要点
- The Verge (Emma Roth, 30 Oct 2026) reports OpenAI, Google, Microsoft, and Anthropic will adopt a shared policy token proposed at the IETF by Anthropic standards technical lead Kevin Kelley, living in robots.txt.
- The four say they will restrict AI crawlers when they encounter the token; Anthropic has already adopted it, Google and Microsoft plan support next year, and OpenAI says implementation is in the “coming months.”
- Microsoft says it will offer granular opt-outs for AI training or AI grounding without affecting Bing search inclusion; Meta’s Andy Stone voiced support for a common signal but the piece does not say Meta will honor it.
The story is less another “we respect creators” pledge than a bid to collapse per-vendor AI-training prefs into one robots.txt token readable across crawlers.
Falsifiable read: if Google and Microsoft ship the shared token in 2027 and OpenAI within the “coming months,” and independent robots probes can verify crawl stops, publishers gain a real cross-lab opt-out; if timelines slip, coverage is product-narrow, or grounding/search carve-outs gut the signal, treat the announcement as an intent letter.
[1]Hard facts (per The Verge): the shared “policy token” is proposed by Kevin Kelley for the IETF and would live in robots.txt, the long-standing file for crawl permissions. Sites today must signal AI-training prefs to each company separately; the control aims at one declaration across multiple AI firms.
Context: publishers face ongoing AI crawlers; RSL tries to charge crawlers; Cloudflare blocks AI crawlers by default. Google already has controls separating AI-search/training use from ordinary search visibility; Microsoft stresses Bing inclusion is unaffected by preference settings.
[1]Evidence is first-line tech reporting of company statements and timelines, plus Meta PR lead Andy Stone’s quote that a common signal is “an important step.” The article does not include an IETF draft number, token syntax, or third-party compliance tests.
Steelman: a unified opt-out lowers compliance cost and reduces arbitrage where one lab honors robots rules another never listed. Beyond the steelman, the piece does not specify breach detection, appeals, or which product paths still legally bypass a “training opt-out” after grounding/training splits.
[1]Strongest steelman: amid lawsuits and crawler arms races, encoding “may we train” as a machine-readable, cross-vendor minimum is more operable than piling up proprietary headers.
Steelman risk: staggered timelines (Anthropic now; OpenAI months; Google/Microsoft next year) create a window for arbitrage; the gap between “restrict AI crawlers” and “search/answers may still cite” can make publishers think they exited all AI use. Meta’s verbal support without a clear honor commitment can inflate the standard’s coverage.
[1]For publishers, a chance to configure crawl prefs once; for the four labs, a governance posture against unauthorized-scraping critiques. For standards readers, watch whether IETF yields interoperable text and whether independent monitoring can reproduce “token ⇒ stop.”
[1]Over the next 6–12 months, judge by reproducible robots probes: if major crawlers still train-scrape after the token is declared, read 30 Oct 2026 as PR sync, not a working default. If probes show consistent stops and grounding/training splits are documented, upgrade the shared signal as content-governance infrastructure.
【No.2】
[1]