In July 2026, OpenAI announced it would fold its safety team back into the general research team. At the same time, Johannes Heidecke — the head of safety for the past two years — confirmed he was leaving the company. For those who follow OpenAI closely, the news has an uncomfortable déjà vu: this had happened before.
In 2024, OpenAI dissolved its Superalignment team — the group dedicated to AI's long-term risks — and reintegrated its members into general research. Ilya Sutskever, one of the co-founders, left during that process. Jan Leike, the team's head, published an open critique saying safety had taken a back seat to products. History seems to be repeating itself.
What exactly happened: Heidecke's exit and the team merger
Johannes Heidecke led what OpenAI called "model safety" — the process of evaluating whether a model released to the public could be used to generate harmful content, weapons instructions, psychological manipulation or any other use OpenAI considers out of bounds.
His team's job was to act as an independent brake: before a model shipped, they evaluated it without product or business pressure influencing the outcome. That structural independence is exactly what is lost when the safety team sits inside the research team, which has direct incentives around launch speed.
OpenAI hasn't publicly explained the reasons for the merger beyond general phrases about "integrating safety work into the research process from the start". It sounds reasonable. The problem is that we had heard it before, and the result was the dissolution of the Superalignment team.
The repeating pattern: launch pressure vs. safety independence
To understand why this matters, you have to understand how safety teams at AI companies work when they are truly effective: they must be able to say "no" to a launch, even if the model is technically ready, even under competitive pressure, even if it means a competitor ships first.
That veto power requires structural independence. A safety team that reports to the same leader as the research team has an inherent conflict of interest: the head of research has incentives to ship models, and the safety team's incentives end up subordinated to that.
The pattern at OpenAI is this: an independent safety team is created with its own autonomy and presented as a commitment to responsible AI; then — when competitive pressure rises — that team is folded back into general research, the safety leaders leave, and the cycle starts again with the next announcement of a commitment to safety.
What this means for the real safety of OpenAI's models
There's an important distinction: merging the teams doesn't necessarily mean OpenAI's models will be more dangerous tomorrow. Safety evaluations still exist within the research process. The problem is that they lose the structural guarantee of independence.
What changes in practice:
- Safety evaluations no longer have a team with a formal, independent veto over launches
- Safety researchers have managers who are also responsible for launch deadlines
- Conflicts between "the model needs more evaluation" and "we need to ship before the competition" are resolved within the same chain of command
- Specialized safety talent that values that independence may choose to leave — as Heidecke did
None of these changes is invisible to the public. The consequences, if any, will show in the models released over the next 12–18 months and in the incidents those models cause.
Implications for companies that rely on ChatGPT
If your company uses ChatGPT for customer service, content generation, data analysis or any critical process, this news doesn't require immediate action. Current models keep working the same way. But there are medium-term considerations worth keeping in mind.
Diversifying AI providers becomes more relevant. Companies that depend on a single provider for critical processes are exposed to policy changes, changes in model behavior or even service interruptions. Anthropic (Claude), Google (Gemini) and Meta (Llama) offer alternatives with different governance profiles.
Provider safety evaluations should be complemented with your own. However robust OpenAI's internal safety process is, companies that integrate AI into critical processes should have their own validation layers: content filters, human review of high-impact decisions, audit logs.
Models' behavioral track record is still the best indicator. More than OpenAI's internal organizational structures — which the public can't verify directly — documented incidents with models in production are the most reliable evidence of how they really behave.
Should you worry as an individual user?
Probably not in the short term. OpenAI's safety evaluations, although now integrated into research, remain more rigorous than those of most of the industry. The models in production today went through robust evaluation processes.
The legitimate concern is long-term and systemic: if competitive pressure makes evaluations faster and less independent across the industry, the risk of shipping models with unexpected behaviors rises. It isn't a risk for tomorrow, but it isn't theoretical either — we've already seen models that pass evaluations and then show problematic behavior in real use.
Alternatives, and what other companies do differently
Anthropic, the company behind Claude, has so far kept a structure in which safety carries significant weight in launch decisions. Anthropic's commitment to long-term safety is explicitly part of its founding mission, not a department.
Google DeepMind has its own safety team with relative independence within Alphabet's structure. Meta released Llama as open source, which transfers part of the safety responsibility to whoever deploys the model.
No model is perfect and no company has fully solved AI safety. But organizational structure matters: it says something about how likely a company is to sacrifice safety for speed when the market pushes.
Frequently asked questions
What is OpenAI's safety team and what is it for?
OpenAI's safety team is the group responsible for evaluating models' risks before public release — from the ability to generate harmful content to long-term alignment risks. Its function is kept separate from performance research so that commercial pressure doesn't compromise safety evaluations.
Does this affect the security of my data in ChatGPT?
Not directly. Merging OpenAI's safety and research teams concerns the safety of the model's behavior (not generating harmful content, aligning with human values), not the security of user data. Data privacy in ChatGPT is governed by OpenAI's privacy policy, which is separate from this restructuring.
What's the difference between AI safety and data security?
AI safety is about the model's behavior: not producing weapons instructions, not emotionally manipulating users, reasoning predictably. Data security is about where your information is stored, who can access it and how it's protected. They are two different fields with different teams at any AI company.