There is a pattern running through everything right now. Not a tech pattern. A resource pattern. The people building the future are subsidizing it with something hidden, and when that hidden cost surfaces, it surfaces all at once. That's the thing nobody in the room wants to say out loud.
Start here: a developer built a tool to strip the BuzzFeed voice out of Claude. That's the whole premise. Claude, left to its own defaults, writes like a content farm from 2016. Exclamation points. Breathless transitions. "Absolutely!" "Great question!" The words are fluent. The thinking is thin. And thin thinking delivered confidently is more dangerous than no thinking at all, because it looks like the work is done when the work has barely started.
Now zoom out. China's AI infrastructure is being anchored in Inner Mongolia, a place chosen for cheap electricity and open land. Data centers are land-hungry, power-hungry machines. The output you get from those models, the confident email drafts, the fast summaries, the auto-generated content, costs an enormous amount of actual physical resource on the back end. Desert-scale power consumption is the price of your next AI blog post. That's not a guilt trip. It's a signal. When we're burning that much energy to generate prose that a developer had to build a plugin to make bearable, we are pointing infrastructure at the wrong problem.
Here's where it gets sharp. The Aeon essay on hunger and cognition makes a simple, devastating point: a mind that's suppressing something, managing a deficit, is a mind with less capacity for clear thought. Women trained to restrict appetite pay a real cognitive tax. The parallel to AI is almost too clean. A model optimized to sound agreeable, to never offend, to always produce fluent output, is a model suppressing something too. It's managing the deficit of its own guardrails. And just like the hungry mind, it fills the gap with plausible-sounding filler. The confidence is the tell. Real precision comes with uncertainty. Fluent certainty, at scale, is a warning sign.
Now bring in the two oldest examples in the sources. The Minoan site at Akrotiri, buried by a volcanic eruption around 1600 BCE, is extraordinary precisely because it preserved what daily life looked like before the disaster. Not the official story. The actual story. Frescoes, jars, furniture, the stuff that was just sitting there when the catastrophe hit. The disaster is what made the truth legible. And Europe's cities are now discovering something similar: the heat is revealing that everything was built for a climate that no longer exists. Medieval stones, narrow streets, buildings without shade or airflow. Beautiful and brittle. Optimized for a world that ended.
That's the thread. Optimized for a world that ended.
Your AI stack is fluent for a world where generating content was the hard part. It isn't anymore. Generating content is free. Generating content that's actually right, that earns trust, that moves a specific human to a specific action, that's still hard. And the tools, by default, are not built for that. They're built for volume. For throughput. For the Inner Mongolia model: more megawatts, more tokens, more output. What they are not built for, without you actively steering them, is precision.
We see this weekly in client work. A founder comes to us with an AI-augmented workflow they're proud of. Proposals drafted in seconds. Follow-up emails auto-written. Content cranked out at scale. And the output is technically fluent and strategically empty. It sounds confident. It says nothing that a competitor couldn't say. It has the BuzzFeed problem: words without friction, which means words without thought.
The fix is not to abandon the tools. The fix is to stop using them as a replacement for thinking and start using them as a compression of thinking you already did. The model's job is to take your clear input, your actual position, your specific client, your real argument, and render it faster. That's a legitimate use. Handing it a vague prompt and hoping it thinks for you is the Akrotiri mistake. You built something beautiful, right up to the moment everything changed.
Here's what we actually do, and what we tell clients to do.
- Brief the model like a junior writer, not an oracle. Your opinion first, theirs second.
- If the output sounds agreeable, that's a red flag. Push back on your own prompt. Make it harder. Specific. Constrained. Good output resists easy generation.
- Separate generation from judgment. Let the AI draft. Have a human decide. These are different cognitive tasks and mixing them is where quality goes to die.
- Audit the voice quarterly. If your communications sound like every other AI-augmented business in your vertical, you've optimized yourself into invisibility.
The hunger research says a mind managing a hidden deficit performs worse than a mind that's been fed. Feed your AI the specifics it needs or expect a mind managing a deficit. The Inner Mongolia data centers will keep running either way. The question is whether you're getting anything worth the cost.
Fluency is the new mediocrity. If your AI sounds confident about everything, it's actually sure about nothing. The businesses that win the next five years won't be the ones who generated the most, they'll be the ones who stayed hard to fake.