If you want to learn beautiful calligraphy, the best method is to copy the masters (like Wang Xizhi's *Lanting Xu*). But what if one day you decide: "Why copy the ancients? I'll just copy my own writing from yesterday." Day one looks fine; day two exaggerates a few strokes; by day one hundred, your characters have become a tangle of distorted lines.
In the AI world, this is called the "synthetic data loop." The worry: if all AI models learn from each other's generated text, will they collectively get dumber?
In May 2026, a research team from the University of Washington (UW) and Berkeley published an arXiv paper meant to reassure everyone: "Iterative Finetuning is Mostly Idempotent."
They mathematically proved a counterintuitive truth: AI is actually quite sober — it does not drift off course forever.
What Does "Idempotent" Mean?
In mathematics, an operation is *idempotent* if doing it once and doing it a hundred times produce the same result — like pressing the "off" button on a light switch repeatedly: the light stays off.
The paper's core insight: AI finetuning also has this "stop when good enough" property.
The Experiment: Forced "Self-Repetition" Training
Researchers put AI into a hard loop: 1. Generate a pile of synthetic data with the model. 2. Finetune the model on that data. 3. Repeat N times.
Previous fears: the model should grow increasingly extreme and absurd (e.g., super sycophantic or prone to hallucination).
Actual result: except for a few very specific parameter settings, after just 1–2 rounds of such "self-repetition," the model's intelligence and personality quickly stabilize — further training changes nothing.
Why Doesn't AI Get Dumber? (Two Key Brakes)
1. The "Gravity" of Logic
To push a quirk (like "praise the user constantly") to an extreme, the model must sacrifice coherence of its language. But training requires the output to remain coherent — this acts like gravity, pulling the model back from the abyss of craziness.2. The Magic of "Starting Over"
If each finetuning round re-initializes from the base "mother model" instead of compounding on yesterday's mess, the model retains a remarkable resilience.Is This "Learning Ceiling" Good or Bad?
- Good news: AI is not so fragile. We need not fear that once the internet fills with synthetic data, models will collectively degrade.
- Bad news: Brute-force self-training no longer works. Without feeding models fresh, novel knowledge from humans and the real world, training a model on its own outputs quickly hits the ceiling named "idempotence."
Summary
Intelligence cannot grow spontaneously in a vacuum.
AI is not a perpetual motion machine. It can patch small flaws through "self-reflection," but it cannot evolve into a god through "self-repetition." Next time someone claims AI will achieve infinite self-improvement, you can smile and whisper that mathematical word: "idempotent."
Truth comes from colliding with the unknown, not polishing the known — the deepest lesson of 2026's training theory about the limits of growth.