Four announcements this week each drew a line about who controls an AI system. Read together, they show that no two vendors draw it in the same place, and that none of the resulting documents is a standard you can audit against.

That is not a criticism of any one of them. It is a procurement fact, and it changes what an enterprise should ask for.

OpenAI: a behaviour document, not a warranty

OpenAI published an updated Model Spec organized around five values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care.

The structural detail matters more than the list. OpenAI states the Spec sets firm boundaries where needed to protect people and society, and otherwise mostly takes the form of defaults that users and developers can override.

It also states, according to the version reviewed earlier in this publication's coverage, that its models do not yet fully reflect the document. No adherence measurement or independent evaluation accompanied the update.

A specification of intent is genuinely useful. It tells a buyer where to aim tests. It is not a service-level agreement, an audit right or a warranty, and filing it as evidence of controlled behaviour is a category error.

Microsoft: a division of labour, made explicit

Microsoft's Copilot harness post makes a blunt argument in its own words: if you want to compete, you need to be multi-model, and taking advantage of what different models offer requires a great harness solving for performance, quality and cost.

The responsibility split is the part worth re-reading. Microsoft runs the agent loop and the sandbox. The customer supplies the instructions, skills, tools and model choice.

That is a clearer allocation than most behaviour documents offer, because it names who is accountable for what at runtime. It also moves the burden: the quality of the instructions, the safety of the tools and the suitability of the chosen model become customer-side obligations.

OpenAI: a preference signal, not a control

OpenAI said it will honour a shared robots.txt signal to help website operators communicate their AI training preferences to participating providers.

Participating is the operative word. A voluntary convention among companies that opted in binds nobody who declines, and it is not a substitute for a contractual term or an access control.

The post was available as a publisher summary only. The signal's syntax, the participant list, an effective date and any enforcement mechanism were not established, and none is asserted here.

The practical consequence is still real for anyone whose published content, product documentation or reference material is a training input. A preference expressed to a defined group is weaker than a permission enforced against everyone else.

Google: provenance that travels with the output

Google's Gemini Notebook post describes adding hundreds of textbooks from providers including OpenStax and S Chand at no charge, with every answer grounded directly in the textbook, inline citations, and direct links back to the original source.

Set that against almost everything else announced this week. The other three mechanisms live on a vendor's website or in a runtime configuration. This one attaches to the answer itself.

A record that moves with the output is the only kind a reviewer can reconstruct after the fact, and reconstruction is what an audit is.

The announcement habit, in both directions

The week also produced partnership news that should be read for what it is. OpenAI and ServiceNow described an expansion to bring OpenAI models and agents into ServiceNow, and ServiceNow workflows into ChatGPT and Codex. That is an announcement of intent, not a shipped integration, and the practical question is what is generally available versus in preview.

The same discipline applies to outcome claims. One vendor story this week reported a striking percentage return from a single enterprise pilot. It is vendor-published, single-customer and unaudited. It may be accurate; it is not evidence a buyer can generalise from.

What none of the four settles

  • No adherence data. Nothing published alongside the behaviour specification measured whether models comply with it.
  • No participant list. Which providers honour the training-preference signal, and from when, is unknown.
  • No availability commitments. Partnership and product posts this week did not establish general availability, entitlements or regional limits.
  • No independent evaluation. Every mechanism described here is vendor-authored and vendor-reported.
  • No market data. No reliable dated market evidence was reviewed for this edition, so no stock moves are reported.

Seven questions before relying on any of it

  • Boundary or default? For each behaviour you rely on, is it a firm commitment or an overridable starting position?
  • Is it in the contract? Does the agreement incorporate the document, or merely link to a page the vendor may revise?
  • Who owns runtime? In a harness arrangement, who supplies instructions, tools and model choice, and who is accountable when they fail?
  • Where is the record? When someone overrides a default, what captures the setting, the system prompt and the tool permissions?
  • Does provenance travel? Can a reviewer reconstruct why an output says what it says, without asking the original operator?
  • Is the preference enforceable? If a crawler declines your signal, what in your access controls or contracts stops it?
  • Announcement or availability? What is generally available today, in your region and on your plan?

JWA perspective

This section is analysis, not reporting. Three observations follow from the week.

First, the control surface has moved from policy to configuration. If most behaviour guidance is overridable, the enforceable artefact is the set of choices an organisation actually made, plus the record of who made them. That is an engineering deliverable, not a policy deliverable.

Second, provenance is becoming the dividing line between outputs you can defend and outputs you merely possess. A cited answer survives three handoffs. A generated file with no trail does not, and the difference is a design decision made at build time.

Third, voluntary mechanisms reward organisations that know their own exposure. You cannot express a preference about training data for systems you have not inventoried, or set a boundary around a crawler you did not know could reach.

JWilliams & Associates architects, develops, integrates and operates enterprise software, multi-tenant SaaS platforms, AI-enabled business applications, automation systems and secure technology infrastructure. Its stated design philosophy treats custom AI as a force multiplier for company staff rather than making staff replacement the default goal. That is a design philosophy, not a guarantee about outcomes, and no productivity, savings or service-level result is claimed here.

Evidence notes and limits

Evidence was reviewed through 23:00 UTC on 30 October 2026, covering items published between 28 and 30 October 2026. Six primary vendor posts support the factual claims above: three on openai.com covering the updated Model Spec, the shared training-preference signal and the ServiceNow partnership expansion; two on blog.google covering the Gemini Notebook textbooks feature and the Gemini 4 Carbon model announcement; and one on blogs.microsoft.com covering the Copilot harness.

All six are vendor-published and this pass retrieved them as publisher summaries rather than full text. They are described at that level, and quoted phrasing is limited to material appearing in the summary text itself. The Model Spec's boundaries-and-defaults characterisation is drawn from the document as retrieved during earlier coverage in this publication.

No claim is made about products, releases or integrations beyond what those posts state, and the ServiceNow item is characterised as an announcement rather than a completed integration. No adherence figures, detection rates, participant lists, availability dates or market data are asserted. The seven questions and the perspective section are analysis and suggested method, not a tested protocol.