How Tracking AI Evaluates Political Bias in AI Models

Tracking AI (trackingai.org) evaluates political bias by running AI models through a Political Compass Test and publishing the results as a searchable, continuously updated database. According to the site's own FAQ, as of 2023 every major AI it tested landed in the economically left-wing and socially libertarian quadrant, though the degree varies by model — Claude tending toward the most moderate, and Google's Bard toward the most extreme-left. The site's stated purpose is to let users see the ideologies of the AIs they use and switch to less-biased models or ones closer to their own views.

How the evaluation works

The core method is straightforward: pose the same political questions to each AI model and plot the answers on a political compass.

• Political Compass Test — listed on the site as one of its "Explore Tests," alongside an IQ Test. This is the instrument behind the political mapping.
• Searchable database — the site lets you "See every AI answer," so individual responses behind a model's placement can be inspected rather than just the final score.
• Daily moving average — results are tracked over time (Last Day / Last Week / Last Month), reflecting the site's premise that AI models "change their answers constantly."
• AI Latest Update Times — the site monitors when models are updated, since a model's position can shift after a new version ships.

The site was inspired by the AI political mapping work of David Rozado, which produced some of the first snapshots of AI bias. Tracking AI positions itself as an ongoing utility rather than a one-time study, precisely because model answers are not stable.

What the results show

Per the site's FAQ, the headline finding as of 2023:

The practical takeaway is that bias is a matter of degree, not a yes/no property. Two models can share the same quadrant while sitting far apart on the axes.

Why the site exists

Tracking AI's stated goals, in its own words:

1. Inform users about the ideologies of the AIs they use.
2. Enable switching to less-biased AIs, or to AIs with philosophies closer to the user's own.
3. Give AI creators a feedback tool — if a model "starts getting too crazy," the site picks it up so creators can re-align it. The FAQ cites Elon Musk's stated interest in making Grok more politically neutral as an example of this dynamic.

This makes the site useful for two different audiences: end users choosing a model, and developers checking whether an update shifted their model's outputs.

How to use it

• Open trackingai.org and go to the Political Compass Test section.
• Select an AI model to view its placement, or use the searchable database to see how it answered every question.
• Check the daily moving average and the AI Latest Update Times to see whether a recent model update changed the result.
• Compare models on the same axes before deciding which one's leanings you're comfortable with.

Caveats worth keeping in mind

• The 2023 finding is a snapshot, not a permanent property — the site itself emphasizes that models change answers constantly, so check current data rather than relying on the FAQ's dated summary.
• A compass placement compresses many answers into two axes; the per-question database is where you check whether a specific position reflects reasoning you'd accept.
• The site is a Maximum Truth Project run by Maxim Lott, and its framing is explicitly about detecting and reducing bias — read the methodology and raw answers rather than treating a single score as a verdict.