The Weirdest Part of the OpenAI Hiring Probe Is the Late-Night Radio Ads

OpenAI advertised some of its tech jobs on late-night radio and made applicants mail in paper applications. That's the opening detail of a Justice Department settlement announced Tuesday, and it is funny in the way that a modern AI company behaving like a 1970s staffing bureau is funny.

The company agreed to pay $3.2 million to resolve claims that it discriminated against U.S. workers in favor of candidates on temporary employment visas. OpenAI denied the allegations. It settled anyway.

To see what's going on, you have to understand a small piece of immigration plumbing called PERM - the Permanent Labor Certification process. PERM is how a company sponsors a foreign worker for a green card. Before the Labor Department will approve the sponsorship, the employer has to show it genuinely tried to hire an American for the job. It has to recruit domestically, post the role, and prove there were no qualified U.S. workers available, willing, and able to take it.

Here's the interesting part. During that PERM recruitment window, the Immigration and Nationality Act bars employers from discriminating on the basis of citizenship or immigration status. Even though the employer is ultimately trying to prove there isn't a qualified American - so it can hire the foreign worker - it can't make the process so onerous that Americans effectively can't apply. The rule is a classic regulatory catch: you have to run a good-faith search while trying to make sure the search comes back empty.

OpenAI, and its subsidiary Statsig, allegedly figured out a way to thread that needle.

The DOJ's investigation found that between June 2023 and at least April 2025, OpenAI did not post certain PERM roles on its external careers website, even though it advertised other jobs there normally. For those specific positions, the company required applicants to mail in paper applications. And it advertised some of them on late-night radio.

Let's sit with the paper-application detail for a moment. OpenAI is a company whose hiring funnel presumably runs through an online system. Everyone who wants to work there clicks a URL, uploads a resume, and enters a digital pipeline. For a handful of roles, someone apparently decided that asking Americans to mail a physical piece of paper was a better recruiting method than the tool the company had already built.

That is not the sort of mistake that happens by accident. It is the sort of mistake that happens when someone realizes the PERM process creates a perverse incentive. The company needs to show it tried to hire Americans. It also wants to make sure the pool of American applicants stays small enough that the green-card sponsorship still goes through. Paper mail and late-night radio are legally compliant recruitment steps that also happen to be terrible ways to attract a software engineer who is browsing the company's own careers page.

The simplest model is this: the PERM program creates a box you have to check, and the incentives around that box reward you for checking it in the least effective way possible. OpenAI didn't break the letter of the recruitment rules. It appears to have optimized them.

The settlement splits into two parts. OpenAI pays $1.2 million in civil penalties and sets aside $2 million in a compensation fund for U.S. workers who can show they were shut out of those roles. The company also agreed to post PERM positions on its public careers site going forward, accept electronic applications, update its anti-discrimination training, revise its employment policies, and submit to future DOJ monitoring.

Fewer than 10 jobs were at issue. The settlement amount seems large for a handful of positions, but the DOJ's own press release says the figure "reflects the harm to U.S. workers when they are shut out of applying for lucrative technology jobs." The penalty is less about those specific openings and more about the precedent: a company that is a magnet for top talent gets a price signal for gaming the recruitment mechanics.

This is the 13th settlement the DOJ has secured under its Protecting U.S. Workers Initiative, which was relaunched in 2025. The department has been running this playbook against companies that use visa sponsorship programs as a way to prefer foreign workers without running afoul of the anti-discrimination rules baked into the process. (In a February 2026 case, the DOJ settled with a separate company that used an AI tool to generate job ads that explicitly excluded U.S. citizens - so the visa-hiring enforcement engine is active on both the analog and digital fronts.)

OpenAI's statement says it settled to "resolve the matter and move forward with our PERM program, which is critical for employees and candidates requiring immigration support." That is a fair summary. The PERM process is genuinely important for tech companies that rely on a global talent pool. The problem is that the same program that enables that reliance also comes with anti-discrimination guardrails, and those guardrails create the incentive to recruit Americans in ways that look like recruitment but don't function as one.

There is an older financial machine this resembles. In banking, you sometimes see products that want both the illiquidity premium and the liquidity marketing pitch at the same time - a fund that promises access but builds in gates, redemption limits, or marking rules that make real access difficult. The gap between the label and the economic reality is where the interesting stuff lives.

Here, the label is "good-faith domestic recruitment." The economic reality, apparently, was "recruitment calibrated to produce the outcome we already had in mind." The paper applications and the radio ads are just the mechanism that bridges the two.

The settlement doesn't tell us how many foreign workers were ultimately hired in place of Americans, or which visa categories were at play. The DOJ didn't specify. That gap matters less than the structural point: OpenAI now knows its hiring plumbing will be monitored, and other companies in the same position are getting a price tag for how far they're willing to push the same optimization.

The basic takeaway is that when a regulatory process forces you to try to prove something you'd rather not prove, there is a whole industry of clever ways to make that proof technically true and practically useless. The government's job is to tell the difference. This settlement is the government saying it can.