When The Times needed a name for its new business podcast, it did something most media companies would find absurd: it asked an algorithm to simulate thousands of readers, then polled them.
The winner? "The Business." Simple, direct, perhaps unremarkably so. But the method behind the choice represents something stranger and potentially more consequential than the result. The London newsroom had used Electric Twin, a year-old startup, to create what the company calls synthetic audiences—AI-powered stand-ins for real people that can supposedly deliver market research insights in minutes rather than the weeks traditional methods require.
Now Electric Twin is putting $14 million behind the bet that this approach might actually work. The company announced the funding on February 11, led by Atomico, with LocalGlobe, Mercuri, and Samos coming in alongside a clutch of notable angels: Marc Andreessen, Slack's Cal Henderson, and perhaps most tellingly, Eric Salama, the former chief executive of Kantar, one of the world's largest market research firms.
That last name matters more than the others. Market research is an industry built on human behavior, tracked painstakingly through surveys, focus groups, and panels that cost thousands and take weeks to organize. It's also an industry deeply skeptical of shortcuts—especially ones that smell like Silicon Valley hubris. Salama's participation suggests Electric Twin might be onto something real, or at least something the old guard is taking seriously.
What They're Actually Selling
Electric Twin's pitch sounds simple until you think about it for more than five seconds. The company builds what it describes as synthetic personas—digital representations of target audiences constructed from demographic data, large language models, and what co-founder Dr. Ben Warner calls "social science methodology." These aren't chatbots role-playing as consumers. The system supposedly ingests context about real populations, then generates thousands of simulated individuals who respond to surveys, evaluate creative work, or participate in focus groups that exist only on servers.
Two products anchor the offering. Merlin handles audience analysis and message testing. Echo runs live synthetic focus groups. The company claims coverage across 173 countries, pulling from what it describes as a library of two million synthetic personas. Users can test product concepts, validate campaign messaging, probe new markets—all without recruiting actual humans or paying panel providers.
The use cases lean toward the practical rather than the revolutionary. A telecom company (Lebara, one of the few named customers) might test pricing strategies. A media outlet might gauge reactions to editorial concepts. A retail brand could explore sensitive positioning before committing to expensive creative production. It's the kind of work that traditionally requires focus group facilities, moderators, and calendar coordination that makes everyone involved slightly miserable.
Alex Cooper, the CEO and co-founder, frames it as an economic transformation. "We're building the world's most powerful synthetic audiences platform," he wrote in the February 11 announcement. The phrasing has that founder enthusiasm common to funding announcements, though perhaps less hyperbolic than the category usually allows.
The Accuracy Question
Electric Twin claims 95.5% accuracy on something called 1-MAE—mean absolute error, for those keeping score. On NDAM, a stricter measure that tracks how responses distribute across answer options, the figure drops to 92%. For context, the company says humans taking the same survey twice typically hit around 94% NDAM. That would put synthetic personas within a couple percentage points of real test-retest reliability.
Which is either impressive or meaningless depending on how you evaluate it.
Warner, who served as Chief Adviser on Digital and Data to the Prime Minister before launching Electric Twin, detailed the validation approach in a technical post alongside the funding news. The methodology involves holding out evaluation questions during persona construction and using private datasets—material theoretically not contaminated by whatever the large language models ingested during training. The distributional accuracy measurement happens at the question level, not just in aggregate, which is admittedly more rigorous than directional hand-waving.
Still, accuracy metrics mean little without understanding the use case. The Times matched its synthetic audience to a holdout group of real subscribers at roughly 92% accuracy, according to Digiday's reporting. Good enough to name a podcast. Perhaps not rigorous enough for, say, a presidential tracking poll or brand health monitoring—applications Electric Twin itself says aren't suitable for synthetic methods.
The company secured ISO 27001 certification in December, signaling it's chasing enterprise customers with serious data governance requirements. That's the market where validation matters and compliance departments ask uncomfortable questions.
The People Behind It
Cooper's background reads like a fictional character's resume. He ran the UK government's COVID-19 testing program during the pandemic's peak chaos—director at No. 10, responsible for the logistical nightmare of mass testing infrastructure. Before that, military command. It's an unusual path into startup land, though perhaps fitting for someone trying to sell the government and large enterprises on trusting machines to simulate their customers.
Warner's credentials are similarly heavy on public sector pedigree, though his academic position at the London School of Economics adds a patina of research credibility the space probably needs. In January, the company hired Leanne Tomasevic from Truth Consulting to lead insights and strategy. That signals Electric Twin understands it's not just selling software—it's embedding with research teams who have spent careers perfecting traditional methodologies and may not appreciate disruption from technologists who just discovered consumer behavior exists.
A Crowded, Confusing Market
Electric Twin isn't alone. Dentsu launched something called Generative Audiences in January. Qualtrics has Edge Audiences. Competitors with names like Evidenza, Artificial Societies, and Aaru are building their own versions. The category suddenly feels both inevitable and undefined.
What differentiates Electric Twin, at least in the marketing materials, is validation with tier-one publishers and that focus on distributional accuracy rather than just "directional insights"—industry speak for "we're pretty sure this is roughly correct." The Times deployment offers proof of concept, though naming a podcast hardly qualifies as high-stakes validation.
The company says it's completed 40,000 evaluations across 155 countries. They acknowledge limitations—synthetic audiences aren't recommended for tracking polls or brand trackers, and the models need regular refreshing to avoid drift. Digiday's explainer noted the platforms require careful guardrails to elicit realistic behavior, which is perhaps understating the challenge of teaching algorithms to think like actual humans with irrational preferences and contradictory opinions.
What Comes Next (Maybe)
Cooper and Warner are betting that traditional market research has a speed problem worth $14 million to solve. Whether synthetic audiences become infrastructure or remain a niche tool for low-stakes testing depends on questions the funding announcement doesn't answer.
Can the technology handle cultural nuance across 173 countries, or does it flatten consumer behavior into statistically plausible but fundamentally generic responses? How does it perform when actual money rides on the research—product launches, rebranding campaigns, market entry decisions? And perhaps most importantly, will research teams trust it when the CEO asks why they're still paying for focus groups?
The Times named its podcast "The Business" based on Electric Twin's synthetic readers. It's a fine name. Whether it's the right name remains unknowable until the real audience weighs in—which is, of course, the fundamental tension at the heart of synthetic research. You're replacing messy, slow, expensive reality with clean, fast, affordable simulation.
Sometimes that's exactly what you need. Sometimes it's just another way to be precisely wrong.
