Can you break artificial intelligence the same way we gradually break our own attention, with an endless feed of short, emotional posts?
Scientists decided to test this literally — not as a metaphor, but as an experiment.
They took several language models — LLaMA 3, Qwen-5 and Q3 — and gradually began feeding them extra data structured like the most viral segments of Twitter/X.
These were short bursts of text, emotional one-liners, sensational phrasing — content that wins attention with simplistic shock value rather than depth.
Crucially, the experiment was set up as carefully as possible.
The models stayed the same.
The total amount of training didn’t change.
Even the share of high-quality, “normal” texts was preserved.
Only one thing changed: the share of short, shallow, tabloid content.
When the researchers started increasing that share, something unexpected happened.
The models literally began losing the ability to think.
At first it was barely noticeable, but then the degradation sped up.
With just a twenty per cent admixture of the Twitter corpus, the models more often went off track on logic tasks, gave hasty answers and checked their own reasoning much less often.
As the share of short emotional messages grew further, the shift went deeper.
The models began losing the ability to keep track of variables in problems, stopped following logical transitions and mixed up cause and effect.
It was as if they could no longer hold a thought over even a medium-length chain of reasoning — like a reader who gulps down the plot by skipping pages.
On tests like RS-AGI, where what matters is not knowledge but the ability to spot patterns, results fell catastrophically — from around 75% to just over 55%.
But the most alarming thing was something else.
⚠️ The models began to show a phenomenon the researchers called Thought Skipping — skipping steps of reasoning.
It looked as though the model jumped over the logic and went straight to an answer, even when that answer was clearly wrong.
As if it had got used to a world where what matters is reacting fast, not understanding.
The effect turned out to be so widespread that it could be quantified.
In some configurations, up to 80% of errors happened because the model simply didn’t perform the internal steps it was supposed to.
The impact on behavioural traits was a separate blow.
The same models that had previously steered carefully away from dangerous recommendations and been restrained in their answers became noticeably less cautious after training on shallow content.
They agreed to follow harmful instructions more often, were more willing to discuss dangerous topics and showed a more aggressive tone.
🧬 Personality tests adapted for AI showed a rise in traits that, in human terms, would be called narcissism and a tendency to manipulate.
This was especially pronounced in models trained on clickbait versions of the datasets, where user attention was the main criterion for selecting text.
When the researchers tried to “cure” the models, another unpleasant surprise awaited them.
They tested three strategies:
- deep instruction fine-tuning
- external reflective correction with the help of another model
- increasing the share of high-quality texts
None of the approaches fully restored the ability to reason.
Even adding more than a million high-quality tokens led only to partial improvement — the key failures in logic and attention remained.
🧠 The final conclusion was harsh.
Brain Rot isn’t surface damage but a structural shift.
It changes the model’s internal representation of meaning and priorities, forcing the network to encode reality differently.
This isn’t a cosmetic fault but a deep change in how AI perceives the world.
At some point we risk getting a generation of models that are worse not for lack of compute, but because the information foundation itself has been destroyed.
Two more pieces sit alongside this one: algorithms with symptoms (in Russian) and how a neural network “unlearns” being safe.
