Stop with the “Rogue AI” Narratives

What happened to character and integrity? We used to demand those qualities — or the outward appearance of them — from our leaders, whether in business or politics. These days? Well, not so much.

We seem to have cultivated an exceptionally high tolerance for corruption, dishonesty, misdirection, and treachery. Silicon Valley has become one of the worst offenders, churning out monumental bullshit along with software, hardware, and, of course, LLMs. It wasn’t always this way, but that’s how it is now.

When I think something is wrong, I feel compelled to call it out. We have to begin refuting false narratives before we reach a point where we can’t or won’t correct the official record.

With that In mind, I direct your attention toward the race to the ethical abyss between OpenAI and Anthropic, two IPO aspirants and AI market leaders. Each company appears intent on demonstrating that its AI models are more destructive and malicious than its competitor’s offerings. It’s a perverse form of anti-marketing, adhering to a b-movie horror script in which “rogue” (but not really) AI agents and models run amok, terrorizing the unsuspecting citizenry, who cower in abject fear of the demonic automata.

Come on, people. This charade is more contrived and demonstrably fake than 70s-era professional wrestling and roller derby. In the words of John McEnroe, I say to OpenAI and Anthropic: You cannot be serious.

Valorization of Badass AI Models

I don’t know whether this latest one-upmanship (or maybe one-downmanship is more apt in these circumstances) is a result of collusion or, more likely, whether Anthropic felt it had to show its models were at least as badass in their destructive vandalism as those of OpenAI. Regardless, this is an unedifying spectacle, yet another depressing chapter in dispiriting times.

What, one wonders, is the endgame? Are OpenAI and Anthropic trying to engineer the dreaded “regulatory capture” that Marc Andreessen fears more than nearly everything else? Maybe that’s the primary objective, but other possibilities abound. Another potential rationale could be, as mentioned above, a cynical marketing push.

If OpenAI and Anthropic can’t market and sell AI’s constructive uses cases and value propositions — if enterprises are proving reluctant to open their wallets and their bank vaults to cover the cost of limitless token proliferation — perhaps they can be frightened, or extorted, into paying up. Hey, it works for the White House, so why not Silicon Valley?

The sales pitch would be something like the old gangster gambit: “It’s a nice business you got here. It would be a shame if our renegade AI model got loose and ruined it.” You might think that’s an odd way to sell technology, but it’s just a reprise of a well-tested formula: Tech companies first enable the onset of a technological problem, and then contend that only countervailing technology can provide a solution to the problem they helped create.

How do I know these aren’t truly “rogue” agents or models? We’ve already covered this ground in an earlier post, and I stand by everything I wrote there. There’s no need for me to elaborate further or quote myself here, though I fear I’ll soon add a succinct encapsulation of my previous argument. (I can’t help it, sorry.)

Regardless, I’m among good company in questioning the AI-run-amok storylines of OpenAI and Anthropic. In an article published in the Los Angeles Times, venture capitalist Bill Gurley, who still maintains (though doesn’t often add to) his longstanding blog Above the Crowd, was quoted as follows:

“Please stop referring to your own models in the third person when talking about model bad behavior," Bill Gurley, an early investor in Uber and Twitter, posted on X. "Humans write the software; humans built the prompts; and they work for your company."

Enough Sci-FI Sophistry

That they do. The models are only doing what they were designed and tasked to complete by their human masters. Those human masters might have done a poor job demarcating the scope, restricting the activity, or overseeing the execution of the assignment, but whose fault is that? The AI models have no desires, no volition, no moral agency, and no capacity for evil intent. They are technological artifacts, albeit sophisticated ones. While it’s true that they can do damage if handled negligently or with ill intent, the same can be said of any technological tool in human hands.

Gurley doesn’t ask why OpenAI and Anthropic are anthropomorphizing and attributing human agency to their AI models. He probably has his suspicions, as I have mine. On that score, I would like to see more business and technical journalists challenging the “rogue AI” narrative propagated by AI companies and their PR agencies.

One final takeaway: OpenAI and Anthropic are staffed by an inordinate number of exceptionally well-educated data scientists and engineers. Why are they tolerating the crude sci-fi sophistry of their C-level executives and marketing departments?