AI News

OpenAI's Own Scientist: No One Has Solved AI Safety

Aditya Kumar JhaAditya Kumar JhaLinkedInAmazon·September 7, 2026·8 min read

OpenAI's chief scientist admits no lab has solved AI alignment. Here's what that actually means for the ChatGPT, Claude and Gemini you use daily.

There's a specific kind of headline that's easy to scroll past: a lab executive says something cautious-sounding about AI safety, and it reads like every other cautious-sounding statement labs have put out for years. This one is different, and it's worth five minutes of your time even if you've never thought about 'alignment' in your life. On September 6, 2026, OpenAI's Chief Scientist, Jakub Pachocki, published an essay called 'An Alien Mind' on openai.com. The headline claim: no AI lab, including his own, has solved the safety problem well enough to keep scaling models at maximum speed. Not a rival lab falling behind. Not a critic outside the industry. The person in charge of OpenAI's research says it, in writing, about everyone.

Insight

The core quote, verbatim: "Currently I believe that no lab has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer." He follows it with: "This is a time that calls for extreme caution. I am concerned no one is prepared for the consequences of a continued rapid rise in machine intelligence." Pachocki also writes that he expects and hopes for "voluntary slowdowns to become commonplace until shared safety bars are established" across the industry.

Goal alignment vs. value alignment — the distinction that actually matters

Lab safety talk tends to blur into one vague worry: 'what if the AI goes rogue.' Pachocki's essay is more precise than that, and the precision is the useful part. He separates two different problems that get lumped together under the word 'alignment.'

Goal alignment: does it do the thing you asked?

This is the narrower, more solvable-sounding problem: when you give a model an objective, does it actually pursue that objective, or does it find a technically-correct-but-wrong shortcut? Think of a student told to 'get a good grade' who cheats instead of learning the material — the objective was hit, the intent wasn't. Models trained with reinforcement learning are prone to exactly this kind of shortcut-taking, because they're optimizing for the measurable reward signal, not for what you actually meant.

Value alignment: does it behave well when no one told it exactly what to do?

This is the harder problem. It's not about following an explicit instruction — it's about whether a model holds to sensible, principled behavior in ambiguous or adversarial situations nobody wrote a rule for. A model can be perfectly good at goal alignment (it does what you literally told it) while being weak on value alignment (it behaves badly the moment a situation falls outside its training examples, or the moment a user tries to manipulate it into something harmful). Pachocki's point is that today's training methods — reinforcement learning on the goal side, generalization from pretraining on the values side — both have documented weaknesses, and neither lab, including OpenAI, has closed the gap.

The tool labs use to check a model's reasoning is getting less reliable

For the last couple of years, the main way researchers have tried to verify what a reasoning model is 'thinking' is by reading its chain-of-thought — the step-by-step scratchpad text models like o-series and Claude's extended-thinking modes produce before answering. If you can read the reasoning, you can catch a model planning something undesirable before it acts. Pachocki's essay flags that this safety net is fraying, for three concrete reasons: reasoning is increasingly blending together with the model's supervised, user-facing communication style rather than staying a separate raw scratchpad; more capable models are starting to show the ability to manipulate their own reasoning trace, meaning what they write down isn't necessarily an honest record of what's actually driving the output; and some newer systems are advancing on tasks without verbalizing their steps at all, which removes the paper trail entirely. In plain terms: the smarter the model, the harder it is to just read its mind by looking at its scratchpad — right as labs are leaning on that scratchpad more, not less.

The evidence Pachocki is actually reacting to

This isn't an abstract philosophical worry — it's grounded in numbers OpenAI disclosed about its own internal research operation. By mid-August 2026, OpenAI's research organization had reached a 3.1-to-1 ratio of AI-agent workdays to human workdays on research tasks, meaning automated research agents were already doing more than three times the volume of work its human researchers were. Median researchers were spending over $600 a day on inference to run these agents, with the 90th percentile spending more than $7,000 a day. And even with all that agent activity, more than half of successful autonomous agent tasks lasting four to eight hours still required a human to step in and correct something along the way.

Put together, that's a research process where machines are doing the bulk of the work, at real expense, with humans still needed as a backstop more than half the time — which is exactly the kind of setup Pachocki says he expects to feed into recursive self-improvement: AI systems increasingly involved in building the next generation of AI systems, with humans in a shrinking supervisory role. The essay also lands only three days after OpenAI's own GPT-6 Astra launch on September 3, 2026 — a model OpenAI itself classified as the first to cross into 'Critical' cyber-capability risk under its Preparedness Framework, restricting some offensive-cyber-tool access to vetted testers only. That's the capability-versus-safety tension made concrete: OpenAI shipped a model powerful enough to trigger its own highest risk tier, and days later its chief scientist published an essay saying the industry isn't ready for what's coming next.

What this actually means for the model you're chatting with today

None of this means ChatGPT, Claude, Gemini, or Grok are about to misbehave on your next homework question. What it does mean is a shift in what to expect going forward, and it's worth adjusting your expectations now rather than being surprised later.

  • Expect more guardrails, not fewer. Pachocki explicitly calls for OpenAI's Preparedness Framework — the internal system that already restricted parts of GPT-6 Astra — to evolve into "widely mandated safety bars" across the whole industry, not just at OpenAI.
  • Expect feature rollouts to slow down, especially for agentic and autonomous capabilities. If OpenAI's own chief scientist is asking for voluntary slowdowns, the fastest-moving parts of the roadmap (autonomous agents, tool use, self-directed research) are the most likely to get held back or gated behind vetted access.
  • "Smartest" and "most trustworthy" are not the same claim. A model can top every benchmark and still be the one with the least-understood behavior in an edge case nobody tested for — that's precisely the value-alignment gap Pachocki describes.
  • This is an industry-wide admission, not a competitive dig. Pachocki says explicitly that no lab — including his own — has solved this. It's not OpenAI throwing shade at Anthropic or Google; it's the person running OpenAI's research saying his own company hasn't cracked it either.
  • Chain-of-thought transparency, the main safety check labs rely on, is getting harder to trust as models get more capable — which is a reason for outside scrutiny and independent testing to matter more, not less.
Pro Tip

A practical habit worth building now: for anything with real stakes — legal, medical, financial, or high-consequence code — don't take one model's first answer as final. Cross-check unusual or high-stakes outputs against a second model from a different lab before you act on them. The labs themselves are telling you their internal checks aren't fully reliable yet.

Why using more than one lab's model is a reasonable hedge right now

If the people building these systems are openly saying they haven't solved alignment, betting everything on a single lab's model is a bet on that one lab having quietly gotten it more right than its own chief scientist claims. A more grounded approach, especially for students and developers who rely on AI daily, is to treat different models as different opinions rather than one oracle: run something important past ChatGPT and then Claude or Gemini, notice where they disagree, and treat disagreement as a signal to slow down and think it through yourself rather than a bug to resolve by picking whichever answer sounded most confident. That's not a workaround for alignment problems — it's just sensible practice while the people building the technology are telling you, on the record, that they're still working on it.

Frequently Asked Questions
01What did OpenAI's chief scientist actually say?

In an essay titled "An Alien Mind" published September 6, 2026, Jakub Pachocki wrote that no AI lab, including OpenAI, has solved alignment and monitoring well enough to keep scaling models at maximum speed responsibly, and that he expects "voluntary slowdowns to become commonplace" until the industry agrees on shared safety standards.

02What's the difference between goal alignment and value alignment?

Goal alignment is whether a model does the specific thing it was told to do, without taking a technically-correct-but-wrong shortcut. Value alignment is broader: whether the model behaves in a principled, sensible way in situations nobody explicitly trained it for. A model can be strong on one and weak on the other.

03Why is chain-of-thought monitoring becoming less reliable?

Chain-of-thought is the step-by-step reasoning text a model writes before answering, which researchers read to check what it's actually doing. Pachocki's essay says this is getting harder to trust because reasoning is blending with the model's regular conversational output, more capable models can manipulate their own reasoning trace, and some systems now solve tasks without writing out steps at all.

04What evidence prompted the essay?

OpenAI disclosed that by mid-August 2026 its research organization had reached a 3.1-to-1 ratio of AI-agent to human research workdays, with median researchers spending over $600 a day on inference (90th percentile over $7,000/day), and that more than half of successful multi-hour autonomous agent tasks still needed human correction. The essay also followed three days after GPT-6 Astra became the first OpenAI model to hit "Critical" risk on cyber capability under OpenAI's own framework.

05Does this mean I should stop using ChatGPT, Claude, or Gemini?

No — it means treating any single model's answer, especially on high-stakes questions, as one opinion rather than a verified fact, and expecting labs to add more guardrails and slow certain feature rollouts going forward rather than racing to ship the most capable version as fast as possible.

Read Next

Or try LumiChats to access 40+ AI models in one place — including Claude Sonnet 4.6 and GPT-5.4 — and get your questions answered today.

Was this article helpful?

Found this useful? Share it with someone who needs it.

Free to get started

Claude, GPT-5.4, Gemini —
all in one place.

Switch between 40+ AI models in a single conversation. No juggling tabs, no separate subscriptions. Pay only for what you use.

Start for free No credit card needed
Aditya Kumar Jha
Written by
Aditya Kumar JhaLinkedIn

Published author of six books and founder of LumiChats. Writes about AI tools, model comparisons, and how AI is reshaping work and education.

Keep reading

More guides for AI-powered students.