The fastest-growing group of Codex users isn't writing code.

OpenAI shared this week that Codex now has more than 5 million weekly active users, up roughly 6x since February. But the more telling number: knowledge workers — people writing reports, drafting contracts, running data analysis — make up about 20% of users and are growing more than three times faster than developers.

OpenAI is leaning into the shift. It rolled out a batch of business plugins hooking Codex into tools like Atlassian, GitLab, Microsoft Suite and Databricks, so it can sit inside workflows knowledge workers already live in rather than being a separate destination.

Virgin Atlantic is the case study OpenAI is putting forward. The airline's leadership said teams are turning workshop ideas into working prototypes in hours instead of weeks, with codebases shrinking by roughly 80% and refactoring jobs that used to take two weeks finishing in about 30 minutes. That kind of compression is what's making non-engineers curious in the first place.

The catch is supervision. When Stanford researcher Andrew Hall asked a similar agent to update one of his old research papers, it pulled new data, ran the analyses and made the figures — but an audit found a lot of errors. The work needed close expert oversight, exactly the kind most knowledge workers can't provide for tasks outside their own field.

OpenAI wants Codex to be the productivity tool for everyone, and the user numbers say it's getting there. The harder problem is that the people getting real value out of it right now still tend to be the kind of users who can spot when the agent is wrong. A lawyer using it for legal work, an engineer using it for code, an analyst using it for data they understand. Knowledge workers pointing Codex at something outside their expertise don't have that backstop, and the audit step is what currently makes these tools usable only to careful users. Whoever closes that gap gets to define what an AI productivity tool actually looks like next year.

After nearly two weeks of delays, Trump finally signed his AI executive order on Tuesday. The whole thing is voluntary.

The order asks frontier AI labs to submit new models to the government for a national security review before release — but Section 3 explicitly states that nothing in it creates a mandatory licensing or permitting requirement. If a lab wants to skip the review, nothing stops them.

It wasn't supposed to be this soft. Trump was set to sign a stricter version on May 21, but pulled out that morning after former AI and crypto czar David Sacks called him directly to talk him out of it. Trump told reporters he "didn't want to do anything that's going to get in the way" of the U.S. leading China in AI. The negotiation that followed was telling: the White House wanted 90 days to review new models before release, industry asked for 14, and the signed version landed at 30 — with no enforcement if a company opts out.

Industry groups predictably loved it, praising the voluntary approach as a way to avoid a patchwork of "conflicting and unworkable requirements." OpenAI's chief global affairs officer Chris Lehane said safety frameworks should come through "democratic institutions" — a careful way of saying Congress, not regulators.

Not everyone's buying it. Senator Mark Warner called the framework lacking "significant substance" and accused the administration of doing "Big AI's bidding." And states aren't waiting around — Florida Governor Ron DeSantis pointed out that an executive order can't preempt state legislation, signaling the patchwork of state AI laws the industry has been trying to head off will keep growing.

The strange part of this order is that Trump's own voters wanted something stricter than what he signed. A Future of Life Institute poll found 79% of Republican voters support the government testing AI models before release, and 87% want the government to be able to block models that pose a national security threat. The signed order doesn't do either. That gap between what the base wants and what the donors wrote is going to define the next round of this fight, and it'll probably get settled by the first state attorney general who decides a voluntary federal framework isn't standing in their way.

The fix for AI misinformation was supposed to be simple: tell people when something was made by a machine, and let them decide what to trust. The research keeps finding labels don't work that way at all.

A new study from Teng Lin at the University of Chinese Academy of Social Sciences found what he calls a "truth-falsity crossover effect." The same AI label pushes credibility in opposite directions depending on whether the content is actually true or false. People doubt real information more when it carries an AI tag — and sometimes give fake content more credit when it doesn't.

That lines up with a study of 877 German Instagram users that found labeled AI images lost about 9 percentage points in perceived authenticity, which is what you'd want. But here's the catch: unlabeled content got a small credibility boost at the same time. Once some posts are flagged as AI, people unconsciously assume the unflagged ones must be fine. Labels don't just flag the labeled content — they quietly endorse everything else.

The frustrating part is that people overwhelmingly want this. When Meta surveyed 23,000 people across 13 countries, 82% said they wanted warning labels on AI-generated content. The demand is real. The effectiveness is the problem.

And the real-world stakes are catching up. Rakesh Dubbudu, founder of fact-checking outfit Factly, said his team is now seeing authentic footage from active conflicts dismissed as AI fakes at scale — once people learn to distrust everything roughly equally, real evidence becomes easy to wave off.

None of this is slowing the rollout. YouTube last week began automatically labeling AI-altered videos, the EU AI Act's transparency rules are already in force, and China has its own labeling regime. The political logic is clear: doing nothing looks worse than doing something.

The case for labels was that they'd help people separate real from fake. The case against them, building paper by paper, is that they're teaching people to distrust everything roughly equally. That's a worse outcome than the one we started with, and it's the one we're scaling. Platforms will keep rolling labels out because the political and PR pressure leaves them no choice. The more interesting question is who builds the second-generation version that actually works before the public concludes nothing online is trustworthy and tunes the whole regime out.