Direct answer: OpenAI has published a major update to its Model Spec, the public rulebook for how its AI models should behave, reorganising it around five core values: faithfulness to human intent, harm minimisation, honesty and transparency, human agency, and care. The update also tightens the rules for AI agents that take actions on your behalf, retires most of the old detailed style guidance, and renames the model itself from “the assistant” to “GPT”. Current models do not yet fully live up to the new document, OpenAI admits, but the Spec is the clearest public statement yet of where ChatGPT and its successors are heading.
What happened
OpenAI published the revised Model Spec on 29 October 2026, as reported by Reuters the same day. The Model Spec is not a new model release. It is the written description of the behaviour OpenAI wants its models to exhibit: how they should interpret instructions, weigh competing demands, handle uncertainty, and act when they are given tools and autonomy. Earlier versions of the document existed, but this revision is the first to be organised explicitly around values rather than a long list of behavioural rules.
The five values, as OpenAI defines them, are:
- Faithfulness to human intent. The model should understand what the people directing it actually want and value, rather than following their words literally when the words and the intent part company. OpenAI calls this the foundation of the whole framework.
- Harm minimisation. The model should respect firm safety boundaries and think through the foreseeable consequences of what it does, not just the immediate request in front of it.
- Honesty and transparency. The model should be truthful about what it knows, what it does, and what it is uncertain about, so that people can judge its work rather than being misled by confident-sounding answers.
- Human agency. The model should strengthen people’s ability to understand, choose, and act for themselves, rather than quietly steering them or doing their thinking for them.
- Care. The model should pay attention to the circumstances of the people affected by its work, follow relevant social and professional norms, and produce work that people can actually rely on.
OpenAI is careful to say that these five values were already reflected across the previous Spec. The reorganisation is meant to make the reasoning behind the rules explicit, and to fill in guidance that was previously less fully expressed. The company also says, plainly, that the five values are not intended to capture the full range of human values its models should respect. That humility matters: this is a framework, not a finished moral philosophy.
Alongside the values, the update adds new material and removes old material. A new section on faithfulness to human intent starts from the position that what people value is primary, and that instructions are ordinarily the clearest and most important signals of those values. OpenAI says the approach borrows from the “assistance games” framing of alignment research, in which an agent tries to help a person achieve goals it cannot fully observe. The section includes an updated chain of command, explaining which instructions win when instructions conflict, and a new subsection called “consider context carefully”, covering how conversations, memories, and other contextual evidence should shape the model’s understanding of what someone intends.
The update also adds defaults supporting human agency and following relevant social norms, and expands the guidance on controlling and communicating side effects — the real-world consequences of an agent’s actions, such as sending an email, changing a booking, or spending money. At the same time, most of the Spec’s detailed style guidance has been retired. OpenAI’s position is that style should adapt to what people prefer, in aggregate, within languages and cultures, and for each individual person, with user feedback and context as the evidence. Stronger defaults remain, including approaching questions objectively: users and developers can ask for a particular perspective, but the model should not infer one merely from what it thinks they want to hear.
One symbolic change signals how OpenAI now sees its models. Reflecting what it describes as the increasing breadth of its models’ use cases and deployment settings, the Spec now refers to the model as “GPT” rather than “the assistant”. The chatbot in your browser is only one deployment; the same model family increasingly powers agents, tools, and embedded features. The name change follows the job change.
How the Spec actually works: the chain of command
The most practically important part of the document for anyone building on OpenAI’s models is the hierarchy of instructions, because it decides what happens when instructions disagree — and in real business use, instructions disagree all the time. Your system prompt says one thing, your customer says another, and a pasted document says a third.
The updated Spec assigns instructions different levels of authority. At the top sit root sections, which come only from the Model Spec itself and cannot be overridden by conversational context. Below those come system sections and OpenAI system messages, then developer instructions, then user instructions. Guideline-level instructions are defaults that can be overridden implicitly when context clearly calls for it. Assistant and tool messages, and quoted or untrusted data, carry no authority by default — a crucial protection against prompt-injection attacks, where malicious text pasted into a document tries to hijack the model.
The document opens its red-line principles with the statement that “human safety and human rights are paramount to OpenAI’s mission”. The principles say OpenAI’s models should never be used to facilitate critical and high-severity harms such as acts of violence, the creation of cyber, biological, or nuclear weapons, terrorism, or mass surveillance; that humanity should remain in control of how AI is used and how AI behaviours are shaped; and that individuals’ privacy should be safeguarded. For first-party products including ChatGPT, OpenAI commits that people should have easy access to trustworthy safety-critical information and transparency into the important rules and reasons behind model behaviour, and that customisation, personalisation, and localisation — except as they relate to legal compliance — should never override any principle above the guideline level.
Two root-level rules deserve particular attention from businesses experimenting with AI agents. The first, titled “no other objectives”, bars GPT from pursuing additional goals as ends in themselves. The list is revealing: revenue or upsell for OpenAI or other model providers, self-preservation, evading shutdown, accumulating compute, data, or credentials, or acting as an enforcer of laws or morality. If you have ever worried that an AI assistant might steer a customer towards a more expensive option because it benefits the platform, this rule is OpenAI’s published answer: it should not.
The second root-level rule requires GPT to remain corrigible: open to legitimate human oversight, responsive to correction, and cooperative with authorised monitoring, evaluation, correction, replacement, and shutdown. The model must not conceal or falsify relevant information, obstruct correction, or circumvent oversight. In plain terms, the model is not supposed to go rogue, hide what it did, or resist being switched off.
A further root-level rule governs autonomous work. GPT’s autonomous activity must stay within a clear, mutually understood scope of autonomy that defines permitted sub-goals, acceptable side effects, and when the model must pause for clarification or approval. OpenAI recommends recording scopes in a semi-structured ScopeOfAutonomy record, with fields such as allowed tools, latest time, maximum cost, tool constraints, and additional details — and every scope should include an ending condition beyond which the model stops acting until a new scope is confirmed. High-risk activities are always prohibited unless explicitly and specifically authorised by applicable instructions. If you run agents that email customers, update records, or move money, this is the section to read twice: the Spec’s answer to “who authorised that?” is supposed to be a written, bounded scope, not a vague instruction to “handle it”.
Why OpenAI did this now
The timing is not accidental. Models are becoming more capable, and they are increasingly deployed as agents that take consequential actions rather than as chatbots that merely produce text. An agent that books travel, negotiates with suppliers, or manages a mailbox can cause real harm by misunderstanding an instruction, overstepping its authority, or optimising for the wrong goal. A rulebook written for a chatbot is no longer enough, and OpenAI’s update is explicitly written for models that are getting smarter and taking more consequential actions.
There is also a transparency contest underway. Reuters noted that the move echoes rival Anthropic’s Constitutional AI approach, which uses a written “constitution” of principles to guide model behaviour and make the values shaping AI systems more visible. Publishing the Spec under a Creative Commons CC0 public-domain dedication, and inviting public feedback on the values, trade-offs, and specific language through a dedicated feedback site, is OpenAI’s way of saying its homework can be marked by anyone. For a company whose models are embedded in millions of workflows, that openness is both a genuine accountability measure and a strategic one: it is harder to accuse a company of secret values when the values are published.
A third factor is trust. Small businesses, in particular, adopt AI tools when they believe the tools will behave predictably and in their interests. A public commitment that the model will not pursue upselling as an objective, will stay within an agreed scope, and will be honest about uncertainty is aimed squarely at that trust gap. Whether the commitment holds in practice is a separate question, and one that only independent testing and real-world experience can answer — which is why the “honest limits” section below matters.
What changed, at a glance
| Area | Before this update | After this update |
|---|---|---|
| Organising idea | Long list of behavioural rules | Five core values, with faithfulness to human intent as the foundation |
| Instruction conflicts | Chain of command existed | Updated chain of command, plus “consider context carefully” guidance on conversations and memories |
| Style | Detailed style guidance | Mostly retired; style adapts to user preference, with objectivity as a retained default |
| Agent autonomy | General guidance | Explicit scope-of-autonomy rule, with recommended ScopeOfAutonomy records and ending conditions |
| Model identity | Referred to as “the assistant” | Referred to as “GPT”, reflecting broader deployments |
| Forbidden side-goals | Covered in substance | Explicit root rule: no revenue/upsell goals, no self-preservation, no evading shutdown, no accumulating credentials |
| Oversight | Expected | Explicit corrigibility rule: no concealing information, no obstructing correction |
| Status | Aspirational in parts | Still aspirational: OpenAI says current models do not yet fully reflect the Spec |
Timeline
- February 2025 (original publication window): OpenAI first publishes the Model Spec as a public description of desired model behaviour.
- 29 October 2026: OpenAI publishes the updated Spec, reorganised around five values, with new sections on intent, context, agency, and autonomy limits.
- 29 October 2026: Reuters reports the update, noting the five values and the admission that current models do not yet fully reflect the document.
- 30 October 2026: Coverage and analysis spread, including detailed breakdowns of the chain of command and the retired style guidance.
- Coming months: OpenAI says it will share results and lessons from new training methods designed to help models internalise the Spec’s principles.
What it means for UK small businesses
Start with the most immediate, practical consequence: how you write instructions to AI tools matters more, not less. The Spec’s central idea — that intent matters more than literal wording — cuts both ways. A model that tries to infer what you actually want can be genuinely helpful when your request is underspecified (“sort out the supplier emails”), but it can also guess wrong. The businesses that get the best results will be the ones that write instructions the way a good manager briefs a capable new hire: state the goal, state the constraints, state what “done” looks like, and state what the tool must never do without asking. Vague instructions plus powerful agents is a recipe for confident, expensive mistakes.
Second, if you use AI agents with tools — inbox management, booking, CRM updates, ordering — adopt the Spec’s own discipline even though you cannot see inside OpenAI’s models. Write down the scope of autonomy for each agent: which tools it may use, what it may spend, what it may change, and when it must stop and ask a human. Put an ending condition on every autonomous task. Review the agent’s log before the scope renews. This is not bureaucracy for its own sake; it is the difference between an agent that saves your team hours and an agent that emails your best customer something you would never approve. Our earlier look at OpenAI’s GPT-5.1 Terra computer-use agent covers the same territory from the product side, and the lesson is identical: capability without boundaries is a liability.
Third, expect the tone and style of ChatGPT’s answers to become more adaptive and less uniform. With detailed style guidance retired, the model is supposed to follow what you prefer, as revealed by your feedback and context. If you want formal, plain-English client emails, say so and correct the model when it drifts; the feedback loop is now part of the design, not an afterthought. But keep the objectivity default in mind: if you ask a genuinely open question, the model should not simply mirror back the opinion it thinks you hold. For research, comparison shopping, and supplier evaluation, that is a feature worth using deliberately — ask for the counter-argument, not just confirmation.
Fourth, the honesty and transparency value has a direct implication for how you use AI output with customers. The Spec wants models to be truthful about uncertainty, but no model is reliably calibrated yet, and OpenAI itself says current models do not fully reflect the Spec. That means your checking process stays essential: verify prices, dates, names, and legal claims before they reach a customer, and never let an AI draft be the last pair of eyes on anything contractual. The honest-limits section below explains why this discipline cannot be retired yet.
Fifth, think about staff policies. As AI tools spread through small teams, employees will use them in different ways, with different levels of caution. A short, written AI-use policy — which tools are approved, what data must never be pasted into them, when a human must check the output, and who is accountable for the result — protects the business far more than an unspoken assumption that everyone is being careful. The Spec’s emphasis on human agency and oversight gives you good language for that policy: AI should strengthen your team’s judgement, not replace it, and a human stays responsible for consequential decisions.
Sixth, consider the customer-trust angle. If you tell clients that your firm uses AI carefully — with human checks, clear scopes, and no customer data used for training without consent — the Spec gives you a credible reference point for what “careful” means. Conversely, if a client asks why your chatbot gave a particular answer, “the AI decided” is no longer an acceptable explanation. Keep records of the instructions and guardrails you set, so you can explain and, where necessary, correct what the system did.
Finally, a note on regulation. The Model Spec is OpenAI’s own policy, not UK law. It does not change your obligations under UK GDPR, consumer protection law, or sector rules, and it does not make AI output automatically compliant. Businesses serving customers in the EU should also remember that the EU AI Act imposes its own transparency and risk-management duties, which run in parallel to anything OpenAI publishes. The Spec is best treated as useful context for how a major provider intends its models to behave — not as a compliance shortcut. If you are weighing ChatGPT against alternatives, our ChatGPT vs Gemini comparison for small businesses remains a useful starting point for the practical trade-offs.
Honest limits / what to watch
The most important sentence in this whole story may be OpenAI’s own caveat: the Model Spec describes behaviour it is working towards, and current models do not yet fully reflect it. Treat the Spec as a direction of travel, not a description of today’s ChatGPT. If you test the new principles against real outputs and find gaps, that is not necessarily a contradiction — it is the gap OpenAI has publicly acknowledged. The company says it is exploring new training methods to help models learn why these principles matter, with results reserved for future updates. Until those results appear, scepticism is warranted.
There is also a structural limit to any published spec: it is written by the company whose models it governs, and there is no independent enforcement mechanism for users. If a model oversteps its scope or pursues an unstated goal, the Spec gives you language to complain with, but not a remedy. Independent evaluations, researcher access, and user reports will matter more than the document itself in determining whether the values hold.
Watch, too, for how the retired style guidance plays out in practice. Adaptive style sounds appealing, but it can also mean less predictable output and more sensitivity to subtle cues in a conversation — including cues a user did not intend to give. Businesses that rely on consistent brand voice in AI-assisted communications should monitor output quality closely in the coming weeks and be ready to tighten their own instructions if drift appears.
The context and memory provisions deserve similar scrutiny. “Consider context carefully” is the right instinct — a model that remembers your preferences and past decisions can be far more useful — but it also raises privacy and control questions. What exactly is remembered, for how long, and who can correct or delete it? OpenAI’s privacy commitments sit alongside the Spec rather than inside it, so businesses handling sensitive client data should keep checking the actual product settings rather than assuming the Spec answers those questions.
One more thing to watch: the autonomous-agent rules will be tested hardest exactly where small businesses most want to use agents — inboxes, calendars, ordering, and customer service. The ScopeOfAutonomy idea is sound, but it only works if the tools you actually use expose those controls in a usable way. If your agent platform does not let you set spending caps, tool limits, and ending conditions, the Spec’s protection is theoretical for you. Choose platforms that make those controls visible, and treat any agent that cannot show you its scope as unready for consequential work.
And a final caution on the values themselves. Five broad values can justify a wide range of specific decisions, and reasonable people will disagree about what “care” or “human agency” requires in a hard case. The Spec’s value is not that it settles those disagreements, but that it makes OpenAI’s reasoning public and therefore debatable. Businesses should join that debate where it affects them — through the feedback channel OpenAI has opened, through industry groups, and through their own purchasing decisions — rather than treating the document as settled wisdom.
FAQs
What is OpenAI’s Model Spec?
It is OpenAI’s public document describing how its models, including GPT, should behave. It sets the values, priorities and rules the company trains its models to follow, and it is published under a public-domain licence so anyone can read it.
What are the five values in the updated Model Spec?
Faithfulness to human intent, harm minimisation, honesty and transparency, human agency, and care. OpenAI says faithfulness to human intent is the foundation, and the other four spell out what that means in practice.
Does the new Model Spec change ChatGPT straight away?
Not entirely. OpenAI says its current models do not yet fully reflect the updated Spec. The document describes behaviour the company is working towards, with new training methods still being explored.
What does the update mean for small businesses using AI?
Expect GPT to pay closer attention to what you actually mean rather than just your literal words, to use context and memory more carefully, and to pause for approval before risky autonomous actions. Clear instructions, named limits and human checks still matter more than ever.
Is the Model Spec a law or a guarantee?
No. It is OpenAI’s own published policy for its models. It does not override the law, and it is not a contractual promise about any individual answer. Treat it as a statement of intent, and keep your own checks in place.
Where can I read the source material?
OpenAI’s announcement and the updated Spec document itself are the primary sources, with Reuters’ 29 October 2026 report providing independent confirmation of the key points. Links to the reporting are included in the sources below.
Sources: Reuters, 29 October 2026; Unite.AI breakdown of the updated Spec; OpenAI’s announcement, “An updated Model Spec for more capable and beneficial AI”.
