October 21, 2026
TechCrunch reported on a new survey finding that more than half of managers say it's gotten harder to evaluate their employees' work since AI entered the picture. That's not a small footnote -- it's the whole pitch for workplace AI quietly cracking in the middle. The sales pitch has always been: AI makes people faster, so output goes up, so everyone wins. Nobody built the second half of that story, which is how a manager is supposed to tell whether a report that took two hours instead of two days is actually better, or just faster and shinier.
This is the quiet cost of adoption that nobody puts in the product demo. When every employee has a slightly different AI workflow -- different prompts, different tools, different amounts of editing afterward -- the variance in how work gets made explodes even as the work itself starts to look more uniform on the surface. A polished deck or a clean-sounding email no longer tells you much about the thinking behind it. Our take: this is a management tooling gap, not an AI problem, and it's going to get worse before anyone fixes it, because most companies adopted AI tools faster than they adopted any way to track how those tools are actually used inside a workflow. Teams that run their work through a connected system -- where you can see the steps, not just the finished output -- are going to have an easier time here than teams relying on a grab-bag of individual AI subscriptions. That's part of why we think workflow automation that keeps a visible trail matters more now than it did a year ago, and it's a problem internal tools are better positioned to solve than one more chat window bolted onto someone's laptop.
Anthropic is reportedly pitching a federal AI safety bill that sounds tough on paper -- it would allow a model to be blocked if it could plausibly contribute to 50 or more deaths. But TechCrunch's reporting flags the catch: the bill would actually bar federal agencies from halting a model's release on their own authority, requiring a judge's sign-off first. That's a meaningful inversion. A policy framed as protection ends up functioning as a procedural shield, moving the decision from a regulator who understands the technology to a court that, in practice, moves slowly and may not.
We don't think this makes Anthropic the villain here -- a company writing its own proposed regulation is always going to shape it in ways that are survivable for its business, and that's a predictable incentive, not a conspiracy. But it's worth being clear-eyed about what's being proposed versus what it's being called. A bill that requires judicial approval before any agency can act is a bill that trades speed for process, and in a fast-moving industry, process delay is often the whole ballgame. This lands in the same week as reports that Anthropic's own cloud partnerships are getting strained, with providers increasingly treating it as a rival even as its infrastructure costs climb ahead of a planned November IPO -- a reminder that the company writing the safety rulebook is also under real commercial pressure to keep shipping. If you want the fuller picture on that compute squeeze, we covered it in our rundown on Anthropic's cloud partnerships and who actually owns the AI boom.
Hugging Face co-founder and CEO Clem Delangue, who appears in Netflix's new OpenAI-hack documentary "AI Gone Rogue," is publicly pushing back on it, saying the film went "full doomer mode" and doesn't accurately represent what happened. That's notable precisely because Delangue isn't an AI skeptic on the sidelines -- he runs one of the most important open-source AI companies in the world, and he's saying a widely distributed Netflix documentary oversold the danger of a real incident.
Our view: this matters beyond one documentary's accuracy. Public understanding of AI risk is being shaped largely by dramatized media, not by the people closest to the incidents, and when someone with Delangue's standing says a popular narrative is exaggerated, business leaders should actually update their priors a little. It doesn't mean AI security incidents aren't real or serious -- they clearly can be -- but it's a useful check against making security decisions based on vibes from a streaming documentary instead of actual incident reports. If your business is thinking through what real AI security practice looks like versus the dramatized version, it's worth a look at how we think about this at our Trust Center and our incident response approach.
Where do you land: is the harder-to-evaluate problem from the manager survey something your own team has felt, or are you seeing the opposite -- that AI is actually making it easier to judge quality once you know what to look for?
Sources