AI giants back a shared training opt-out for websites

Google, Microsoft, OpenAI and Anthropic are backing a common way for website owners to say whether their content may be used for AI training. The proposal could simplify publishers’ controls, but implementation is still ahead and the details differ between providers.

Today’s podcast also covers OpenAI’s police-referral disclosure, confirmed ChatGPT allowance changes, citations in Google Sheets and an unusual video model developed using research on living neurons.

Listen and follow on Spotify, Apple Podcasts or Amazon Music. Explore more editions in the podcast archive.

The October 30 announcements from Google, Microsoft and OpenAI describe a shared signal in robots.txt. An Anthropic engineer also confirmed that implementation is under way.

The aim is to reduce the need for publishers to maintain different training preferences for every provider. Existing company-specific controls remain relevant, and the underlying proposal is still an individual Internet-Draft, not a finalized IETF standard.

Microsoft plans to honor its new preferences from February 1, 2027. Google targets 2027; OpenAI says the coming months. Publishing the signal today does not mean every participating company is already applying it.

Training preferences also differ from blocking a crawler or deciding whether a page can inform an AI-generated answer. Microsoft’s system separates training from that answer-time use, known as grounding. Google’s implementation maps the shared opt-out to Google-Extended unless a more specific instruction applies.

For publishers, the useful next step is understanding those distinctions before changing existing settings. One common signal does not guarantee identical product behavior across companies.

In its October 30 announcement, OpenAI says it made more than 200 law-enforcement referrals in the third quarter of 2026. That is a count of referrals, not a count of convictions or independently verified attacks prevented.

The company plans to publish its first Community Safety Report in early 2027. Its existing process description explains how automated detection is followed by human review and how cases involving imminent, credible harm to others can be escalated.

The disclosure raises questions about safeguards and accountability. Reporting on outcomes and errors will be important when assessing how well the process works.

OpenAI’s October 29 release notes confirm that Pro 100 and Pro 500 receive twice their previous Work and Codex allowance. Pro 200 users already on the higher allowance retain it, while users on the lower allowance are brought up to match. A planned reduction has been cancelled.

Those distinctions matter: not every Pro 200 subscriber receives a doubling of their previous usage. The Codex guidance explains that eligible features can draw on shared allowances and that consumption varies with the task and model. Users should check their own dashboard rather than assume a fixed number of coding jobs.

Google’s October 28 announcement adds individual web-source citations to values generated with the AI function in Sheets. The change is part of a gradual rollout to eligible plans that began October 20.

The feature guide explains how to inspect a generated cell’s sources. That creates a more useful route back to the evidence, although a citation still needs checking for relevance and accuracy.

For a team comparing products or gathering information, the benefit is not simply a faster-filled spreadsheet. It is being able to investigate a questionable entry without starting the search again.

The Biological Computing Company launched Mischief 1 on October 30. Built on Wan 2.2, the image-to-video model is promoted as five times faster than its base model. That is a company claim, not an independent finding from our own testing.

TBC’s technology explanation makes the key distinction: living neurons are studied in its laboratory, and useful discoveries are translated into software. Customers run the resulting model on conventional AI hardware.

Mischief 1 is listed on fal. Creators evaluating it should compare usable output, consistency and the cost of repeated attempts, rather than speed alone.

It is an unusual research route to a familiar goal: making useful AI less expensive to run.

Follow AI World Today: Daily News for the next edition.