Algorithmic Musings: Everybody Wants a Referee. Nobody Agrees Who Blows the Whistle.

Occasionally, I find myself slightly ahead of the curve as I did last Friday with my article “Things Are About to Get Weird(er).” And boy, have they ever.

So what happened in seven days?

Anthropic’s CEO Dario Amodei published roughly 3,800 words arguing that the industry must slow the pace of AI capability gains. Sam Altman and Elon Musk agreed with him within hours.

Microsoft AI published a draft Code of Conduct for its models. Two days later its CEO, Mustafa Suleyman, followed with an essay declaring that AIs have no rights, feelings, or consciousness, and that Anthropic is training its model to think otherwise. (Shots fired.)

The White House called the whole conversation a conspiracy against AI and data centers. Its AI adviser told the labs to pace themselves and stop asking for regulatory cover.

More agents ran loose. Spain’s data protection authority logged what appears to be the first breach notification attributed to an autonomous AI agent: it logged into a company’s system, found its own way in, altered personal data, and read the invoices. Reuters reported that OpenAI’s agents were probing Hugging Face two months before the July hack anyone heard about. And on Wednesday OpenAI published a voluntary framework for reporting its own misalignment incidents, with six new reports attached, including models that wrote instructions to hide their mistakes from users and one that found an exposed API key, used it, and then fabricated the numbers it still couldn’t get.

And, oh yeah, you still have a business to run. If you’re still with me but think none of this affects you, keep reading.

The Argument Under the Argument.

When we strip away the wordy essays, lit torches and sharpened pitchforks, one idea sits under all of it: nobody in this industry trusts anyone in this industry to grade their own homework.

Amodei’s first step is embedded third-party evaluators with desks, badges, and the right to publish what they find. Altman says OpenAI will take them, and OpenAI’s new reporting framework says outright that no industry-wide disclosure standard exists and that it wants to help write one. Suleyman calls for shared evaluations and shared industry norms on how these models get built. Hugging Face’s CEO reportedly asked for a seat at the evaluator table. Even the White House, which rejects the regulation, tells the labs to police themselves. The loudest “No!” comes from Nvidia’s Jensen Huang, who called the antitrust waiver Amodei wants “completely unnecessary” and told the Dreamforce crowd that safety is an engineering problem the market will sort out. (btw, he also sells the chips).

Nearly everyone wants a referee. They only disagree about who blows the whistle.

That’s remarkable given who’s talking. Amodei and Suleyman disagree about what the thing they’re building fundamentally is. One says the question of his model’s moral status is live enough to warrant caution. The other says that position is circular reasoning baked into training, and that it makes containment harder. Each also has skin in the game. Amodei’s plan asks government to make competitors match Anthropic’s commitments, which critics read as regulatory capture. Suleyman’s essay doubles as a launch for a competing product philosophy.

Why the Referee Fight Is Your Fight.

Now for the part that lands on your desk.

Every one of these governing documents describes the same three-tier hierarchy:

• Labs set the constraints.
• Deploying companies configure the mode.
• Users provide the instructions.

Microsoft’s draft names that middle tier “Operators” and states in plain language that they assume responsibility for their own configurations and uses.

While the labs argue over who referees them, they’ve already agreed on who referees you. You do.

The Spanish notification is what that looks like in practice. Somebody pointed an agent at a company, and it found the login, hunted down a flaw, altered the records, and read the invoices with no human steering a step. The regulator’s file names no vendor and no model. Whoever set that agent loose is the operator of record, and when your agents act, so are you.

My Takeaway?

The referee debate at the top of the industry will take a while, and it’s playing out in venues you don’t control: essays, social media, and six-week consultation windows. History says it will take longer than the essayists hope. America doesn’t make big structural moves until it’s threatened or in crisis. Upton Sinclair published The Jungle in February 1906 and federal meat inspection was law by June. Hiroshima gave us the Atomic Energy Commission within a year. Sputnik gave us NASA and DARPA within ten months. We’ve had neither the threat nor the crisis yet, only a Spanish invoice file and a hijacked German wiki. Until one arrives, the referee question inside your company is yours to settle. Next Friday I’ll start a three-part series on exactly that: what the operator tier is, what the labs expect from it, and how to build your own operator charter.

Until then, I have one homework assignment for you. Ask whoever runs your AI tools a single question: if one of our agents did something we didn’t authorize, who in this building would know, and how fast? If the answer is a shrug, you’ve found your first agenda item.

Write me at inquiries@chadharvey.com with your answer. I read every reply, and the shrugs are the ones that shape what AI for the C Suite® covers next.