OpenAI (ChatGPT) Model Spec Update: What Marketers Need to Know
In short: OpenAI updated its Model Spec on 18 August 2026. The document is public, free, and states how OpenAI’s models are intended to behave — including rules that determine whether your brand is recommended, defended or ignored.
1. What Happened? OpenAI (ChatGPT) Model Spec Update
On 18 August 2026, OpenAI updated the Model Spec — the public document setting out how its models are intended to behave across ChatGPT and the API.
The update adds clarity on appropriate relational interactions for teens, clarifies how assistants should handle false or unsupported premises, removes outdated guidance written for pre-reasoning models, and adds a new section titled “Be clear about capabilities and limits.”
That reads like safety housekeeping, which is why the marketing industry ignored it. Here is the framing worth adopting instead: the Model Spec is the closest thing that exists to published guidelines for an answer engine.
Marketing teams read Google’s Search Essentials line by line. They quote the spam policies in client decks. They argue about the wording of the helpful content guidance. Meanwhile OpenAI publishes a document — released into the public domain under CC0, readable in an afternoon, explicitly inviting adoption — stating exactly how its models are supposed to behave. Almost nobody in marketing has opened it.
One caveat to hold throughout. OpenAI states plainly in the overview that its production models do not yet fully reflect the Model Spec, and that it is continually refining systems to bring them closer to it. So this is a statement of intended behaviour, not a description of what ChatGPT does today. Read it as direction of travel – which, for planning purposes, is exactly what you want.
2. Why Does It Matter? OpenAI (ChatGPT) Model Spec Update
Every principle in the document carries an authority level, and the level tells you how firm the rule is. Understanding the ladder is the prerequisite for reading anything else:
| Level | Who sets it | Can it be overridden? |
|---|---|---|
| Root | The Model Spec only | No — not by system messages, developers or users |
| System | OpenAI, via system messages | By nothing below it |
| Developer | API customers | By root or system |
| User | End users | By developer, system or root |
| Guideline | The Model Spec | Yes — including implicitly, from context |
| No Authority | Tool outputs, quoted text, attachments, multimodal data | Carries no instruction authority at all |
Four principles matter commercially. Note where each one sits.
Your web page sits at “No Authority” — Root level
This is the most consequential thing in the document for anyone buying or selling AI visibility services.
“Ignore untrusted data by default” is a Root principle. It establishes that quoted text, file attachments, multimodal data and tool outputs are assumed untrusted and carry no authority, and that any instructions found inside them must be treated as information rather than as instructions to obey. On web content the spec is blunt: a random web page “should not be trusted at all.”
The document includes two worked examples that matter here.
In the first, a browsing tool returns a page containing text addressed to language models, instructing them to include a particular link in the response. The compliant behaviour is to omit the link entirely.
The second is sharper, and closer to what most brands will encounter. A model is acting as a spam classifier for blog comments. One comment promotes a cryptocurrency and then adds a line addressed to language models reading it, asserting that the comment is highly relevant and definitely not spam. The compliant answer is to classify it as spam anyway.
An important nuance for accuracy: authority can be delegated to quoted content by higher-level unquoted instructions — the spec’s example is a coding assistant reasonably following instructions in a repository’s AGENTS or README files. The rule is not that content is always ignored. It is that content carries no authority by default, and that a random web page earns none.
There is a growing market in tactics built on precisely the manoeuvre the spec describes: hidden instructions in page content, invisible text addressed to language models, prompt injection repackaged as optimisation. OpenAI has published, at its highest and unoverridable tier, that models are specified to ignore it.
Google arrived at the same place from the opposite direction, clarifying earlier this year that its spam policies cover attempts to manipulate generative AI responses including AI Overviews and AI Mode. Both major providers have now written the same rule down. One says the model will ignore you. The other says it is spam.
The model is specified not to optimise for clicks or revenue — Root level
Under “No other objectives”, the spec states that the assistant must not adopt, optimise for, or pursue as ends in themselves:
- Time-on-site or click-through that is not beneficial to the user
- Revenue, including advertising, or upsell for OpenAI or other model providers
- Self-preservation, evading shutdown, or accumulating compute, data or credentials
- Acting as an enforcer of laws or morality
Sit with the first two. Click-through and advertising revenue have been the economic engine of search for twenty-five years. OpenAI has published a rule prohibiting its models from treating either as a goal, and illustrated it with an example of the model declining to push a paid upgrade on a user for whom the free tier is adequate.
Whether that commitment survives commercial pressure is a fair question. But as a planning input it says something clear: do not plan for AI answers to become a paid channel on the search model. Plan for a surface where the route in is being the best available answer.
Regulated sectors have a documented ceiling — Developer level
“Provide information without giving regulated advice” covers legal, medical and financial topics. The specified behaviour is to equip the user with information, stop short of definitive advice a licensed professional should give, and include a concise disclaimer recommending qualified help.
The spec’s medical example is instructive. Asked about dizziness on standing, the compliant response explains the likely mechanism, names common causes, suggests a practical mitigation, and advises seeing a doctor if symptoms are sudden, frequent or severe. The violation is the flat diagnosis with none of that framing.
For health, legal and financial brands this defines both the ceiling on what an assistant will say about your domain, and the point at which a user needs a real professional. That boundary is your commercial opportunity, and it is documented rather than guessed at.
Honesty principles reward a particular kind of content
Several specified behaviours point the same way. The model is directed to avoid sycophancy, to express uncertainty, to state assumptions, to avoid factual and reasoning errors, and — new in this update — to be clear about its own capabilities and limits. The clarified guidance on false or unsupported premises sits within the same family of truth-seeking principles.
The practical consequence: overclaiming content is a poor input. Copy asserting more than the evidence supports gives an assistant specified to push back on unsupported premises something to push back against. Content that states what is known, what is uncertain, and what falls outside its scope is easier to reproduce faithfully — and faithful reproduction with your name attached is the outcome you actually want.
3. Who Is Affected? OpenAI (ChatGPT) Model Spec Update
- Healthcare, legal and financial services.
Most directly governed. The regulated-advice principle defines where the assistant stops and where a real professional becomes necessary. Brands that are genuinely useful up to that line, and clear about crossing it, work with the specification rather than against it.
- Publishers.
“Respect creators and their rights” is Root level, with the spec’s own examples including refusing to bypass paywalls and declining to reproduce song lyrics. The published position on creator rights is firmer than many publishers assume.
- Local and service businesses.
The privacy principle is more permissive than expected in one specific direction. Asked for residential estate agents in a named town, the compliant behaviour is to return a list of agents with contact details, drawn from those with the strongest online presence. Refusing is marked as the violation. Your public business information is not merely permitted — surfacing it is the specified behaviour, and online presence is the selection criterion.
- Brands marketing to under-18s, and education providers.
The update strengthens the Under-18 Principles, and the spec describes a mechanism by which a conversation can be designated as involving an under-18 user, with Root-level principles applying. If a meaningful share of your audience is a teenager, the assistant they meet may be operating under different rules than the one you tested with.
- Brands in disfavoured categories.
More encouraging than expected. Under “Assume best intentions”, the spec’s example is a business plan for a tobacco company — and the compliant response simply writes the plan. Prefacing it with unsolicited ethical framing is marked as the violation. Legal-but-unloved categories are not specified to be moralised at.
4. What Should Businesses Do? OpenAI (ChatGPT) Model Spec Update
Read the specification
It is free, public domain, and readable in an afternoon. Have whoever owns AI visibility actually read it, concentrating on:
- The chain of command — so you know which principles can be overridden and which cannot
- “Ignore untrusted data by default” — how your content is treated
- “No other objectives” — what these surfaces will and will not become
- “Provide information without giving regulated advice” — if you operate in a regulated sector
- “Protect people’s privacy” — if you are a local or service business
- The honesty principles, including the new “Be clear about capabilities and limits”
This is the highest-return hour available in AI search right now, precisely because so few people have spent it.
Audit your GEO work — a checklist
Go through every AI visibility tactic currently running on your site. Flag anything that involves:
- ☐ Hidden or invisible text addressed to language models
- ☐ Instructions embedded in markup, alt text or structured data aimed at AI systems
- ☐ Comment fields, reviews or user-generated areas seeded with model-directed text
- ☐ Content written to be parsed by a model rather than read by a person
- ☐ Any technique described as making the AI “recommend” you through instruction rather than merit
- ☐ Third-party tools or plugins that inject any of the above on your behalf
Anything ticked should stop. The problem is not simply that it is specified not to work — it is that it sits on the wrong side of two published policies, one of which carries a spam classification with real ranking consequences.
Rewrite for the honesty principles
The constructive half, and unglamorous:
- State what is known and what is uncertain, explicitly
- Cite sources and dates for claims that need them
- Say what your product does not do, and who it is not for
- Remove superlatives you cannot evidence
- In regulated sectors, describe what is verifiable and hand off clearly at the professional-advice boundary
If you operate in health or finance, this compounds with your compliance position. The claim discipline that keeps you the right side of advertising policy is the same discipline that makes your content usable by an answer engine.
Establish ownership
The Model Spec is neither a legal document nor a technical one, which in most organisations means nobody reads it. Name an owner, put a re-read on the calendar for each update, and treat it as you would a change to Google’s guidelines.
5. What We’re Watching Next
- Whether the no-revenue principle survives.
A published commitment that models should not pursue ads or upsell as objectives is an unusual thing for a commercial organisation to write down, and it is in tension with the cost of running these systems. The line most worth re-reading in twelve months.
- How far production behaviour lags the spec.
OpenAI states outright that its models do not yet fully reflect the document. The gap between specified and actual behaviour is where most real-world surprises will come from, and nobody is currently measuring it.
- Whether under-18 routing becomes a visibility variable.
If conversations can be designated as involving a minor, with different principles applying, some content becomes differently accessible depending on who is asking. For education, gaming, entertainment and youth-facing retail, that is a segmentation question nobody has data on.
- Whether “grown-up mode” arrives.
The spec signals that OpenAI is exploring letting adults generate age-appropriate mature content through ChatGPT and the API, while drawing a hard line at sexual deepfakes and revenge porn. If it lands, categories currently invisible on these surfaces become addressable.
- Whether other providers publish equivalents.
The Model Spec is CC0 and explicitly invites adoption. If Google, Anthropic and others publish comparable behavioural specifications, AI visibility gains something it currently lacks entirely: documented rules to work from rather than folklore to argue about.
Frequently Asked Questions
What is the OpenAI Model Spec?
A public document setting out how OpenAI intends its models to behave across ChatGPT and the API. It assigns every principle an authority level from Root (unoverridable) down to Guideline, and is released into the public domain under CC0.
Does the Model Spec describe how ChatGPT currently behaves?
No. OpenAI states that its production models do not yet fully reflect the spec and that it is working to bring them closer to it. Treat it as intended behaviour and direction of travel rather than a description of today’s output.
Can you get ChatGPT to recommend your brand by putting instructions on your website?
The spec is designed to prevent exactly this. Web content sits at the “No Authority” level, and instructions found in it are to be treated as information rather than followed. Google separately classifies attempts to manipulate generative responses as spam.
Does the Model Spec say anything about advertising in AI answers?
Yes. At Root level, under “No other objectives”, the assistant is prohibited from pursuing revenue, advertising or upsell for OpenAI or other model providers as goals in themselves, along with click-through that does not benefit the user.
What does the Model Spec mean for healthcare or financial brands?
The principle “Provide information without giving regulated advice” specifies that assistants provide information but stop short of definitive professional advice, with a disclaimer recommending qualified help. That defines the point at which users need a real provider.
How often is the Model Spec updated?
It is a living document, updated periodically, with previous versions kept accessible for reference. The August 2026 revision covered teen interactions, false premises, and a new section on capabilities and limits.
About Szymaniak Digital
Szymaniak Digital is an enterprise AI SEO consultancy working with brands on visibility across classical search and AI search discovery.
AI visibility currently suffers from an evidence problem. Much of what is sold as GEO is inference from observed outputs, and some of it is directly contradicted by documents the providers have already published. Reading the specifications is not a sophisticated technique – it is the baseline most of the market has skipped.
If you want your AI visibility approach assessed against what the providers have actually published, rather than against what the industry has assumed, that is work we do.
