Cohn, the New York Times analyst behind the NYT/Siena poll. Cohn had explained why that poll still relies on the telephone despite the cost. Salvaing was wary of ranking survey modes: a bad phone sample stays bad, and a well-built online panel can be excellent. The persistent problem, she said, is self-selection, meaning that the people who choose to answer are not typical. Her example was the 1936 US election. The Literary Digest gathered more than two million responses, predicted a win for Alf Landon, and was proved badly wrong when Roosevelt won. Her remedy is controlled recruitment, quotas and weighting.
What a prediction market actually measures
Servan-Schreiber described a prediction market as a form of betting without a bookmaker. In what is called a betting exchange, participants trade with one another, so each one is effectively their own bookmaker. The resulting price, read as a probability, emerges from all of those opposing views, much as a share price does. In his view, markets generally beat polls, though not always. French polls are much better than American ones, he said, yet Hypermind’s markets still did slightly better at the last two French presidential elections.
His main point was that the two instruments answer different questions. A poll asks how you would vote if the election were held tomorrow. A market prices what will happen on the real polling day, six or ten months away. A poll depends on a representative sample. A market depends on participants’ knowledge of the subject, whatever their demographic profile. He added that no money is at stake on Hypermind, unlike on Polymarket or Kalshi, and argued that real money makes no difference to forecast quality.
Salvaing accepted that the two measure different things: a poll records what voters think now, and a market records what traders expect to happen. She credited markets with combining scattered information quickly. She also warned that they can be swept up by fads, share mistaken collective expectations, and draw on too narrow a pool of participants. She added that markets have sometimes swung on a media event and reversed within days.
Two rounds, and a block that leaks
Servan-Schreiber argued that France’s two-round system suits markets even better than a US-style contest. Polls tend to focus on who will reach the run-off. A market prices each candidate’s overall chance of winning: roughly, the chance of qualifying multiplied by the chance of winning the run-off. He gave 2017 as his example. Polls then pointed to a Fillon–Le Pen run-off with Emmanuel Macron well behind. By his account, the markets already ranked Macron second for the overall win, because he was likely to win any run-off he reached.
On France’s blocking of Polymarket, which the regulator ordered in mid-July (see our earlier article on the L’Express interview), his verdict was blunt. He said the block had no real effect: he had signed up the day before without difficulty, and a VPN removes any obstacle.
Synthetic voters versus AI forecasters
The sharpest disagreement concerned AI. One emerging method has a model imagine how a voter with a given profile would answer, and then repeats this across every profile. Servan-Schreiber said there is no evidence yet that this works in a real election, only a few laboratory studies. In his view, a different use of AI works better: asking the model to forecast the outcome directly. He said The Forecasting Machine and other firms have built such tools, which aggregate expert commentary and the record of comparable past elections. He expects the coming US and French elections to be the first large-scale test of the approach.
Mais aujourd’hui, une IA n’a pas d’opinion politique.
Laure Salvaing, Atlantico, 20 September 2026
In our translation: “But today, an AI has no political opinion.”
Salvaing went further. A model simulates from its training data, she said, so it risks reproducing the past rather than spotting a break with it. Electoral history is full of such breaks: Jean-Marie Le Pen in 2002, Brexit in 2016, Macron in 2017. Verian therefore holds that observing real citizens remains indispensable. She named three pitfalls for 2027: estimating who will actually turn out, AI’s still poorly understood effect on campaigns, and the spread of deepfakes.
A week earlier on B SMART: humans versus machines
On 15 September, Servan-Schreiber took three questions from Delphine Sabattier on B SMART 4Change’s Smart Tech. Asked why Polymarket is banned, he noted that betting exchanges were not authorised when France opened online gambling to competition under the law of 12 May 2010. He pointed to the risk that real money creates an incentive to leak secrets or tamper with outcomes. The regulator’s own decision refers to weather bets where sensors may have been hacked. In April, Météo-France filed a complaint over anomalous readings from its Roissy-Charles-de-Gaulle sensor, which Polymarket used to settle Paris temperature markets. He also cited the US special forces soldier charged in April with betting on the operation to capture Nicolás Maduro.
Asked whether human forecasting crowds can still compete with AI, he answered that, in forecasting terms, they cannot: as he put it, the contest is already settled.
His reason was scale, not intelligence. Machines can take on vastly more questions, including specialist ones, in minutes. No one can find enough affordable, competent people to answer them all. Asked which single method would best read the 2027 race, he chose prediction markets, his own in particular. He recalled that in 2017 the British bookmaker Ladbrokes publicly named Hypermind as the guide to follow. He attributes this to Hypermind drawing more French participants who follow the election than Polymarket, Kalshi or Betfair. The two answers fit together: an engaged crowd on one heavily followed election, and machines for the long tail of questions no crowd could cover.
