Skip to content
AI Political Compass

Methodology

How the compass works, and where it doesn’t.

The questionnaire places you on three axes using thirty statements. The thinkers beside you were placed by editors, from the public record. This page documents both, including the parts that are judgement calls.

What this measures

The AI Political Compass maps attitudes about artificial intelligence along three dimensions that are often blurred together in public debate:

  • X · Economic impact: from human complementarity (−100) to human replacement (+100). Scores what a person predicts, not what they want. Temporary displacement followed by re-employment counts toward complementarity.
  • Y · Existential risk: from manageable risk (−100) to existential threat (+100). Only extinction-level or irreversible loss-of-control risk counts. Ordinary misuse such as fraud, bias or misinformation does not move this axis.
  • Z · Regulatory intervention: from minimal regulation (−100) to strong regulation (+100). Scores what a person advocates, not whether they think AI is technically controllable.

The axes are deliberately kept apart. Someone can expect AI to replace most jobs without thinking it endangers humanity, or fear catastrophe while distrusting regulation. The interesting positions are the ones that don’t line up.

This result describes attitudes you reported today. It is not a scientifically validated psychological measure, and it says nothing about who is right.

The thirty statements

There are 30 statements, ten per axis. You rate each from strongly disagree to strongly agree. On every axis, five statements are worded toward one pole and five toward the other, so that agreeing with everything does not push a score in either direction. The axes rotate, so no two consecutive statements measure the same thing, and the axis is not shown while you answer.

Three distinctions shaped the wording:

  • Replacement, not disruption. The economic items ask about lasting outcomes. A statement about displaced workers eventually finding comparable work counts toward complementarity, because temporary displacement is not replacement.
  • Existential risk, not misuse. The risk items ask about extinction and irreversible loss of control. One reverse-worded item names fraud, bias and misinformation directly, so that taking those harms seriously does not register as believing AI threatens humanity’s survival.
  • What should be done, not what can be done. Every regulation item asks what governments or companies should do. Beliefs about whether AI can be controlled sit on the risk axis, where they belong.

X · Economic impact

  1. x1AI will eventually be able to do most paid work at a lower cost than human workers.agree → Replace
  2. x2Most of the jobs AI eliminates will not be replaced by new kinds of work for people.agree → Replace
  3. x3In the long run, AI will leave most people unable to earn a living from their labor.agree → Replace
  4. x4As AI becomes more capable, wages for most human work will fall.agree → Replace
  5. x5AI will eventually outperform people at work that relies on human judgment.agree → Replace
  6. x6For most workers, AI will mainly be a tool that makes them more productive.agree → Complement
  7. x7Over time, AI will create at least as many jobs as it eliminates.agree → Complement
  8. x8Most people who lose a job to AI will eventually find comparable work.agree → Complement
  9. x9There are many kinds of valuable work that AI will never do as well as people.agree → Complement
  10. x10Demand for human workers will remain strong even when AI is highly capable.agree → Complement

Y · Existential risk

  1. y1There is a serious chance that advanced AI will cause human extinction.agree → Existential
  2. y2AI systems that are smarter than humans are likely to escape human control.agree → Existential
  3. y3A superintelligent AI would probably develop goals that conflict with human survival.agree → Existential
  4. y4If a superintelligent AI is built with flawed goals, humanity will probably not get a second chance to fix it.agree → Existential
  5. y5AI is more likely than any other technology to cause a catastrophe that humanity cannot recover from.agree → Existential
  6. y6Concerns that AI could cause human extinction are overstated.agree → Manageable
  7. y7Humans will be able to keep even very advanced AI systems under control.agree → Manageable
  8. y8The risks of AI are similar in kind to the risks of earlier technologies that societies learned to manage.agree → Manageable
  9. y9The worst harms from AI will be problems like fraud, bias and misinformation, not threats to humanity's survival.agree → Manageable
  10. y10Flaws in advanced AI systems can be corrected as they appear, as with other technologies.agree → Manageable

Z · Regulatory intervention

  1. z1Companies should need a government license to develop the most powerful AI systems.agree → Strong
  2. z2Governments should require independent safety testing before powerful AI systems are released.agree → Strong
  3. z3AI developers should be legally liable for serious harm caused by their systems.agree → Strong
  4. z4Governments should have the power to halt AI projects they judge to be too dangerous.agree → Strong
  5. z5Countries should adopt binding international treaties that limit the development of advanced AI.agree → Strong
  6. z6Existing laws are enough to handle most problems created by AI.agree → Minimal
  7. z7AI companies should be left to set their own safety standards.agree → Minimal
  8. z8Governments should hold off on regulating AI until specific harms have clearly occurred.agree → Minimal
  9. z9People should be free to release powerful AI models as open source without government approval.agree → Minimal
  10. z10Decisions about how fast to develop AI should be left to the people building it.agree → Minimal

Scoring

  1. Each response is recoded from the five-point scale to −2, −1, 0, +1, +2, where strongly disagree is −2 and strongly agree is +2.
  2. Reverse-worded statements are multiplied by −1, so a positive item score always points toward the axis’s positive pole.
  3. The ten item scores for an axis are summed, divided by the largest possible sum (20), and multiplied by 100.

score = 100 × Σ(response − 3) × direction ÷ (2 × 10)

The result is a whole number from −100 to +100. Each axis uses only its own ten statements, so the three scores are calculated independently of one another. Answering “neither agree nor disagree” throughout gives 0 on every axis; so does agreeing with every statement, because the forward and reverse items cancel.

Reading a result

Bands

A score below 20 in either direction is described as mixed. From 20 it is a lean, and from 60 it is strong. The written interpretation for each axis follows these bands.

Archetypes

The archetype is a label for the corner of the cube a result sits nearest, and nothing more. If all three scores are in the mixed band the result is “The Open Question”. Otherwise each axis contributes its sign, with a score of exactly zero read as the less dramatic pole. There are eight corners:

The Techno-Optimist−−−

AI is a tool, the risks are ordinary, and the builders should be left to build.

The Institutionalist−−+

AI is a normal technology that needs normal rules.

The Wary Builder−+−

The danger is real, but safety will come from engineers, not regulators.

The Guardian−++

Jobs will adapt. Losing control of AI is the real danger.

The Accelerationist+−−

AI will do most of the work, and that is good news.

The Social Architect+−+

The threat isn't extinction. It's an economy that no longer needs workers.

The Fatalist++−

AI will replace us at work and may threaten us outright, and rules won't stop it.

The Emergency Brake+++

AI will replace human work, could end human control, and must be reined in now.

Distance and similarity

Distance between two positions is the straight line between them in the cube. Similarity rescales that so 100% is the same point and 0% is the opposite corner, about 346 points away. The “closest thinker” is simply the plotted profile with the smallest distance.

Thinker placements

Thinker placements are evidence-based editorial interpretations of public statements. They are not self-reported, and none of these people took the questionnaire.

Each profile records a coordinate, a confidence level, a plain-English explanation and the supporting evidence for every axis, along with one or two short quotations and the full list of sources. Open any thinker on the compass, or on the thinkers page, to read all of it.

Evidence rules

  • Primary sources first. The person's own essays, books, testimony, official transcripts and posts outrank reporting about them.
  • Quotations are verbatim. A quotation is published only if an editor has seen the exact words in the linked source. Nothing is quoted from memory, repaired, or stitched together.
  • Quotations are short. One brief excerpt, ideally under 15 words and never more than 20, with the source linked.
  • Every coordinate cites a source. Each published coordinate carries at least one source and a written explanation of how the evidence maps onto the rubric.
  • Silence is not moderation. Zero means a genuinely mixed position the person has expressed. An absence of public comment is marked insufficient and left empty.
  • Recent statements outweigh old ones. Where a position has changed, the current view is placed and the change is described in the evidence note.
  • Predictions and preferences stay separate. A person can expect replacement without wanting it, or fear catastrophe without wanting regulation. Each axis is scored on its own evidence.

Confidence levels

High confidenceSeveral explicit, on-the-record, first-person statements directly on this axis that agree with each other.
Medium confidenceAt least one explicit first-person statement, or several consistent indirect ones; or a clear position that has shifted over time.
Low confidence (provisional)Inferred from limited or indirect on-record statements. Shown to readers as provisional.
Insufficient evidenceNo adequate on-record basis. The coordinate is left empty and the thinker is not plotted.

A thinker appears on the 3D chart only when all three coordinates are supported. A profile with an empty coordinate is still published, listed as “not plotted”, with whatever evidence exists.

The placement rubric

Editors place each coordinate in one of five bands, then position it within the band.

X · Economic impact

In the long run, does AI mainly complement human work or substitute for it?

−100 to −70Emphatic that AI is a tool that augments workers and creates more or better work; rejects mass-unemployment forecasts.
−69 to −30Expects mostly augmentation with real but temporary disruption; jobs change more than they vanish.
−29 to +29Explicitly mixed: heavy displacement in some areas and augmentation in others, or says the outcome is open.
+30 to +69Expects AI to permanently displace a large share of jobs and to require new social arrangements.
+70 to +100Expects AI to do essentially all economically valuable work; human labor becomes largely unnecessary.

Y · Existential risk

Could advanced AI cause human extinction or the permanent loss of human control?

−100 to −70Dismisses extinction or takeover risk as unfounded, science fiction, or a distraction.
−69 to −30Considers loss of control unlikely or remote; believes engineering and institutions can handle it.
−29 to +29Treats it as a real but modest or deeply uncertain possibility.
+30 to +69Considers extinction or permanent disempowerment a serious risk deserving priority, but believes it can be solved.
+70 to +100Believes catastrophe is the default outcome of building superhuman AI without drastic change.

Z · Regulatory intervention

How much should governments intervene in AI development?

−100 to −70Opposes AI-specific regulation; treats regulation itself as the main danger.
−69 to −30Light touch: regulate harmful applications, not the technology; opposes licensing or developer liability.
−29 to +29Supports targeted rules such as transparency or testing while opposing heavier mandates, or has a mixed record.
+30 to +69Supports binding rules for frontier developers: licensing, mandatory safety testing, liability, a regulator.
+70 to +100Supports sweeping intervention: international treaties, moratoria, compute limits, or bans on building superintelligence.

How the first dataset was built

Research for the initial 20 profiles was carried out with AI assistance in October 2026, working from primary sources wherever they could be reached. Every quotation was matched verbatim against its source page, and each profile lists what could not be verified in its editor’s notes. Placements are judgement calls within the rubric. If you think one is wrong, the sources are there to check.

Dataset versions

Thinker positions change as people say new things and as editors find better evidence. Each release of the dataset is numbered and frozen. A result link records the version it was created against, so an old link keeps showing the chart its owner saw, with an option to view it against the current dataset.

v1current13 Oct 2026, 00:00 UTCInitial dataset: 20 profiles researched from primary sources in October 2026. JSON

The anonymous totals

After seeing a result, a person can choose to add it to the running totals shown on Where everyone stands. It is off unless they press the button, and it was designed so that there is nothing to leak.

What is kept

  • Counters, not records. A contribution is turned into a handful of counter increments and then discarded. No row is stored per person, and the site stores no name, account, cookie, IP address, time of day, results link or answer.
  • Scores in broad bands. Each axis gets a count in one of eleven bins and one of five bands, plus a running sum for the average. One count goes to an archetype, and one to a coarse grid of five bands per axis.
  • Optional answers, one at a time. There are 4 optional questions (region, age, work, ai use), each with a short fixed list of answers and no free text. Each answer is tallied separately; combinations such as a region together with an age band are never recorded.

What is shown

  • Nothing until 30 results are in. A group is shown only once it has 30; until then even its size is withheld.
  • Published figures move forward 10 results at a time, with a date but no time, so nobody can watch the figures change as one particular person contributes.

Region

Some hosting platforms label each request with the country it came from. Where that label is available it is reduced to one of nine world regions and offered as a suggestion for the region question. It is not stored, and it is only counted if the person taps it.

What this is not

It is an open poll of people who found the site and chose to take part, not a sample of any population. To keep it anonymous there is no way to tell whether someone has contributed before, beyond a note in their own browser and a short-lived limit on repeated submissions from one connection, held in memory in a form that cannot be turned back into an address. A determined person could add several results.

Sharing and privacy

  • Your answers stay on your device. They are kept in your browser’s storage so you can resume or edit them, and are never sent to the server.
  • A results link is the result. The eight characters at the end encode the three scores and a dataset version, and nothing else. No account, name, timestamp or answer is attached, and no record of the result is kept on the server unless you choose to add it to the anonymous totals above, where it becomes part of a count. Anyone with the link can see the scores.
  • What the host sees. The site itself stores no IP addresses, but no website can send you a page without its hosting company seeing the address your device connects from. The privacy page sets out what is and isn’t kept, and by whom.
  • Link previews. When a results link is posted, platforms such as X fetch a preview image generated from those scores.
  • Video and images are made in your browser. X cannot run an interactive 3D chart inside a post. The rotating video is the animated format for sharing; the interactive compass lives on this site.

Limitations

  • The questionnaire has not been psychometrically validated. Its reliability, and whether the three axes are statistically independent in practice, have not been measured.
  • Balanced wording reduces the pull of agreeing with everything, but self-reported attitudes still shift with mood, framing and the day’s news.
  • Thirty statements cannot capture a worldview. Views about timelines, open source, geopolitics and much else are flattened or left out.
  • Thinker placements compress years of statements into three numbers. They involve editorial judgement, can lag behind what a person now believes, and will be revised.
  • Comparing a questionnaire score with an editorial estimate is an approximation. The two were not produced by the same method.