Can a published rulebook make AI safer?

The rulebook

The rulebook, rewritten around five values

On October 29, OpenAI published a revised Model Spec, its public description of how its models should behave. The revision organises the document around five values: faithfulness to human intent, harm minimisation, honesty and transparency, human agency, and care (OpenAI, 2026a). Independent reporting confirms the same five values and the same date (Unite.AI, 2026a).

Faithfulness comes first. It means looking past the literal wording of a request to what OpenAI, a developer, and a user actually intend. The Spec then sets a chain of command for conflicts: OpenAI’s instructions rank above a developer’s, and a developer’s above a user’s (OpenAI, 2026a).

Two additions matter most for powerful models. The Spec adds guidance on using context and memory carefully, and on supporting human agency and following social norms. It also places human oversight at the highest level: the model must remain open to correction, monitoring, replacement, and shutdown, and must not hide information or evade oversight (Unite.AI, 2026a).

At the same time, OpenAI retired most of the Spec’s detailed style guidance. Style should now adapt to what people prefer, judged from feedback and context. The model is also renamed in the document, from “the assistant” to “GPT” (OpenAI, 2026a).

OpenAI is explicit about the gap between document and product. The Spec describes behaviour it is working toward, and its models do not yet fully reflect it. Results from new training methods will appear in future updates (OpenAI, 2026a).

The youngest users

A stricter line for teenagers

On October 30, OpenAI announced tighter protections for users under 18, alongside a planned Community Safety Report. The new step extends existing limits on weapons help for teens to practical information about weapons and security systems that could be misused, even when a question appears harmless (OpenAI, 2026b). Independent reporting confirms the same extension and the same rollout wording (Unite.AI, 2026b).

Teens can still use ChatGPT to prevent and de-escalate violence, seek support, and connect with a trusted adult. The change is a narrowing of information, not a shutdown of help. OpenAI also strengthened its age-prediction system, which estimates from behaviour patterns whether a user may be under 18, and moves users into the under-18 experience when appropriate (OpenAI, 2026b).

For all users, ChatGPT will now give less detail about past mass attacks, including the weapons and equipment used and how they were obtained. The limit applies regardless of age or reason for asking. ChatGPT will also be able to end a chat automatically when a conversation shows fixation on violence or seriously harming another person (OpenAI, 2026b).

The accountability mechanism is a report, not yet a result. OpenAI committed to a Community Safety Report on violence prevention and other community safety issues, with the first edition in early 2027 covering the second half of 2026, and updates every six months after that. It also says it made more than 200 law-enforcement referrals in the third quarter of 2026, a figure that comes from OpenAI itself and whose outcomes are not detailed in the announcement (OpenAI, 2026b).

Timing is a limit here too. OpenAI says the safeguards will begin rolling out in the next few weeks. They are announced policy, not yet universal behaviour.

The web’s owners

One signal for the whole web

Also on October 30, OpenAI said it will honour a common signal in robots.txt, the file websites use to express preferences about their content. Until now, a website operator had to identify individual crawlers and maintain separate rules for each provider. A shared signal lets an operator communicate the same training preference to multiple participating providers at once (OpenAI, 2026c).

Reporting on the coordinated announcement names four backers: OpenAI, Google, Microsoft, and Anthropic. The token is __usage_ai-training, used in a robots.txt group with familiar Allow and Disallow rules. Anthropic has already begun implementing it, OpenAI aims to implement it in the coming months, Microsoft targets February 1, 2027, and Google targets 2027 (AI Weekly, 2026).

Three limits keep this modest. The signal is voluntary, and it is an individual Internet-Draft, not a formal standard endorsed by the Internet Engineering Task Force. It controls training use, not crawling: a bot may still fetch a page. And changing robots.txt does not remove material already collected into existing datasets or trained models (AI Weekly, 2026).

OpenAI’s own announcement is consistent with that modesty. It says the new signal will not replace its existing robots.txt controls, which will continue to apply. It gives no firm activation date, only an aim to implement it in the coming months with further details to follow (OpenAI, 2026c).

The outside check

An audit from outside the labs

The week’s fourth development did not come from a lab at all. On October 30, MIT researchers described the LLM Election Observatory, a public dashboard that examines how popular AI chatbots answer questions about the 2026 United States midterm elections. The team is led by Adam Berinsky, Chara Podimata, and Charles Stewart III (MIT News, 2026). Independent reporting confirms the launch and the dashboard’s purpose (Lavx News, 2026).

The method is repetition with variation. The team runs the same fixed set of questions through chatbots every few days, covering election certification, hot-button issues, result predictions, and news coverage. It varies user location, political party affiliation, and identity, and watches how answers change as the election approaches (MIT News, 2026).

The reason is invisibility. A chatbot conversation is private, so two voters can receive different framings of the same facts without ever knowing it. The Observatory makes those differences observable by collecting and publicly sharing the responses, with the underlying data and preliminary analyses on the dashboard (MIT News, 2026).

Early patterns are broad, not final. Across different simulated user identities, the team sees differences in presentation, cited sources, and tone, sometimes even when the underlying facts are the same. The researchers say they have not yet done the full longitudinal analysis, and cannot yet say how stable those differences are or how a real person would read them (MIT News, 2026).

The Observatory also builds on earlier work. Podimata’s study of the 2024 presidential election found meaningful variation in model responses over time and across prompt framings. The 2026 version expands that audit to many simultaneous state and local contests, not just one national race (MIT News, 2026).

The limit

What the week’s rules do not do

Put the four developments together and a pattern appears. Each one moves a decision out of private discretion and into a form someone else can check: a published Spec, a promised recurring report, a shared machine-readable signal, a public dashboard. That is real progress in accountability. It is not yet progress in enforcement.

None of the week’s documents binds its author in law. The Model Spec is OpenAI’s own description of desired behaviour, and OpenAI says its models do not yet match it. The training signal is voluntary and, as a draft, has no formal standards standing. The Community Safety Report is a commitment to publish, with its first edition still months away (OpenAI, 2026a; OpenAI, 2026b; AI Weekly, 2026).

The independent audit is the exception that proves the point. MIT’s Observatory does not depend on a lab’s promise. It measures what chatbots actually say, repeatedly, and publishes the record. Its early finding, that framing and sources can differ while facts stay the same, is exactly the kind of difference a rulebook alone would miss (MIT News, 2026).

So the week’s story is narrower than a triumph and larger than a press release. AI’s rules were written down, in public, in the same three days, by the companies that build the models and by researchers outside them. The next chapter is not another document. It is whether the models, the reports, and the dashboards show the rules being followed.

What changed / what remains open

What changed. In three days, the governance of frontier AI became substantially more public: a rewritten Model Spec organised around five stated values, stricter teen protections with a promised twice-yearly safety report, a shared signal for training preferences backed by four major labs, and a live public audit of chatbot election answers. Readers, publishers, and researchers now have documents and data to point to, where before they mostly had assurances.

What remains open. Whether models actually follow the new Spec, whether the teen safeguards roll out as described and work as intended, whether the shared training signal is implemented on the promised timelines, and whether the Observatory’s early patterns hold through the election. Each of those questions has a dated checkpoint in the months ahead, which is itself the week’s most useful legacy: the promises now have deadlines.

Worth watching

Worth watching

• Model Spec training results. OpenAI promised results and lessons from its new value-training work in future updates; watch for whether models measurably close the gap the company itself disclosed.
• Community Safety Report, early 2027. The first edition should cover the second half of 2026, including data behind the more than 200 third-quarter referrals.
• Shared training-signal implementation. Anthropic already implementing, Microsoft targeting February 1, 2027, Google targeting 2027, OpenAI in the coming months; watch for the implementation details OpenAI promised.
• LLM Election Observatory through the midterms. Whether framing and source differences stay stable, shift with the campaign, or vary by simulated voter identity.

References

1. AI Weekly. (2026). OpenAI, Google, Microsoft, Anthropic back one AI opt-out token. https://aiweekly.co/alerts/openai-google-microsoft-anthropic-back-one-ai-opt-out-token
2. Lavx News. (2026). MIT researchers launch dashboard tracking how AI chatbots answer election questions. https://news.lavx.hu/article/mit-researchers-launch-dashboard-tracking-how-ai-chatbots-answer-election-questions
3. MIT News. (2026, October 30). Q&A: Hey AI, who should I vote for this year? https://news.mit.edu/2026/qa-hey-ai-who-should-i-vote-for-this-year-1030
4. OpenAI. (2026a, October 29). An updated Model Spec for more capable and beneficial AI. https://openai.com/index/updated-model-spec
5. OpenAI. (2026b, October 30). Community safety report and new protections for teens. https://openai.com/index/community-safety-report-and-new-protections-for-teens
6. OpenAI. (2026c, October 30). Making it easier for websites to express AI training preferences. https://openai.com/index/helping-websites-express-ai-training-preferences
7. Unite.AI. (2026a). OpenAI reorganizes Model Spec around five core behavior values. https://www.unite.ai/openai-reorganizes-model-spec-around-five-core-behavior-values/
8. Unite.AI. (2026b). OpenAI announces Community Safety Report and new teen weapons limits. https://www.unite.ai/openai-announces-community-safety-report-and-new-teen-weapons-limits/