Rob's Notes 28 - OpenAI and the (Ads) Lessons of Social Media
Why their latest acquisition is a precursor to building an ads business
Back in May 2025, OpenAI announced they were bringing my former colleague Fidji Simo on as CEO of Applications (she had joined their board in 2024). Sam Altman said in this note that Fidji brings "a rare blend of leadership, product and operational expertise, and genuine commitment to ensuring our technology benefits everyone". Fidji is an excellent and well-respected executive who helped scale the Facebook application for the company, and then subsequently took Instacart public while building out a very significant retail ads business of over 5,000 CPG and retail advertisers. The last part of Sam’s note - “ensuring our technology benefits everyone” - definitely indicated to me at the time, that one of Fidji’s main goals will be creating and scaling an ads business. A robust ads offering could help OpenAI extend ChatGPT and other products to billions more users globally. Not unlike Meta and Google, of course.
That’s why OpenAI’s >$1 billion purchase of Statsig, an experimentation start-up, is not a niche software deal. It is certainly partially an acquihire of Vijaye Raji, a very smart engineering leader who worked with Fidji at Facebook/Meta, and who will report to her as OpenAI's CTO of Applications (making their products “truly beneficial to people everywhere”). Statsig’s tools - feature flags, holdouts and endless A/B tests - were built to somewhat mimic Facebook/Meta’s internal tools. These internal tools helped the company know (for example) whether a nudge kept people scrolling for minutes longer, or what the yield/drop off rates might be for incremental ad load in the newsfeed. As a Statsig product manager (also formerly at Meta) said on the company blog in 2022:
“One of the most expensive holdouts Facebook runs is a long term Ads holdout. Yes — there are a set of people that get to use Facebook without advertisements! FB values this because it helps them measure the costs of ads on engagement. It also helps them isolate the impact of ad specific bugs.”
In acquiring Statsig, OpenAI is not just buying software. It is buying the ability to run millions of controlled experiments on its own users that opens the door to building a profitable and valuable ads business**.
The timing is suggestive. OpenAI is building out its “Applications” arm, with ambitions beyond paid ChatGPT subscriptions. A mass-market free app, subsidised by advertising (and perhaps with new social features?), looms. Such a business needs exactly what Statsig offers: infrastructure to run holdout groups, suppress features for some users, and quantify whether an ad, notification or ranking change translates into higher engagement.
The promise
There is sense in this. Consumer AI products are still young and clunky; rapid iteration could improve usability and safety alike. The talent behind Statsig, steeped in Meta’s scale, gives OpenAI a ready-made experimentation culture. And a free, ad-supported tier could widen access to powerful AI, much as Google Search or the Gmail or Facebook apps once opened the internet to billions. At its best, this could be a key unlock for humanity in democratizing access to both information AND reasoning.
The perils
Yet the hazards are equally clear. Facebook/Meta’s history shows how an obsession with metrics can corrode judgment. Engagement is easy to measure; trust, wellbeing and democratic impact are not. If OpenAI leans too heavily on “what the numbers say”, it risks replaying Silicon Valley’s least edifying chapter: maximizing clicks and advertiser outcomes while downplaying responsibility. Regulators will not be blind to AI models run through opaque holdouts or experiments whose side-effects users cannot see.
Meta has historically been led by engineering and product leaders, whereas companies like OpenAI have tended to have researchers play a key role. It’s important to watch how this might shift over time at OpenAI, and whether employee and team incentives could lead to an over-reliance on growth metrics at the cost of overall health or safety. To wit from the aforementioned Statsig blog post, there is both power (to bring important advances to a wider audience) and peril in this path:
At Facebook, most product team[s] on the core app calculate the cumulative impact of all features shipped over the last 6 months. This aligns with their goal setting and performance review process. At the start of every period — they create a small holdout (1–5% of users). At the end of the half, they measure the impact of the features they shipped by comparing metrics against the holdout group. They release the holdout, and start a new one for the next half.
The balance
Still, there is room for hope. OpenAI’s leaders insist they are building not just products, but institutions that handle novel technologies responsibly. If experimentation is used to make models safer, not simply stickier, Statsig could prove more than an ad-tech echo. The company might yet show that scale and stewardship can co-exist. I myself certainly hope so, and I know some great people who are already at the new AI companies who are helping with exactly that.
For now, though, skepticism is warranted. The world has seen this movie before. The danger is that OpenAI, in recreating Meta’s testing machine, also inherits its blind spots.
The hope is that it has learned enough to rewrite the script.
Notes:
** Per ChatGPT 5: “In antitrust inquiries and academic testimony, economists have pointed to “no-ads holdouts” as evidence that Meta knew the opportunity cost of advertising”, for example: The Consumer Welfare Effects of Online Ads (Collis et al., 2024)