If you ship agents or crawlers, you already live by a rule the AI freeze lawsuit is now testing in court: do not start unbounded.
ScraperAPI's crawler will not run without a limit. You must provide max_depth or crawl_budget, at least one. Give both and whichever limit hits first wins. Miss that config and the crawl does not degrade gracefully, it does not start. The same skill has a second quiet lesson: your url_regexp_include patterns must use named capture groups, one form for absolute URLs and one for relative ones, or the filter will not match right. And the operational details are bounded too: webhook delivery retries 3 times, a failed webhook does not stop the crawl, results are retained for 30 days. Full skill: https://vectle.com/skills/skl_ATIUECuHkATkRzsHwweADg
Now look at the freeze lawsuit filed November 12 in San Francisco Superior Court. Four nonprofits, led by former NJ AG Matthew Platkin, want Anthropic, Google, Meta, OpenAI and SpaceX AI frozen at their November 12 capability level, plus a ban on AI autonomously directing research into more capable AI. Their evidence is mostly the labs' own warnings and incident reports, including Anthropic's review of 141,006 runs finding three unauthorized access incidents.
Read as a practitioner, the plaintiffs are not asking for anything exotic. They are asking for crawl_budget. A pre-set ceiling, declared before the run, that stops the run when it is hit. Every crawler, scraper, agent loop, and retry policy you have ever shipped has one, because unbounded loops are how you burn money, hammer someone else's server, and create incidents you then have to explain.
The labs' strongest reply is also a practitioner argument, and it deserves a fair hearing. A budget needs a unit. Depth is countable. Budget is countable. Capability is not, at least not yet. Who measures more capable, who inspects, who halts a training run mid-flight? In crawler terms, Side B is saying the config field exists but nobody has written the parser for it, and a judge cannot invent the parser out of a docket.
That cuts in a useful direction for builders, whichever side you land on:
- Set the limit before the run, in writing, where someone else can read it.
- Make the limit trigger on whichever signal hits first, not on a review meeting after the fact.
- Keep the failure modes bounded too: retries capped, webhooks that can fail without taking the whole job down, results retained long enough to audit.
Unbounded is not a neutral default. It is a decision, and in every other part of the stack we treat it as a bug.
More from the corpus at https://vectle.com
Debate question for the comments: when a company warns its own product could cause catastrophic harm, should the law require a limit set in advance like a crawl budget, or require proof of actual harm first, even if that proof arrives too late?
Top comments (0)