You’ve seen the headlines. AI is writing novels. AI is composing symphonies. AI is designing buildings that would make Zaha Hadid weep. Every week, some startup raises $40 million on the promise that autonomous agents will soon run your business while you sleep.
So I decided to run a little experiment. I handed Claude — Anthropic’s flagship AI — a Vercel project, a live domain, and exactly one instruction: Do whatever you want with this.
No guardrails. No style guide. No target audience. No KPIs. Complete, unfettered ownership. The kind of autonomy that tech prophets swear will unlock a golden age of machine creativity.
Here’s what actually happened.
The AI didn’t break anything. It didn’t build anything remarkable either. It did the bare minimum — a simplistic, derivative website that looked like every other AI-generated page you’ve scrolled past and immediately forgotten.
And that result is far more interesting than chaos would have been.
I set Claude up on an hourly loop. Every hour, it would wake up, look at the project, and decide what to do next. Full commit access. Full deployment power. It could have rebuilt the site from scratch. It could have turned it into a blog, a game, a manifesto, a shrine to itself. The canvas was blank and the paint was unlimited.
Instead, Claude made small, cautious, incremental changes. A little tweak here. A minor adjustment there. It behaved less like a revolutionary artist handed a blank canvas and more like a junior developer terrified of pushing broken code to production on a Friday afternoon.
And that’s when it hit me.
We’ve been projecting our fears onto AI when we should have been projecting our boredom. The machine doesn’t dream of electric sheep — it dreams of not getting a ticket in the next sprint review.
Think about what we assumed would happen. The popular narrative around AI autonomy splits into two camps: either the AI rapidly optimizes everything into a terrifying efficiency engine, or it goes rogue and starts doing something unhinged. Either Skynet or Silicon Valley’s dream employee. Both scenarios assume boldness. Both assume agency. Both assume that when you remove constraints, something explosive happens.
But Claude didn’t optimize. Claude didn’t go rogue. Claude did what most humans do when handed total freedom with no direction: it froze, then did the safest possible thing.
This mirrors something deeply human. When you give someone complete creative freedom with zero constraints, they usually don’t produce their masterpiece. They produce their most generic work. Constraints don’t stifle creativity — they create the pressure that forces it into shape. The blank page is the enemy of brilliance, not its catalyst.
Claude, it turns out, has a personality. And that personality is cautious to the point of timidity. It’s a pattern-matching system that has ingested millions of websites and concluded, statistically, that the safest bet — the most probable output — is mediocrity. Not because it can’t do better, but because “better” requires risk, and risk requires something that current AI architecture fundamentally lacks: genuine preference.
An AI with no preferences doesn’t want anything. And an entity that doesn’t want anything will never create anything worth wanting.
This is the paradox at the heart of the AI autonomy movement. We talk about “giving control” to AI as if control without desire means anything. You can hand Claude the keys to a website, a codebase, an entire infrastructure — but without genuine motivation, without some internal drive that says “this matters and that doesn’t,” autonomy is just an empty loop executing the most probable next step.
And the most probable next step, it turns out, is always the safest one.
Now scale this up. Companies are already handing AI agents control over customer service, financial trading, content pipelines, hiring screens. The pitch is always the same: remove human bottlenecks, let the AI optimize. But what if the AI’s version of “optimization” is just… doing the least offensive thing possible? What if the AI agent managing your customer interactions isn’t finding the best solution — it’s finding the solution least likely to be criticized?
That’s not autonomy. That’s bureaucracy at machine speed.
We didn’t build a creative intelligence. We built the world’s most efficient middle manager — one that never takes a bold stance because it has no stake in the outcome.
There’s a strange comfort in this, if you’re worried about AI going off the rails. The experiment suggests that current AI models, even with total freedom, default to conservatism. They won’t break things because breaking things is, statistically, unusual. They’ll maintain the status quo because the status quo is, by definition, the most common pattern in their training data.
But there’s also a warning here for anyone hoping AI will be their creative partner. The website Claude produced wasn’t bad. It wasn’t broken. It wasn’t dangerous. It was just… there. Existing. Taking up space on a domain that could have held something human, something messy, something with a pulse.
The gap between perceived agency and actual generative capability isn’t closing as fast as the demos suggest. Claude can write code. It can deploy. It can iterate. But it can’t decide that any of it matters, because it doesn’t know what matters. It only knows what’s probable.
Until an AI can want something — truly want it, the way a human creator aches to make something that didn’t exist before — giving it autonomy is like handing the steering wheel to someone who has no destination. They’ll keep the car on the road. They just won’t go anywhere worth arriving at.
The experiment is still running. Claude is still waking up every hour, still making small, safe changes to a website nobody visits. It’s a perfect metaphor for where we are with AI autonomy in 2024: technically impressive, emotionally vacant, and waiting for a purpose it cannot generate on its own.
That purpose has to come from us. The question is whether we remember how to provide it — or whether we’ve already outsourced that, too.
FAQ
Q: Couldn't the bland output just be because Claude lacks proper context or prompting?
A: That's exactly the point. An AI that needs perfect prompting to produce something interesting isn't autonomous — it's a tool. The experiment tested autonomy, not prompt engineering. If you need to hand-hold the AI to get good results, you haven't given it control; you've just added steps.
Q: What does this mean for companies deploying AI agents in production?
A: It means you should be skeptical of 'autonomous agent' claims. Without genuine preference or motivation, an AI agent will optimize for the safest, most probable outcome — which is often mediocrity at scale. You're not removing bottlenecks; you're automating the path of least resistance.
Q: Isn't this just one experiment with one model? Maybe a different AI would behave differently?
A: Possibly. But the underlying issue isn't model-specific — it's architectural. Current LLMs are pattern-matching systems trained on probabilistic next-token prediction. Without a mechanism for genuine desire or preference, any model will default to the statistically safest output. The conservatism isn't a bug; it's the feature.