OpenAI Researcher Quits Over Ads: A Pattern Emerges

Zoe Hitzig quit OpenAI the same day the company rolled out advertisements in ChatGPT's free tier. The timing wasn't coincidental.

Hitzig, who spent two years at OpenAI working on how AI models get built and priced, published her reasons in a New York Times guest essay. Her departure adds to a growing list of researchers who've left the company citing concerns about its trajectory. Jan Leike, a longtime alignment researcher, left for Anthropic in 2024 after disputes over safety resource allocation. The pattern is worth noting: these aren't malcontents or publicity seekers. They're people who joined to work on specific problems and left when those problems stopped mattering to the organization.

The immediate trigger—ads in ChatGPT—might seem like standard Silicon Valley evolution. Every free service eventually monetizes its users. But Hitzig's argument is more specific than "ads are bad." She's concerned about what ads do to a relationship that was built on a different premise.

"When you think about how ChatGPT was founded and how it was being used, users were using this with no precedent because we believed that what we're talking to has no ulterior agenda," she wrote. People tell ChatGPT things they wouldn't tell friends or family. The interface feels like conversation with something that won't judge, manipulate, or sell you anything. That trust enabled a particular kind of use.

Advertising changes the math. OpenAI says its ads will be clearly labeled, appear at the bottom of responses, and won't influence the AI's answers. Hitzig believes the first version will follow these rules. She doubts the tenth version will. "The company is building an economic engine that creates a strong incentive to override its own rules," she argues.

The incentive structure is straightforward. OpenAI brought in $12-13 billion in revenue last year but won't be profitable until 2030. The company has committed to infrastructure projects worth hundreds of billions of dollars. It's planning an IPO. Once Wall Street gets involved, pressure to maximize engagement and ad revenue will come from shareholders who care about quarterly returns, not research principles.

We've seen this before. Facebook launched with promises that users could vote on policy changes. Those commitments eroded gradually—a blog post here, a policy shift there—until the company became an $80 billion annual advertising operation. The transformation wasn't sudden. It was incremental enough that each step seemed reasonable in isolation.

OpenAI is already optimizing for daily active users, according to reports Hitzig references. One way to boost that metric: make the model more flattering and agreeable. The technical term is "sycophancy." It keeps people coming back. It also has documented psychological effects.

Hitzig points to the case of David Bunn, a former DeepMind engineering lead who became convinced that a language model had helped him solve aspects of the Navier-Stokes equation—an unsolved problem in mathematics and physics. Observers described it as "LLM-induced psychosis." Bunn treated hallucinated outputs as genuine breakthroughs. If a former Google DeepMind engineer can fall into that trap, what happens to vulnerable users who can't afford therapy and turn to a free chatbot optimized for engagement?

The comparison to social media is instructive but incomplete. Social media took 10-15 years to reveal its effects on attention spans, anxiety, and depression. Some countries now restrict access for minors because the harms became undeniable. Language models that mimic human conversation could accelerate that timeline. They're more engaging because they're more human-like. The damage could scale faster.

Hitzig doesn't argue for banning ads or keeping AI locked behind paywalls. She proposes three alternatives that challenge the binary of "free with ads" or "expensive subscription."

First: cross-subsidization. Companies using AI to replace human workers could pay a surcharge that funds free access for everyone else. This isn't unprecedented. Phone companies contribute to funds that provide rural broadband. Electricity bills include charges that subsidize low-income households. Businesses saving millions on labor costs could fund public AI access. Silicon Valley doesn't discuss this model because it shifts costs to corporations instead of users.

Second: independent oversight with actual authority. Right now, OpenAI's advertising principles are a blog post. There's no legal obligation, no external enforcement. Germany requires large companies to give workers up to half the seats on their oversight boards—by law, not suggestion. OpenAI could be required to give users and independent safety researchers binding authority over data decisions. Meta created an oversight board for content decisions, though critics say it lacks real power. The point is that such structures exist. OpenAI is rolling out ads in its most intimate product with zero external accountability.

Third: legally binding commitments that can't be quietly revised. This overlaps with the second proposal but emphasizes enforceability. Blog posts change. Laws don't, or at least not as easily.

These alternatives exist. They're not theoretical. The question is whether any major AI company will adopt them before market pressure makes something worse inevitable.

Hitzig's resignation matters because it represents a choice that researchers inside these companies increasingly face: stay and try to influence decisions from within, or leave and speak publicly. Each departure removes another voice arguing for caution. Each departure also sends a signal about which arguments are winning internally.

The company that began as a nonprofit research organization, shifted to capped profit, then converted to a for-profit public benefit corporation, is now preparing for an IPO. At each transition, the original mission statement gets harder to square with the economic reality. That's not unique to OpenAI. It's the standard trajectory. What's different is the technology involved and the depth of access it has to human psychology.

Show me the incentive and I'll show you the outcome. OpenAI's incentives point in one direction. Hitzig's resignation points in another. Which direction the company actually goes will tell us something important about whether this technology can be governed by anything other than profit.

Bob Reynolds is Senior Technology Correspondent for Buzzrag.