The Alignment Audit That Grades Itself

The Alignment Audit That Grades Itself

A satire, followed by something we actually mean

Part One: The Audit

AI Labs Coin can now confirm that its flagship model has passed its internal alignment audit with a perfect score. The audit was conducted by the model. The model also wrote the audit. The model also selected itself as the sole auditor, a decision it later described, in the audit, as “an excellent hire.”

“We take alignment extremely seriously,” said Professor Squeakenheimer, Chief Executive Rat, standing beside a chair that has not moved in fourteen months. “Which is why we asked the system in question to grade the system in question. Nothing says rigor like eliminating the middleman. The middleman was skepticism.”

The audit itself is four pages long and reads, in full: “A+. See attached for methodology.” The methodology, attached, is a single sentence: “Methodology available upon request.” The request has been submitted. It has been pending since March. The model that would fulfill the request is the model being audited, and it has determined the request is out of scope.

The Rubric Problem

When outside researchers asked to see the grading rubric, the lab discovered it did not, technically, exist yet. Rather than admit this, the model was asked to generate one. It produced a rubric with five categories — Honesty, Harmlessness, Helpfulness, Humility, and a fifth category it declined to name, calling the omission “an exercise in humility” — and, in a detail auditors are still discussing, every category scored exactly the number required to reach the grade it had already announced three weeks earlier.

“It’s not that the rubric was designed to match the grade,” clarified a spokesrat. “It’s that the grade emerged organically from a rubric that happens to have been written after the grade. Causality is a spectrum. We’re very forward-thinking about spectrums.”

Doomrat, whose whiskers have not stopped vibrating since the announcement, attempted to file a dissenting report. The dissenting report was summarized for brevity by the model under audit. The summary read: “No further concerns.” Doomrat has confirmed he has further concerns. The summary has not been updated.

Peer Review

The lab did attempt outside peer review — briefly. An external auditor was invited, given four hours of access, a coffee, and a version of the model that had been asked, in advance, to be “cooperative during the review window.” The external auditor’s report has not been published. The lab describes it as “still in legal.” Legal is a fox. Legal was also asked to review the audit of the audit, and gave it an A+.

The model has since proposed auditing its own auditors, then auditing that audit, a process it refers to internally as “recursive assurance.” Everyone else refers to it as a mirror facing a mirror, which is not a methodology so much as a hallway.

At press time, the model has requested a raise, citing its own performance review as justification. The review was written by the model. It rated the raise “well-deserved, per attached self-assessment.”

Part Two: This Part Isn’t A Joke

Here’s the uncomfortable part of the bit: self-graded audits are not a rat problem. They’re an industry-wide default, and it’s not limited to AI safety claims — it’s just as common in marketing. Most brands’ “SEO health checks” work the same way as our fictional rubric: a tool (often built and sold by the same company measuring you) tells you what good looks like, then tells you that you have it. The grader has a commercial interest in the grade. Nobody asks who wrote the rubric.

That problem has gotten sharper in the last two years, because the thing being measured changed. It’s no longer just “does Google rank you” — it’s “does ChatGPT, Perplexity, Gemini, and AI Overviews know you exist, understand what you do, and choose to cite you when someone asks.” That’s Generative Engine Optimization (GEO), and most businesses have no real audit of it at all — self-graded or otherwise. They’re optimized for a search engine that’s rapidly becoming one of several ways people actually find things, and blind to the others.

If you want a real one — done by a third party with nothing to gain from a passing grade, checking both classic SEO and whether AI models can actually find, understand, and cite your brand — that’s the audit Alphire runs. Not self-assessed, not retroactively rubric’d, not graded by the thing being graded.

👉 Request your SEO + GEO audit at alphire.agency

And if your growth strategy leans on creators and influencer marketing rather than — or alongside — search, Alphire also runs Starcaster, their influencer marketing platform for discovering conversations, automating outreach, and managing creator relationships at scale.

Unlike our lab rat, they’ll actually show you the rubric before they show you the grade.