Language Models Agree With Each Other, Not With Readers

The Great AI Echo Chamber vs. The Quiet Reader

Imagine you are in a room full of people trying to guess what a mysterious, invisible book is about. If you ask a group of humans to read a page and highlight the most important sentences, you'll get a messy, beautiful chaos. One person highlights a funny joke, another marks a sad fact, and a third circles a word they like the sound of. They are all reading for their own reasons, with no instructions, no prizes, and no one telling them what "important" means. This is how real reading works: it's personal, messy, and full of disagreement.

Now, imagine you replace those humans with a fleet of super-smart computers, all trained on the same massive library of the internet. You ask them the exact same question: "Which sentences in this story matter?" You might expect them to be just as messy as the humans, or perhaps even more diverse because they are different machines. But here is the twist: these computers are starting to sound exactly like each other. They are all picking the same sentences, with a level of agreement that no group of humans ever reaches. This paper investigates that strange phenomenon. It asks: Are these AI models just copying each other, or have they somehow all decided on a single "correct" way to read? And if they agree so much with each other, do they agree with us?

The Experiment: A Highlighting Battle

To find the answer, the researchers set up a unique game. They didn't hire people to act like test subjects or pay them to follow rules. Instead, they looked at a real social platform where thousands of regular people had been highlighting text on 120 different web articles for their own personal reasons—maybe to remember a fact, quote a friend, or just because a sentence felt right. Crucially, these readers couldn't see each other's highlights; they were reading in total isolation.

Then, the researchers asked 18 different AI models (from big companies like OpenAI, Google, and Anthropic, ranging from small to super-powerful "frontier" models) to read the same 120 articles and pick out the top 20% of sentences that seemed most important. They compared how much the AIs agreed with each other versus how much the real humans agreed with each other.

The Big Discovery: AIs Are Best Friends (But Not With You)

The results were startling. When two different AI models read the same article, they agreed on which sentences were important about 2.3 times more than two random human readers agreed with each other.

To put that in perspective:

• Two humans reading the same page might share about 4.1 highlighted sentences out of a typical set of 14.
• Two AI models reading that same page shared about 8.7 sentences.
• Even more shocking: The two most powerful AI models from rival companies (GPT-5.4 and Claude Opus 5) agreed with each other 5.1 times more than two humans did. In fact, they agreed with each other almost as much as a single human agrees with themselves when reading the same text twice!

The paper calls this "convergence." It means the AIs are all marching in lockstep, picking the same "best" sentences, while humans are all marching in different directions.

What It's NOT (Ruling Out the Wrong Explanations)

The researchers were very careful to make sure this wasn't a trick. They tested and ruled out several obvious reasons why the AIs might be agreeing so much:

• It's not because they are the same company: Models from totally different labs and countries agreed just as much as models from the same company.
• It's not because they are "deterministic" (robotic): Even when the same AI was asked the exact same question twice, it didn't always pick the exact same sentences (it only agreed with itself about 27% of the time). But when it talked to a different AI, they agreed way more than that.
• It's not because of the prompt wording: Changing the words of the instruction slightly didn't stop them from agreeing.
• It's not because they are copying a "correct" answer: The task didn't have a right or wrong answer. The AIs were just picking what they thought was important.

The "Register" Mystery: Do They Like Different Words?

One might guess that AIs agree so much because they all love long, fancy, scientific sentences, while humans love short, emotional ones. The researchers checked this by looking at 11 different features of the sentences (like length, punctuation, and word types).

They found that when they matched sentences by length and position, there was no difference. The AIs and the humans were picking sentences that looked exactly the same on the surface. The AIs weren't choosing "robot words"; they were choosing the same types of words as humans, just a completely different set of sentences. The distance between what AIs pick and what humans pick is growing, but it's not because they are speaking different languages.

The Ceiling: AIs Are Running Out of Room to Be Human

Here is the most interesting part of the story. The researchers asked: "As AI gets smarter, does it start agreeing more with humans?"

The answer is yes, but only up to a point.

• Newer, smarter models do agree more with humans than older, dumber ones.
• However, they hit a "ceiling." Even the best AI in the study only reached about 84% of the agreement level that a perfect "crowd consensus" (the average opinion of many humans) could reach.
• Meanwhile, the agreement between two AIs keeps climbing, passing the human ceiling and going way higher.

It's like the AIs are learning to read like humans, but they are also learning to read like each other even faster. They are becoming so good at mimicking a specific "ideal" reader that they are leaving actual humans behind.

The Procedure Problem: A Fair Fight?

The paper also admits a quirk in how the game was played. When a human highlights text, they might mark 20 sentences, but the researchers only let them keep the top 14 (a random selection of their best). The AI, however, was forced to pick its absolute best 14 sentences. This gave the AI an unfair advantage, like a basketball player who gets to pick their best 14 shots while the human has to pick 14 shots at random from a pile of 20.

When the researchers made the AI play by the human's rules (picking randomly from its top choices), the gap between AI and human agreement shrank by half. But even then, the AIs still agreed with each other significantly more than humans did. The gap didn't disappear; it just got smaller.

The Takeaway

The paper concludes that language models are not just tools that help us read; they are becoming a new kind of reader entirely. They are converging on a shared, internal way of understanding text that is much more uniform than how real people read. While they are getting better at understanding us, they are getting even better at understanding each other. The result is a world where AI models might agree with a rival AI from a different company more than they agree with the person reading the text. It's a powerful, slightly eerie echo chamber where the machines are all humming the same tune, while the humans are still singing their own unique songs.