Brain with noise and downward arrow. Image generated by Google Gemini 3 Pro.

Is It Dumb?

Crowdsourced tracker for when LLMs feel "dumbed down"

About This Project

LLM providers are known to change model behavior, reduce token usage, modify thinking budgets, or adjust reasoning capabilities during high loads or for testing purposes. These changes are often silent and not obvious to the user. The only symptom a user might notice is that a specific LLM they rely on feels "more dumb" than they are used to.

Our goal is to track these fluctuations through community statistics. If we see a spike in "dumbness reports" for a specific model, we can deduce that the provider might have "dumbed down" the model or is experiencing performance degradation.

Why This Works

Even when a model has not been dumbed down, we expect some statistical noise: users may occasionally report a model as "dumb" due to subjective perception or edge cases. Let's call this the baseline probability:

P(report | model = not dumb)

However, if a model has been dumbed down, we expect significantly more reports:

P(report | model = dumb) > P(report | model = not dumb)

With enough community participation, a spike in the timeseries graph becomes a strong signal that something has actually changed with the model, rising well above the false-positive baseline noise.

Report Trends by Provider

Track reported quality concerns across all providers and models

...

Loading charts for all providers...