AI: The World's Biggest Game of Telephone — And We're All Losing

By Chris Daily

You played telephone as a kid. One person whispers a message. It passes down the line. By the end, the message is garbage. Everyone laughs.

Nobody's laughing now.

Because AI is playing the same game. Except the stakes aren't a silly misheard phrase. The stakes are the quality of information, knowledge, and language that future generations will have access to. And the message is already falling apart.

How the Game Works

Here's what's happening. AI models get trained on human writing — books, articles, websites, conversations. That training is what makes them smart. But now, AI is generating so much content that it's flooding the internet. And new AI models are training on that content.

Round one of telephone: fine. Round nine: pure noise.

Researchers at Cambridge, Toronto, and Oxford published this finding in Nature in 2024. They called it "model collapse." They took an AI model and trained it on its own outputs. Then trained a new model on those outputs. Then again. And again.

By generation nine, the model stopped making sense. It had been asked to write about medieval architecture. Instead, it produced lists of imaginary animals. Blue-tailed rabbits. Species that don't exist. The original message — coherent, detailed, human — had completely disappeared.

That's not a glitch. That's math. The message always breaks down when you pass it through enough hands.

The Edges Die First

Here's the part that should make you angry.

When the telephone game breaks down, it doesn't break down evenly. The weird parts go first. The specific parts. The rare parts. The parts that made the original message worth saying.

It works the same way in AI. Research shows that model collapse doesn't destroy the common, average, predictable stuff right away. It destroys the edges. The minority languages. The unusual writing styles. The cultural knowledge that doesn't fit the mainstream mold.

Think about what that means in the real world. A dialect spoken by a small community. A literary style that doesn't look like a press release. A way of understanding the world that comes from lived experience outside the dominant culture. These things don't slowly fade. They get cut off fast, in the early rounds of the game, while everything looks fine on the surface.

One researcher called this "Habsburg AI." The Habsburgs were a royal family that kept marrying within their own bloodline. After generations of it, their descendants were physically and mentally broken. The last one couldn't chew his own food. That's not dramatic storytelling. That's what happens when a system keeps feeding on itself with no outside input.

Same thing is happening here. Except instead of a royal bloodline, it's human knowledge and expression. And instead of one family, it's the entire internet.

The Telephone Line Is Already Broken

How bad is the contamination? Between 50 and 74 percent of new web content is now AI-generated. Not human-written. AI-generated. More than half. Possibly three-quarters.

And here's the kicker: none of the major AI training datasets filter for it. Not one. The people building the next generation of AI models are scooping up the whole web and training on it — synthetic content and all.

The telephone game is now playing itself.

Every time an AI generates content, that content hits the web. The next AI trains on it. That AI generates more content. That content hits the web. The next AI trains on that. Around and around, with a little more of the original message lost each time.

There's a name for this: the Ouroboros. It's the ancient symbol of a snake eating its own tail. It used to be a metaphor. Now it's a product roadmap.

Who Actually Pays the Price

Let's be clear about who wins and who loses here.

The big labs — the ones with billions in funding — will probably be fine. They have their own private data. They employ teams of people to clean and curate their training sets. They can afford to buy access to high-quality human-written content, and they will.

Everyone else? The smaller developers. The open-source community. The researchers in countries that can't compete in a bidding war over clean data. They train on whatever's free and available — which is the contaminated web. Their models degrade first and fastest.

The people with the most money have already built walls around the good stuff. Everyone else gets the copy of a copy of a copy.

Meanwhile, we're running out of the original source material entirely. According to Epoch AI, at the rate AI is consuming human-written text, we could exhaust the supply by 2027. We built machines that burn through human creativity faster than humans can create it. And when the tank runs dry, we're feeding them each other's exhaust.

We Knew This Would Happen

The honest truth? This isn't surprising. We've watched this exact pattern play out before.

Overfish a lake and the fish disappear. Drain an aquifer and the land dries up. Strip a forest and the topsoil washes away. Every time, the people doing the damage had short-term reasons that made perfect sense. Every time, the cost got passed to everyone else.

This is the same thing. Generating AI content is cheap. It scales fast. There's money in it. So everyone does it. But the thing they're cheaply generating at scale is slowly ruining the shared resource that made their product work in the first place.

The researchers call this the tragedy of the commons. The profit goes to individuals. The damage goes to everyone.

And no one is stopping it. Regulations are coming — the EU requires AI content labeling by August 2026 — but the web is already more than half synthetic. Detection tools top out at 90% accuracy on a good day. The message has already been passed down too many lines.

What We're Really Talking About

Strip away the research papers and the technical jargon, and here's what's actually happening.

AI was built on human expression. Real writing. Real thought. Real experience. The chaotic, specific, contradictory, irreplaceable record of billions of people actually living their lives and putting words to it.

That raw material cannot be faked. It cannot be generated. It has to be lived first.

When we replace it with AI output and then train the next AI on that output, we're not making something smarter. We're making a machine that whispers to itself in an empty room, convinced it still has something to say.

The telephone game only works if someone at the start has a real message. We had one. We're throwing it away.

And every round we keep playing, more of it disappears.

Sources: Shumailov et al., Nature 2024; Dohmatob et al., ICLR 2025; Alemohammad et al., ICLR 2024; Gerstgrasser et al., arXiv 2024; Guo et al., NAACL 2024; Epoch AI; Graphite/Ahrefs web content estimates.

---

Knowing what you're reading is becoming a real skill

If a growing share of the internet is a copy of a copy, the ability to evaluate what's actually in front of you — instead of just accepting it — is one of the most useful skills you can build right now. That's exactly what Think Critically with AI is built around: judgment and evaluation, not just tool usage.

See the course →