Altman Called AI Ads a 'Last Resort'

Introduction

In May 2024, at a Harvard fireside chat, Sam Altman said this about putting ads inside AI: "Ads plus AI is sort of uniquely unsettling to me." He went further. "I kind of think of ads as a last resort for us for a business model."

Twenty months later, OpenAI launched the ad business he described as a last resort. The targeting fuel is your conversation history. And on April 30, 2026, the company quietly rewrote its privacy policy so that free users' data started flowing to Meta's platforms by default, with no announcement and an opt-out buried in settings.

From "Uniquely Unsettling" to Live in Twenty Months

Start with the words, because Altman put them on the record himself at a recorded public event. At Harvard he didn't hedge. "I will disclose, just as a personal bias, that I hate ads," he said, arguing ads "somewhat fundamentally misalign a user's incentives with the company providing the service." By June 2025 he'd softened to "I'm not totally against it." Seven months after that, the announcement came.

On January 16, 2026, OpenAI published a blog post titled "Our approach to advertising," authored by Fidji Simo, the company's CEO of Applications. Simo spent nearly a decade at Meta running advertising products and the Facebook app before a stint as Instacart's CEO. A year before the ad launch, Altman hired her. In March, OpenAI also brought on Dave Dugan, a longtime Meta ad executive, to lead its Global Advertising Solutions team. So the man who called ads a last resort went and hired the people who built one of the biggest ad machines on the planet.

The product went live on February 9, 2026, with ads at the bottom of answers for Free and Go users. By early May, OpenAI had a self-serve Ads Manager in beta and had expanded ads to the UK, Mexico, Brazil, Japan, and South Korea. The phrase "last resort" implies you've already exhausted the other resources, but OpenAI hadn't run out of anything. It's the most-used AI product on earth, with 900 million weekly users as of late February, the most valuable consumer franchise in tech. The ad business got built on top of all that, not in place of it.

What the System Reads to Pick Your Ad

OpenAI is unusually direct about the targeting, which makes the privacy framing easier to check against the product. The "Testing ads in ChatGPT" post states the methodology in one sentence: "During the test, we decide which ad to show by matching ads submitted by advertisers with the topic of your conversation, your past chats, and past interactions with ads."

"Your past chats" is the part to sit with. ChatGPT's whole pitch is that you bring it the questions you wouldn't ask a person, the diagnosis you're scared to Google or the debt you're trying to dig out of. Per OpenAI's own Help Center documentation, when ad personalization is on, the targeting signals expand to include your stored Memory and your full chat history. Memory is on by default, which makes the personalized version the effective default state for a free user, not an opt-in upgrade.

There's a real safeguard worth stating plainly: OpenAI blocks ads near health, mental health, and politics conversations, and it does not show ads in Temporary Chats. Your conversations themselves aren't handed to advertisers either, but the signal pulled out of those conversations is what runs the whole thing. The privacy policy classifies your prompts and uploads as Personal Data the company collects, and says it may personalize ads for Free and Go users using conversation content, past chats, and stored memories, "subject to your settings." That intimacy is exactly what makes ChatGPT's targeting different in kind from Google's, which reads a one-off search query and never sees a stored, multi-turn relationship with you.

The April 30 Change Nobody Got Told About

The most consequential move came with no blog post at all. On April 30, WIRED compared OpenAI's old and new privacy policy text directly and found two material edits.

The old policy said: "We don't 'sell' Personal Data or 'share' Personal Data for cross-contextual behavioral advertising." The new one cut the second half. It now reads: "We don't 'sell' Personal Data. Depending upon your choices, we may share limited data with select marketing partners for purposes of promoting our products and services to you on third-party properties." OpenAI also renamed a data category from "Vendors and Service Providers" to "Vendors, Service Providers, and Marketing Partners," with new language about sharing data with partners who are explicitly "not service providers."

WIRED then tested it. Two free accounts had marketing cookies on by default; two paid accounts, Plus and Enterprise, did not. Adweek documented the same update and described data now moving in both directions: outbound, OpenAI shares cookie and device IDs with marketing platforms including Meta for off-platform retargeting of its own products; inbound, advertisers send purchase data back. The policy spells out the inbound half in writing: "We may receive information from advertisers and other data partners... For example, we could receive information about purchases you make from these advertisers."

These cookies are easy to confuse with the in-ChatGPT ad personalization, so keep them separate. They're about OpenAI advertising itself to you across the web through Meta's infrastructure, and tracking whether those ads worked. Both mechanisms expanded in 2026, but only the in-chat one got a blog post. OpenAI's response to the criticism was narrow and technically accurate: "Nothing about our policy of not sharing people's conversations or other private user content with advertisers has changed." That's true, and also beside the point, because what flows to Meta is your cookie and device IDs, not the text of what you typed.

Who Benefits

Follow the cash and "why now" stops being a mystery. OpenAI posted a $13.5 billion net loss in the first half of 2025 alone, and the research firm Sacra pegs its 2026 cash burn at roughly $27 billion against an operating loss near $14 billion. By its own internal projections, the company doesn't turn cash-flow positive until around 2029 or 2030, and subscriptions and API fees don't close a gap that size on their own. Advertising is the third revenue stream, and OpenAI built it on the one asset Google and Meta don't have: a confessional relationship with hundreds of millions of people.

The roadmap OpenAI showed investors makes the stakes concrete. Per an Axios scoop confirmed by Reuters, the company projected $2.5 billion in ad revenue in 2026, climbing to $11 billion in 2027 and $100 billion by 2030, a target that assumes 2.75 billion weekly users by decade's end. The early numbers say it's already working: OpenAI told Reuters the US pilot cleared $100 million in annualized revenue in six weeks. With an IPO reportedly targeted for late 2026 at a valuation near $1 trillion, a working ad business is the difference between a story public-market investors will buy and one they won't.

Microsoft benefits without lifting a finger. It holds roughly a 27% stake in OpenAI's for-profit entity, valued around $135 billion at the October 2025 restructuring and higher after the March 2026 round pushed OpenAI's valuation to $852 billion. Every dollar of ad revenue that lifts that valuation accrues to Microsoft's equity. Then there's the ad-tech layer, which is where this stops being OpenAI acting alone.

The Machine Behind the Ads

OpenAI didn't wire 17,000 advertisers into your chat history by itself. On March 2, Criteo announced it was the first ad-tech partner integrating with the pilot. Criteo's own press release does the bragging: it "activates more than $4 billion in annual media spend and works with 17,000 advertisers globally." It also dropped the number that explains why every advertiser wants in. Users referred from AI platforms like ChatGPT convert "at approximately one and a half times the rate of other referral channels," measured across 500 US retailers in February 2026.

That 1.5x is why advertisers care so much. A person who asks ChatGPT what to buy and then clicks through is worth half again as much as a normal ad click, because they've already stated their intent in plain language. Criteo is the pipe connecting that intent to an advertiser ecosystem, and by May it reported more than 1,000 brands running ChatGPT campaigns, with holding companies Dentsu, Omnicom, Publicis, and WPP signed on. The most intent-rich ad surface anyone has built, fed by what people thought were private conversations, is now wired straight into the same machinery that runs the rest of the ad industry.

What can you do about it? Less than you'd want, but not nothing. In Settings under Data Controls you can turn off personalized ads, which limits targeting to your current conversation instead of your stored history and Memory, and there's a separate marketing privacy control for the cookie-sharing piece WIRED flagged. A Temporary Chat shows no ads and stores nothing. Two honest caveats: opting out doesn't make ads disappear, it just stops them keying off your chat archive, and per OpenAI's Help Center, ad data can be retained up to 30 days after you delete it. The controls exist, but every default is set to collect, so you have to dig the switches out yourself.

A Choice, Not an Inevitability

The clearest sign this was a choice and not gravity is that OpenAI's competitors looked at the same trade-off and walked away from it. Perplexity, one of the first AI platforms to run ads back in 2024, killed its ad business by February 2026, with executives saying ads were "misaligned with what users want" and risked making people "just start doubting everything." Anthropic spent Super Bowl money on a campaign tagged "Ads are coming to AI. But not to Claude." Altman himself called those ads "so clearly dishonest." So the whole industry was openly arguing about whether to do this, and OpenAI's answer was to hire Meta's ad people and build the surface anyway.

Then there's the document OpenAI signed before any of this. The OpenAI Charter states: "Our primary fiduciary duty is to humanity... but will always diligently act to minimize conflicts of interest among our employees and stakeholders that could compromise broad benefit." An advertising business is a structural conflict of interest, the exact kind the Charter promised to minimize. The company now earns more when it learns more about you, and it learns about you through the questions you trusted it with. OpenAI's own ad principles claim it prioritizes "user trust and user experience over revenue," which is hard to square with a $2.5 billion revenue target built on conversation data.

There's a sign the bet may be costing something already. Adweek, citing Sensor Tower, reported ChatGPT app uninstalls surged 413% year-over-year in March 2026 and 132% in April. The causes are tangled, a Pentagon deal controversy hit in the same window, so I won't pin it all on ads. The timing does line up with the rollout, though, and it suggests at least some users noticed what changed.

The Bottom Line

OpenAI took the most intimate data set any company has ever assembled, the unfiltered questions of 900 million people about their bodies and their money, and turned it into ad inventory. It did this against its CEO's own recorded words and its founding Charter, while two named competitors were publicly refusing to. The math behind the choice isn't hidden. There's $27 billion in cash burn this year, an IPO to sell into by late 2026, and a $100 billion revenue target by 2030 that needs a confessional to feed it.

The open question is what "minimize conflicts of interest" means once the conflict is the business model. OpenAI says ads don't influence ChatGPT's answers, and right now that may be true. But the incentive to make your most trusted source of information also your most profitable ad surface is now permanent, sitting inside the product, waiting on the next quarter's numbers. So the next time ChatGPT recommends something, you have no way to know whether you got the best answer or the best-paying one.