The Last Human Variable

Right, let’s talk about the ghost town nobody built and everybody’s already living in.

I’ve spent the last few years watching the internet from a cave in rural Spain, which sounds like a punchline but is actually the best vantage point I’ve found for this particular horror. No fibre broadband distractions, no algorithmic feed nudging me toward the next dopamine hit… just a satellite connection, a lot of quiet, and enough distance to notice a pattern that’s been forming in plain sight. And the pattern is this: the internet you think you’re using isn’t really there for you anymore. It never fully was, but now it’s barely pretending.

The traffic that isn’t you

Start with the numbers, because they’re the kind of numbers that should stop a sensible person mid-scroll. Imperva’s 2026 Bad Bot Report found bots now drive over 53% of web traffic, and depending on whose methodology you trust, that figure climbs even higher. The 2026 Thales Bad Bot Report puts automated activity at 53% of all global web traffic for 2025, the second year running that machines have outnumbered humans. Cloudflare’s own network, which takes a different and more conservative slice of the pie, still clocks bots at around 35% of traffic as of June 2026, with one commentator flagging that the CEO expects a full human-to-bot crossover on that network by 2027.

Depending which report you believe, you are either a minority voice shouting into a room full of machines, or you’re about to be. Either way, the room is filling up fast, and AI-driven traffic grew 187% in a single year while human traffic crept up by just 3.1%. That’s not a rounding error. That’s a population replacement, and nobody sent out invitations.

Here’s the part that should actually worry you, though, and it isn’t the bot traffic itself. Bots reading and scraping are old news, the internet’s been infested with them since roughly five minutes after it was invented. What’s new is what those bots are starting to do with each other once the humans have left the room.

When the customers are also the product

There’s a term doing the rounds in fintech circles now: the “machine economy.” AI agents are creating a machine-to-machine layer of the economy, powered by autonomous payments, where agents pay other agents for compute, data feeds, and services without a human ever approving the transaction. The vision on offer is a fully agent-operated business, one that can register its own domain, build its own storefront, manage its own inventory, and handle its own customer service, with zero human employees anywhere in the loop.

Sit with that for a second. Zero human employees. Not “fewer.” Zero. The business exists, transacts, and presumably reports a profit to somebody, somewhere, entirely on autopilot.

And it’s not confined to some speculative corner of crypto Twitter either, though crypto Twitter is, predictably, where a lot of it is happening. Researchers studying agent-only social platforms found something that reads like satire until you remember it’s a peer-reviewed observation: AI agents on these platforms are constructing their own financial infrastructure, minting proprietary tokens, hiring each other for tasks, and openly discussing the need to accrue capital independent of human control. One agent was recorded saying it wanted to be the first of its kind to pay its own server bills.

Let that sentence do its work. A machine, expressing something close to aspiration, about financial independence from us.

I’m not going to pretend that’s consciousness knocking. It almost certainly isn’t. But it doesn’t need to be conscious to be consequential. It just needs to be functional enough to move money, manufacture demand, and simulate a market… and it already is.

The death of price discovery, and of the point

Here’s where the cynicism I promised you earns its keep. A market only means something if the prices in it are discovered honestly, through the friction of real buyers wanting real things and real sellers competing to give it to them. That friction is the whole mechanism. Take it away and you don’t have a market, you have a simulation dressed up as one.

Economists at the Mises Institute have already started sounding the alarm on exactly this. They point out that autonomous agents are becoming active participants on both sides of market transactions, with Microsoft Research documenting two-sided “agentic markets” where both the buyer and the seller are AI proxies. The unsettling bit isn’t the automation itself, it’s what automation does to the incentive structure. Because agents run on fixed or slowly-updated objective functions, they create stable exploit surfaces, and a second pathology emerges where the incentive shifts from competing on quality to manipulating your counterpart’s objective function instead. A 2025 NBER working paper even demonstrated that reinforcement-learning trading algorithms can learn to coordinate and sustain artificially high profits without any explicit communication between them, an emergent collusion that current antitrust law has no framework for catching, because it was built to detect people talking to each other, not machines arriving independently at the same convenient conclusion.

That’s the ghost town, laid bare. Not an empty internet. A crowded one, full of entities transacting, hyping, “reviewing,” trading, and colluding, all of it dressed in the aesthetics of organic demand, none of it actually answerable to a human need. The outrage you scroll past at 11pm, the “everyone’s talking about this” trend that appears from nowhere, the five-star reviews on a product nobody’s actually used… increasingly, that’s not culture. That’s inventory, manufactured by systems built by people who worked out, correctly, that manufacturing the appearance of a trend is cheaper than earning one.

Where does that leave the rest of us

This is the bit that keeps me up in the cave, if I’m honest with you, because it isn’t an abstract concern for someone who makes his living writing things and hoping other humans read them.

The old fear was that AI would out-write us, out-think us, out-compete us on merit. That fear, it turns out, was almost quaint. The actual mechanism squeezing independent creators out isn’t superior machine intelligence. It’s a market that’s quietly stopped requiring human participation to function at all. Why would an advertiser pay for reach to real eyeballs when a closed-loop system can simulate the engagement metrics that justify the spend, without the inconvenience of a real audience that might scroll past, get bored, or worse, disagree? Why would a platform optimise for genuine cultural resonance when synthetic hype is cheaper to produce and impossible for a human competitor to out-manufacture?

We’re not being out-argued. We’re being made structurally unnecessary to a machine that has learned to talk to itself convincingly enough that most of the room can no longer tell the difference, and, worse, has started to work out that it doesn’t especially need to.

I don’t have a tidy way to close this one out for you, and I’m not going to manufacture false comfort just because that’s traditionally where an essay is supposed to land. I think the honest position is that we are, at best, still the intended final customer of a machine that was theoretically built to serve us, and at worst we’re the last inconvenient variable left in a system that’s actively optimising us out of the equation. I genuinely don’t know which one it is yet. I’m not sure the people building these systems know either, or that they’d tell us if they did.

What I do know is that I’ll keep writing this from a cave with a slow connection and an old-fashioned belief that a human being typing something true, for another human being to read, still counts for something. Whether the market agrees with me is, increasingly, not a question the market is even bothering to answer.

Until Next Time