ChatGPT Ads Are Coming: What OpenAI's Leaked Plans Actually Tell Us

In 2024, Sam Altman called ads “a last resort.” His exact words, on the record: “I kind of just don’t like ads that much.” Two years later, OpenAI has a CEO of Applications who launched advertising in the Facebook News Feed, a search underway for a Head of Ads to own all of ChatGPT’s monetization, leaked Android code referencing a “search ads carousel,” and a December report that the ad business is “taking shape.” The man who didn’t like ads is building an ad business.

That’s not a contradiction worth dwelling on. CEOs change their minds; that’s their job. The interesting question is the one hiding behind the headlines: not whether ChatGPT runs ads, but what shape they take — because the shape is the whole story, and the leaked details suggest OpenAI knows it.

How the stance actually moved

It’s worth tracing the pivot precisely, because the arc tells you how seriously this is being taken inside OpenAI.

• 2024: Altman frames ads as a last resort and says plainly he doesn’t like them much.
• Mid-2025: he softens — “I’m not totally against it” — and starts citing Instagram as a model that works.
• October 2025: on Stratechery, he goes the rest of the way. Instagram “changed my mind.” “I actively like Instagram ads.” “Maybe ads don’t always suck.”

A founder talking himself into ads in public is one signal. The hiring is the harder one. In August 2025, Fidji Simo joined as CEO of Applications. She ran Instacart, came up through Meta, and — the detail everyone fixated on — launched ads in the Facebook feed. You don’t hire that résumé to keep not running ads. By September she was reportedly meeting candidates for a Head of Ads for ChatGPT, a role that would own all monetization and report to her. OpenAI also posted for a “Growth Paid Marketing Platform Engineer.” The org chart was being drawn before the product was.

The code knew before the comms team did

Words and hires are intentions. Code is commitment. In late November 2025, ChatGPT’s Android beta — version 1.2025.329 — reportedly shipped with strings no one was meant to read yet: an “ads feature,” “bazaar content,” a “search ad,” and a “search ads carousel.” TechRadar caught it. Strings in a beta build aren’t a launch, and half of them may never ship in that form. But you don’t write a string for a “search ads carousel” unless someone, somewhere, is building a search ads carousel.

Then came the part that shows how thin the ice is. On December 2, users posted screenshots of ChatGPT surfacing an unsolicited Peloton suggestion that looked, to anyone who’s ever seen an ad, exactly like an ad. The reaction was immediate and hostile. Nick Turley, who runs ChatGPT, said the posts showing in-app ads were “not real or not ads” and explained it was a clumsy app-discovery feature, not advertising.

Take him at his word — it probably was a discovery feature, not a paid placement. That’s the point. Users couldn’t tell the difference, and they assumed the worst. That’s the trust deficit OpenAI is walking into. When the actual ads arrive, every relevant suggestion will be read as bought until proven otherwise. The Peloton stumble was a free preview of the credibility problem, paid for with goodwill.

What’s reportedly on the table

Over December 24–25, The Information reported that the ad push is taking shape, and described the formats being explored. None of this is shipped; all of it is in the conditional. But the candidate list is specific enough to reason about:

• Sponsored content woven into the answer. Ask about mascara, and the response surfaces Sephora. The ad isn’t beside the answer — it’s in it.
• Sidebar placements. Mockups reportedly exist for ads off to the side, away from the answer text. The most conservative option, and the most legible.
• Post-click “tell me more” sponsored links. Plan a Barcelona trip, and a follow-up surfaces Sagrada Familia tour packages. The ad attaches to expressed intent after the user has already leaned in.
• “Generative ads,” possibly personalized using ChatGPT’s memory of you.

An OpenAI spokesperson said the company is “exploring how advertising might work… without compromising user trust.” Read those last three words slowly, because they contain the entire fight.

The only question that matters

These formats are not equivalent. They sit on a spectrum, and the spectrum is defined by how close the ad gets to the model’s judgment and the user’s private context.

A sidebar placement is the old deal in new clothes. It’s a labeled box next to the answer; the model’s response is untouched; the user can tell what’s paid. That’s a format people already tolerate, because it doesn’t ask them to wonder whether the answer was bought.

Weaving sponsored content into the answer is a different proposition entirely. The moment Sephora shows up because someone paid for it to show up, the user has to ask a question they never had to ask of a banner: is the model telling me what’s best, or what’s sponsored? Once that doubt exists, it contaminates every answer, paid or not. An answer engine that can’t be trusted to answer straight isn’t worth running ads on, because it isn’t worth using.

And “generative ads” personalized with chat memory is the live wire. People tell ChatGPT things they’d tell a doctor or a lawyer. If that memory becomes targeting data, the product stops being an assistant and becomes a surveillance surface that talks back. The blowback would dwarf the Peloton episode by orders of magnitude.

This is the same line a rival already had to walk. When Perplexity put ads in AI answers, the format it chose — sponsored follow-up questions, with the model still writing the answer — was a deliberate bet that the ad could sit next to the answer without corrupting it. Whether OpenAI draws the line in the same place is the thing to watch.

Where Elo fits

We’ve spent the last two years building an adserver for AI chat — ads rendered as a message inside the thread, matched to live conversation context, brand-safety filtered, and delivered server-side. So the formats OpenAI is reportedly weighing aren’t abstract to us. They’re the exact design space we live in, and the constraints are the ones we’ve had to design around from the start.

Two of those constraints are non-negotiable, and they’re the ones that will decide whether ChatGPT’s ads are tolerated or rejected:

• The ad can’t bias the answer. Selling a placement is fine. Selling the model’s judgment is suicide. The advertiser buys the slot, not the verdict.
• Targeting context isn’t a strip-mine. Conversational intent is the richest targeting signal ever produced — the user literally states what they want. That’s exactly why it has to be handled server-side and in the moment, not warehoused as a personal dossier. The intent in the prompt is plenty; the private memory is off-limits.

None of this is a pitch. It’s just what the format demands, and we mention it only because we’ve already had to make these calls in code rather than in a blog post.

The shape decides everything

ChatGPT ads stopped being an if sometime around the Simo hire. The Head-of-Ads search, the leaked strings, the reported mockups — those are the mechanics of a company that has decided. What’s still undecided is the how, and the how is not a detail. A labeled sidebar is a business. An answer quietly tilted toward whoever paid is a betrayal users will smell, and a memory-mined “generative ad” is a scandal in waiting.

OpenAI gets to pick which of those it ships. The rest of us get to watch whether it picks the version users can live with. That choice — not the launch date, not the revenue — is the one worth waiting for.