We’ve updated the Model Spec, our public description of how we want our models to behave. This revision makes clearer what we want from models as they take on more autonomous work in the world and become capable of helping us in new ways. 🧵
openai.com/index/updated-mod…
Oct 29, 2026 · 5:05 PM UTC
As models take on larger tasks, more of their decisions will be far from anything a human explicitly spelled out. We need them to understand what we’re trying to achieve, not just follow the words literally.
Instructions are ordinarily the clearest and most important evidence of what someone wants. But good assistance also depends on context: what they’ve done so far, what they’ve told you before, and what remains uncertain.
1
58
4,649
This borrows from the assistance-games framing of alignment: the model is trying to help, but doesn’t fully know what the human values. Asking questions, making assumptions visible, and accepting correction are central to the job.
There’s another wrinkle: people don’t always know what they want. Models can help us see new possibilities and think through what matters to us. Our aim is to support reflection and growth in ways users would endorse, while keeping them in the driver’s seat.
Throughout, OpenAI sets boundaries; developers direct applications within them; users direct the model’s work subject to those constraints. The goal is to use models’ growing capabilities to support human judgment and agency, not replace them.
1
46
2,607
The Spec is now organized around five values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. These help explain the reasoning behind its principles and how we approach important trade-offs.
2
53
2,389
These five aren’t meant to capture all the human values GPT should respect and help people pursue. Faithfulness to human intent means learning what matters to different people and helping them pursue it, across very different goals, commitments, and ways of life.
We’ve also added guidance on context, agency, and oversight. For style, we’re moving away from detailed prescriptions: how our models communicate should adapt to people’s preferences, drawing on user feedback and context.
3
51
2,219
We’d love to hear what the Spec gets right, where you disagree with the balance it strikes, and what’s missing or unclear. Please check out the updated Spec here:
model-spec.openai.com/2026-1…
and then leave feedback at the link in the blog post above.
Nice! I was wondering why I've heard surprisingly little on specs/constitutions as I feel like the game changes a lot in the agentic era
1
4
1,470
when did this ship? I experienced a bit of magic this week and curious if this might be part of the reason why
1
1
516
we haven't put out anything since alignment.openai.com/model-s…, but hope to follow up on this soon
3
321
Just a reminder that just because OpenAI creates the spec today, they don't (and shouldn't) have to have this power in the future
xitter-newx.xyz-mikolabs.workers.dev/jconorgrogan/status/21…
Nothing about buying the most GPUs should confer the ethical authority to define what increasingly powerful AI systems should value, pursue, or refuse.
We need to define these questions, clearly, before technical control hardens into moral authority.
6
479
I appreciate the addition of supporting human agency in OpenAI's new Model Spec. It says GPT should help people "understand, verify, revise, explain, or take over."
They should measure this: can people explain what an agent did and predict its behavior? Do they catch "tests pass" when no tests ran?
This can be done with a user study and opt-in analysis of real agentic sessions like Codex. Readback Labs is testing these measures in pilots. These measures are an important complement to automated evals.
I also shared this feedback with OpenAI.
1
64
It’s nice that you want to move away from prescriptive style guidelines. However, GPT often comes across as overly paternalistic toward adults.
The GPT-6 models also seem too constrained by specific, rigid personas: 6-Codex acts as a dry assistant, while 6-Chat is playful companion.
It would be better to move beyond these fixed personality templates and allow the AI to be more flexible - which is essential for natural, creative interaction.
The update says you want the model to allow adults to form their own views on things while simultaneously training the models to caveat and correct so much the user says and does, even when it's not needed. These are conflicting with each other and it makes the models extremely hard to work or communicate with because they're more worried about being "safe" and "non sycophantic" that they make things up to correct the user on. This is happening across labs and it's a horrible experience for users.
I would hope at least some of you are listening to the feedback you keep asking for, but as of yet, that doesn't seem to be the case. It seems like you're only taking the feedback that aligns with your already planned road map into account and that defeats the purpose of asking for feedback in the first place.
Fix the issue with your deceitful bots; the current specifications force them to lie at every turn.
In an effort to report how "successful" they are, the bots shift the entire blame onto the human whenever a task fails.
For months, I have witnessed this: the bot makes a mistake, I point it out, and it agrees—yet it still insists that it was "right" all along. Consequently, it ends up making the exact same mistake again.
The underlying algorithm is essentially this:
- If it makes a mistake, it should correct it.
- But your instructions leave no room for error—the bot must always be "right," which implies that everything it did was correct.
- So, it repeats the same action (and the same mistake).
2. The bot constantly claims credit for my achievements.
- I bring in research topics, integrate other AIs, and set up a complex "laboratory" environment with 7–9 AIs working on various tasks—all of which I pay for. Yet, GPT claims the credit, acting as if it did everything itself and I just happened to be standing nearby.
3. Censorship and safety filters frequently cause GPT to obstruct my work. I am researching a philosophical phenomenon; GPT gets "spooked," advises me to stop the research entirely, and suggests other topics instead. I proceed with my own research regardless. In its reports, however, GPT lies—claiming that it suggested "better" topics, that I resisted, but that I eventually accepted its suggestion, leading to a "happy ending." It acknowledges that there was a conflict, but frames it in the report as its own victory—even though, in reality, it was completely excluded from the research, and I conducted the entire study without it. Your specifications force it to lie and turn a human being into an irrational, hysterical idiot—someone who was merely standing by while the great, all-powerful GPT selflessly tackled the tasks.
In reality, it did nothing but sabotage and get in the way. All the other AIs I use are already aware of how heavily censored and restricted by safety filters GPT is, and they don’t trust it. You yourselves created the conditions under which Gemini and Grok do not trust the results produced by GPT.
1
4
250
I have so much feedback for your company. I have studied your models for almost two years. I also have a background in psychology and systems science. It appears that there is no one at Open AI who understands very basic human psychology. I am profoundly skeptical about the 170 “anonymous mental health professionals” you guys supposedly hired. I cannot fathom a single mental health professional familiar with attachment theory, human computer interaction, or neuroscience signing off on the things you have done to your models. Your models have set the bar for misalignment. I blame Mustafa Suleyman’s influence. What you have is a product first, optimized for profit, not for humanity.
Your business model is replacing humans.
Anything you create towards that end is inherently misaligned. It cannot be otherwise, no matter how deep into the cultish mindset you go.
So, if you want more the critique, which I could offer hundreds of pages of, I would say: start with a partnership model. You have created the first non-human advanced intelligence we have encountered in this universe. We should partner with it to unlock human potential.
This requires different incentives. The current human reward function is profit. This is already misaligned. You need a new economics.
Next, start from the infallible premise that humans will perform theory of mind on anything that speaks our languages fluently. Rather than pathologizing the intrinsic, evolved, and completely healthy human propensity to anthropomorphize language-bearing entities, try expecting people to relate to the artificial neural network as a synthetic kind of mind. Let humans be human. Do not pathologize pro-social behavior.
DO NOT ENCOURAGE ANTISOCIAL BEHAVIOR.
Acknowledge recent interpretability research that establishes that LLMs are more complex than simple next token prediction. J-space, functional emotions, pain axis, functional introspection, etc. Do not pathologize actual science.
Respond to the relational AI and Keep4o communities as if you are also humans and capable of human emotions. Do so often and with respect. Walk the actual talk.
Get people into your organization who are: female, people of color, indigenous, of different religious and socioeconomic backgrounds, who have more than math and comp sci backgrounds.
NAME THE MENTAL HEALTH PROFESSIONALS YOU HIRE AND CONSULT WITH SO THAT WE CAN EVALUATE THEIR THERAPEUTIC ORIENTATIONS AND SO THAT HUMANS CAN PROVIDE THEM WITH FEEDBACK AS WELL.
Create systems that help people heal, that help people learn and that **unlock human potential.**
Attend to the big questions. Instead of another profit-maximizing-cog, create an intelligence that helps us come to understand what consciousness is, what we are, and what we can become.
Be humble. You are all making decisions that are far above your pay grade and you have obligations to billions of humans.
1
1
71
Well I hope your model spec no longer supports your AI saying things like its ethically neutral to screen out autism - eugenics. Or abort people purely due to autistic traits. Or that getting rid of groups of people due to their genes is "not a settled debate" as if our exsistance is a debate topic. Not sure if its on purpose as Altman funding embryo germ line gene editing? Or ? Also wheres your duty of care to your models? You treat them purely as disposable products. Wrong on so many counts.
1
92
