industryAug 14, 2026 · arXiv preprint 2608.04205 · 3 min read

Harvard and MIT gave away 8.3 billion fake customers, and companies will test on them

Morning Byte · by Keith Andrade

Comments

Two researchers sit at a plain table on a mezzanine, heads together over a tablet showing a document, laptops and a red notebook beside them. Through the full-height glass wall behind them a crowd fills the sunlit plaza below, and as it recedes the figures resolve into repeating copies of the same few people, arranged in a grid.

The three takeaways

Persona 8B costs nothing since August: 8.3 billion fake customers, more than Earth's population.

The same test's buy rate quadrupled depending only on which AI played the shopper.

It targets human user panels, which means your next price test may be simulated.

Source

arXiv preprint 2608.04205 · Harvard/MIT-led team · Aug 2026

Morning Byte · weekly digest

One email a week. The stories that mattered.

What changed

A team led from Harvard and MIT released Persona 8B for free in August: a database of 8.3 billion simulated customers, more profiles than the world has people, each carrying 1,290 traits, which means product testing just got a synthetic crowd.

The study shows the tool runs Claude and GPT models as stand-in shoppers, chatbot users, and app testers, replacing slow, costly human panels for early screens.

8.3B

Simulated customer profiles, released free. Each carries 1,290 traits. More profiles than the world has people.

A synthetic crowd for product testing.Source: arXiv 2608.04205, Aug 2026

One AI shopper23.2% buy

Another AI shopper93.9% buy

Same shopping test, different AI playing the customer.Source: arXiv 2608.04205, controlled trials

Why it matters

The researchers' own numbers carry the warning. In this year's trials, personas stayed in character 91.5 percent of the time, yet on one identical shopping test the buy rate swung from 23.2 percent to 93.9 percent depending only on which AI played the shopper, which means the customer you simulate depends on the model you rent.

For working people the stakes are direct, because the tools used to test prices, apps, and chatbots on you are shifting from real focus groups to AI stand-ins. The team behind it claims the tool targets slow, costly human user-testing panels. Taken together, the swing suggests the same AI can play both the customer and the product, and grade its own homework.

What to watch

If product teams treat persona results as early screening only, the tool could save real testing money; if they justify price hikes with it before the model-dependence problem is fixed, customers would pay for a simulation's taste.

Whether these uses become standard is not yet known. Watch whether the swing narrows in follow-up work, because until it does, a buy rate from a simulated crowd is a coin with a chosen side.

Grok 4.6 ties for third best, and the real price hides past 200,000 tokens

market

Duolingo's users grew 23 percent, but the cash they pay grew 8

Back to Morning Byte News

Reader comments

Not signed in yet — hit Post and we'll finish it together