AI polling with "fake people" secures nearly £7m from early Klarna backer
A UK start-up led by a former No 10 adviser has raised almost £7m to build AI-driven polling that simulates respondents instead of calling them. The backing comes from an early investor in Klarna, signalling growing confidence that synthetic populations can speed up opinion research and scenario testing.
The company's pitch is simple: model a representative population, ask it questions, calibrate it to real-world data, then iterate fast. Supporters say it cuts cost and time while avoiding low response rates; critics worry about bias, drift, and misuse.
What "polling fake people" actually means
Think of it as an agent-based survey pipeline rather than a traditional panel. Each "agent" is assigned demographics, attitudes, media diet, and behavioral tendencies, then prompted to answer survey items or to react to stimuli.
- Seed and scope: Build a synthetic population mirroring census and high-quality survey benchmarks.
- Response modeling: Use large language models conditioned on profiles to generate answers and rationales.
- Calibration: Weight and fit to known distributions (e.g., RIM weighting, post-stratification, MRP) to reduce bias.
- Scenario testing: Introduce policy changes or messages and observe simulated shifts over time.
- Uncertainty: Report error bars from resampling, model ensembles, and parameter sensitivity.
Why investors care
- Speed: Turnaround in hours instead of days or weeks.
- Coverage: Test long-tail subgroups that are hard to reach with live panels.
- Exploration: Run many what-if scenarios before committing to costly fieldwork.
- Cost discipline: Reduce spend on repeated tracking while keeping human studies for critical checkpoints.
What would convince scientists
- Transparent benchmarks: Side-by-side comparisons with gold-standard surveys on identical instruments.
- Error reporting: MAE/Brier scores, calibration curves, and reliability across time and topics.
- Pre-registered forecasts: Publish predictions before events (elections, referenda, policy reactions) and score them after.
- Ablations: Show how performance changes when removing demographics, media priors, or behavioral features.
- Bias audits: Demographic and ideological fairness checks; distribution shift detection; guardrails for hallucination.
- Replicability: Frozen model snapshots, versioned prompts, and open summaries of weighting schemes.
Limits and risks
- Model bias: LLM priors can skew opinions; calibration reduces but doesn't erase this.
- Shift over time: Culture and news cycles move faster than model updates.
- Question design: Poor prompts or leading wording will amplify error.
- Misuse: Synthetic outputs presented as real responses can mislead stakeholders.
- Ethics: Clear labeling and documentation are non-negotiable.
Where synthetic polling helps research teams right now
- Instrument testing: Pre-test question wording, order effects, and scale choices before live fielding.
- Scenario analysis: Message/ad concept screening across many subgroups without blowing the budget.
- Rare-population studies: Early-stage reads where live recruitment is slow or privacy-sensitive.
- Policy prototyping: Rapid reactions to draft proposals to prioritize which to test with humans.
- Privacy-preserving analysis: Shareable synthetic microdata for collaboration and teaching.
Practical workflow you can adopt
- Run synthetic first; human validate later on the few best hypotheses.
- Combine with MRP or hierarchical models to structure uncertainty and subgroup estimates.
- Use stable benchmarks (census, high-quality national studies) for regular recalibration.
- Publish a short methods note with each release: data sources, model version, prompts, weighting, and known limits.
- Track performance over time with a standing scorecard; retire underperforming configurations.
Performance claims to watch
The company says Mr Warner's modeling previously anticipated a landslide win for Donald Trump in the 2024 US election while many pollsters expected a close race. Treat this as a claim until third-party evaluations and out-of-sample scores are available.
Why this funding round matters
Securing almost £7m from an early Klarna backer gives the team runway to scale data pipelines, validation, and governance. If they publish rigorous benchmarks and accept independent audits, synthetic polling could become a standard pre-test before human fieldwork, not a replacement for it.
Further reading
Upskill your team
If you're building a synthetic-to-human research pipeline, structured training helps. See practical courses by role at Complete AI Training, including AI Research Courses and AI Design Courses.
Your membership also unlocks:
