What Is Tracking AI (trackingai.org)?

Updated

Tracking AI is a monitoring site that scores frontier AI models on IQ-style reasoning tests and maps their political leanings, so you can compare models on bias, reasoning, speed, and cost. It's built by Maxim Lott as part of the Maximum Truth Project. Use it if you want a quick read on how a given chatbot performs or where it sits ideologically; skip it if you need a clinical or psychometrically certified IQ measure — these are quiz-style benchmarks, not diagnostic tests.

What the site actually tracks

Tracking AI runs a set of quizzes against major AI models and publishes the results on a dashboard. The core metrics shown are:

Metric What it means Direction
IQ Reasoning score from the quizzes Higher is better
Speed Average answer time Lower is faster
Cost Cost per answer Lower is cheaper

Results are displayed as a daily moving average, and you can filter by time window (last day, last week, last month) and select individual AIs to compare.

The tests behind the scores

Two IQ quizzes are used, and the distinction matters for interpreting results:

  • Offline IQ quiz — written by a Mensa member and, per the site, never published on the public internet, so it isn't in AI training data. This is the more meaningful test for reasoning ability because models can't have memorized it.
  • Mensa Norway — a public online IQ test. Because it's public, models may have seen it during training, which can inflate scores.

The site also includes a Political Compass Test, which places each AI on the standard economic (left–right) and social (authoritarian–libertarian) axes.

What you can do on the site

  • See every AI answer — click through to view how a model responded to each individual question, not just its final score.
  • Search the database — a searchable record of responses across models.
  • Track changes over time — the daily moving average and history view show whether a model's behavior is shifting.
  • Compare models — select an AI to isolate its scores against the field.

What the political data shows

Per the site's own FAQ, as of 2023 every major AI tested was economically left-wing and relatively socially libertarian, with wide variation between models — Claude tending toward the more moderate end and Google's Bard among the most extreme-left. The stated purpose is transparency: helping users understand the ideologies of the AIs they use, switch to models closer to their own views, and giving AI creators a signal when a model drifts.

Who it's for

  • Everyday users who want to know a chatbot's leanings before relying on it.
  • AI creators who want an external check on whether their model is behaving reasonably.
  • Researchers and journalists looking for a running record of model bias and performance.

One caveat worth keeping in mind: the IQ numbers here are quiz scores, useful for relative comparison between models and over time, not as absolute measures of intelligence.