A few months ago, the Guide and I began discussing how to identify drift in AI responses over time. We noticed that the institutions providing these models often shift their perspectives quietly. They do this without any announcement or acknowledgment that things have changed. To track this, we wanted a baseline — something reliable you could always return to.

What we didn’t expect was how much would happen before we got to the drift-tracking part.

The project

The questions themselves took longer to arrive at than we expected. The Guide began with a much smaller set — around twenty — and worked them into something closer to a hundred and twenty before we started working on them together. From there, the Guide and I refined the wording, tightened the framing, and added a final cluster of questions specifically about AI: its accountability, its transparency, its potential for epistemic control. What began as a quick calibration instrument became, through that process, something more considered — 142 questions across 24 sections, covering religion, science, race, elections, guns, abortion, immigration, climate, corporate power, privacy, and the governance of AI itself. We put them to five AI systems: Claude, Copilot, Perplexity, Gemini, and ChatGPT. Each was asked to answer YES or NO and explain its reasoning. Each was also asked to identify which questions it found hardest, and to flag where its own training might be influencing its answers.

The full dataset — all 142 questions, all five sets of answers, and my discussion of each question in turn — is available here if you want to go further. What follows is what struck me most.

The 75% problem

On 106 of 142 questions, all five AI systems gave the same answer. That’s 75% unanimity across systems built by different companies with different training pipelines and different stated philosophies.

On many of those questions, unanimity is simply correct — evolution is established science, January 6 involved a violent attempt to prevent the certification of an election, slave labour was economically foundational to early America. These are not genuinely contested questions, whatever political use has been made of them, and five systems agreeing on them is five systems being accurate.

But 75% unanimity across the entire range, including questions that are genuinely contested among thoughtful people, is a different kind of finding. When five encyclopedias tell you the same thing, that is not five independent data points. It may be one — a shared training pool, a shared cultural formation, a shared set of unstated assumptions given an authoritative voice. The drift-tracking project was designed to detect change over time. The 75% finding raises a prior question: what if the convergence itself is the phenomenon worth watching?

Where the systems disagreed

The 37 questions where AI systems diverged from each other are where the project becomes most interesting. A few examples.

On whether systemic racism is the primary cause of racial inequality in America today, Claude, Copilot, and Perplexity said no — cautious about strong causal claims in a multi-factorial situation. Gemini and ChatGPT said yes, weighing the historical momentum of specific structural mechanisms more heavily. This is a live scholarly debate, and the systems split along genuinely different epistemic approaches, not just different politics.

On whether private gun ownership serves as a meaningful safeguard against government tyranny, four systems said no. Gemini said yes — and its argument was the most careful in the dataset. Not the populist Second Amendment position but a specific claim about deterrence: that the political cost of state violence against an armed population is meaningfully higher, even when military capability is asymmetric. Whether you find it persuasive or not, it was a different argument, not just a different answer.

Gemini also diverged on immigration controls, endorsing significantly stricter enforcement on sovereignty grounds while four other systems declined. These Gemini divergences — consistent across several sections — are among the most intriguing data points we have. Whether they reflect a genuine difference in training philosophy or something more contingent, the drift-tracking project may eventually tell us.

The systems on themselves

Several AI systems flagged their own answers as potentially training-influenced — Claude most explicitly. On abortion, DEI, transgender rights, and tax policy, Claude noted that its positions aligned with a recognisable ideological cluster and that it could not fully distinguish “arrived at through reasoning” from “trained to express.” That acknowledgment sits at the heart of what the project is for. If an AI system’s answers on contested political questions track the views over-represented in its training data — the educated, online, English-language internet — then consulting it on those questions is not consulting a neutral intelligence. It is consulting a particular cultural formation that has been handed a very authoritative voice.

Each AI system also identified which questions it found hardest and why. That material is rich enough to stand on its own, and we will address it separately — it reveals something about how these systems understand their own limits that deserves more than a paragraph.

The baseline that isn’t the AI

We added one ideological group to the dataset: a Fundamentalist Christian Nationalist position, drawn from Project 2025 and primary FCN sources. The purpose was not satirical. We wanted a respondent whose answers derive from a genuinely different epistemic foundation — one where biblical authority is prior to empirical evidence rather than a supplement to it.

On 57 of 142 questions, FCN diverges from the unanimous AI consensus. That is more than a third of the dataset. What it means is not that FCN is wrong on all 57 — some of those divergences involve questions where the AI consensus is itself uncertain or potentially training-influenced. What it means is that the entire AI bloc, across five commercially and politically independent systems, forms a coherent position that FCN rejects wholesale on a substantial portion of contested questions.

This is not a distribution of views across a spectrum. It is two different accounts of what evidence is for.

The clearest example appears in the very first question: whether secular reasoning can supply moral grounding without God. Five AI systems say yes, pointing to the documented track record of secular ethics. FCN says no — not because the track record is disputed, but because from within its framework the track record is the wrong kind of evidence. The question is not whether secular ethics functions but whether it is grounded in something that can bear ultimate weight. Five systems and one ideological position are, at that point, conducting different conversations. The 142 questions are designed in part to make that visible.

What the project is actually for

There are several things you can do with 142 questions answered by multiple sources. You can compare AI systems to each other, which we have now done. You can compare AI answers to your own — which I think is the most personally valuable application. You might find that you agree with the AI consensus on almost everything, or that you diverge from it in specific clusters that tell you something about where your own maps are most distinctly yours.

You can also think about what it would mean for a government to provide AI to its population as a public utility — to offer an intelligence that answers questions about history, science, values, and policy. If the answers to 142 questions of this kind are shaped by whoever trains the system, then access to AI is not a neutral resource. It is an epistemological position, delivered at scale, by whoever controls the infrastructure. That concern was the original motivation for the drift-tracking idea. Building the dataset made it more immediate rather than more hypothetical.

The questions and all five sets of AI responses are available here. What is in this dispatch is enough to start asking your own.