"Not an Industry Body," so Says the Industry

The Economist published a plan this week for keeping frontier AI safe. Demis Hassabis, co-founder of Google DeepMind, wants a referee for the most powerful AI models. He told The Economist the overseer cannot be a government agency; too slow, wrong resources. He also said, quote, "it's important that it's not just an industry body."

His solution? A private agency. Funded by AI companies. Modeled on FINRA.

Huh? Make that make sense.

FINRA IS an industry body. It’s straight up the textbook definition of one. FINRA stands for the Financial Industry Regulatory Authority. It’s a private, not-for-profit organization funded by the member firms it oversees. Meaning, Wall Street pays FINRA's bills. And that’s the model Sir Demis holds up as the answer for AI. He ruled out industry self-regulation, t AI regulation, FINRA model AI, Demis Hassabis AI safety proposal, frontier AI oversight, AI self-regulation, regulatory capture AI, independent AI audit, ISO 42001, AI governance, AI assurance, AI benchmarks, AI compliance hen proposed industry self-regulation with better branding.

And yes, we have to talk about the "Sir." The tech press bows so low for this man you can hear vertebrae popping. We are beyond the floor and digging towards Hades. Knighthoods don’t confer regulatory wisdom. They confer a title. When the coverage leads with deference, scrutiny picks up their toys and goes home early. So, let's do the scrutiny.

FINRA's Actual Record

You deserve the receipts, so here they are.

• FINRA formed in July 2007 from the merger of NASD's regulatory arm and NYSE regulation.
• 14 months later, Lehman Brothers collapsed and the financial system nearly followed.
• 17 months later, Bernie Madoff confessed. His broker-dealer had operated under NASD and then FINRA oversight for decades.

The largest Ponzi scheme in history ran its final act on FINRA's watch. Madoff himself had chaired the NASDAQ stock market in the early 1990s. He wasn’t hiding from the club. He was in da’ club.

For years, years critics have hammered FINRA over conflicts of interest, opaque governance, and an arbitration system that keeps investor disputes out of public courtrooms. So, there’s no words to describe the ridiculousness of it all when Demis lovingly labels his approach as the "elegant" precedent for governing artificial intelligence. The Wikipedia entry alone is a syllabus in regulatory capture.

Who Picks the Test Picks the Winner

Here’s the part of the proposal that should worry you most, and it’s buried deep in the details. Under the Hassabis plan, this new agency selects the benchmarks that decide which models count as "frontier." The labs fund the agency. The agency writes the test. The labs take the test.

You already know how that story ends because you have watched it in finance, in accounting, and in every industry that graded its own homework.

Arthur Andersen audited Enron while collecting consulting fees from Enron. The credit rating agencies stamped AAA on securities issued by the firms paying their invoices. When the entity setting the standard depends on the entities meeting the standard, the standard bends.

The auditor cannot be the vendor. We built Fusion Collective on that principle, and this proposal violates it in the first paragraph.

What This Means for You

Washington is reportedly close to announcing its preferred oversight model and if the FINRA template wins, "certified safe" will mean "approved by a body the labs pay for." Your board will see that certificate and relax. And it should NOT.

Do these four things now:

1. Ask every AI vendor who audits their models and who pays that auditor. If the answer traces back to the vendor, the audit is pure marketing fodder.
2. Demand independent, 3rd-party assessment language in your AI contracts before any federal scheme sets a lower than dirt bar as the default.
3. Map your own AI systems against ISO 42001 or NIST or another comprehensive framework today. Independent international standards exist. Use them while independence still means something.
4. Put one question on your next board agenda: if our AI vendor's regulator is funded by our AI vendor, what’s our actual assurance?

The Clock Is Real while the Answer Is Not

I’m going to agree with Demis on ONE thing. The threats are serious and the hour is late. Cyberattacks driven by AI are already here.

Urgency is legitimate and urgency is also how really bad structures get approved. "Something must be done, this is something, therefore do this" is precisely how finance got FINRA, and finance got 2008 anyway.

Demand better. Demand oversight with no revenue line from the overseen. The gold standard for AI assurance will come from independence, verification, and boring, rigorous audits.

Checkbox theater with a royal accent is still checkbox theater.

Checkbox theater ends here.