I asked nine AIs to map Middle Eastern politics

The political compass you know — left to right, authoritarian to libertarian — came out of a very specific argument. “Left” and “right” are just where deputies happened to sit in a French assembly in 1789. The economic version we use today is a story about the state and the market: how much should the government own, tax, redistribute. It’s a good map for the place and the century that drew it.

It travels badly.

Drop that map on the Middle East and most of the interesting arguments fall straight through the cracks. The loudest fights in the region aren’t about marginal tax rates. They’re about religion and the state. About the regime — are you with it or against it. About which way a country faces: Tehran, Washington, Riyadh, no one. A party that reads as “moderate” in one country is the regime’s party; the same word across the border belongs to the people the regime put in prison. “The left” in Cairo, Beirut, and Baghdad is three different animals wearing one label.

I kept wanting a map that took the region on its own terms, and I couldn’t find one I trusted. So I built it. It’s called Tayyar — تيّار, Arabic for “current,” as in a political current — and it lives over here.

Small confession up front: I read these platforms in the languages they’re written in. Arabic is my first language and Hebrew is around me every day, so the whole thing renders right-to-left and keeps the original text sitting next to the translation. That mattered to me more than it probably should have.

§The awkward part: who decides where a party sits?

Here’s the problem with any political map. Someone has to place the dots. And the instant a person places them, you are looking at that person’s politics at least as much as the party’s. One coder is one worldview. A roomful of coders is slow, expensive, and usually still shares a worldview.

So Tayyar does something that sounds, I’ll be honest, slightly unhinged the first time you say it out loud: it asks a panel of large language models to score the parties.

Not one model — a panel. Nine of them, five built in the United States and four in China. Each one reads a party’s documented positions, against a fixed rubric, and places it on every axis from −10 to +10, with the party’s name kept out of the way wherever possible so it’s scoring the substance, not the brand. The published number is the median of the panel — the nine models, plus a hand-coded baseline of my own tossed into the same pot as one more anchor.

I know how that lands. Asking chatbots to grade political parties is not self-evidently a serious thing to do. Two reasons I think it holds up anyway:

First, it shows its work. Every single score opens up to reveal the reasoning behind it and the exact passages the model leaned on. Click any cell and you can see why. It’s reproducible — run it again, get the distribution again. A lone human expert can’t hand you their priors as a file you can audit.

Second, and this is the part I didn’t expect to enjoy: when the models disagree, that isn’t noise to be averaged away. It’s a measurement. More on that in a minute.

None of this makes the numbers true. They are estimates of documented positions, not verdicts on anyone, and emphatically not endorsements — the entire thing is built to describe what a party says and does, never whether it’s good. The panel agrees with itself at around 0.86 on the standard reliability scales: solid, not scripture. This is a v0.2. Treat it like a draft that argues back.

§What the data actually says

As I write this, Tayyar holds 98 parties (83 of them placed on the full axis system), 274 politicians, across 20 countries and 16 axes, grounded in 286 source documents. Here’s the part I find genuinely interesting.

Look at where the bars are widest. The region splits hardest on religion and the state, on the regime, and on alignment with the West. Those are the axes where parties are flung clear across the scale — a real argument, with populated opposite ends.

Now look at the bottom of the chart. The economy — statist versus market, the exact thing a Western compass treats as the axis — is the tightest well-covered axis in the whole dataset. Parties bunch up near a mild-statist center and mostly don’t fight about it. The question the imported map makes the main event is, here, close to settled.

That one fact is the whole reason a borrowed left-right line flattens the region. It spends its entire horizontal resolution on the question everyone roughly agrees about, and crushes the three or four questions they’d actually come to blows over into a vague “social” smudge.

A couple of other things fall out of the numbers:

• The closest thing to a regional consensus is the Palestinian question. It’s the most lopsided axis in the dataset — parties lean hard toward pro-Palestinian-rights positions, more one-sidedly than on any other single issue.
• There’s a quiet, broad lean toward the centralist state (against federalism), and a more sobering one toward patriarchal traditionalism on gender. Those are consensus directions, not fault lines. The region argues loudly about God and the regime; it largely agrees on the strong central state.

§Now zoom into one country

Averages across twenty countries hide as much as they show. So let me do the thing the tool is actually for, and zoom into one place. Israel — which carries 27 parties in the dataset, 19 of them placed — is the most internally divided country in the whole set, and it divides in a way that turns out to be very revealing.

Start with the headline regional finding: the Palestinian question is the region’s most one-sided axis. Parties across twenty countries lean hard, together, toward pro-Palestinian-rights positions. Now look at the same axis inside Israel alone:

That flip is the whole point of looking country by country. The question that unites the region is the question that splits Israel down the middle, all the way to both poles. And it isn’t alone — Israel’s two deepest internal divides are exactly the two you’d guess if you’ve followed its politics for five minutes: the security cleavage (the Palestinian question, a standard deviation of 7.8, parties stretched from −10 to +10) and the religious cleavage (state and religion, 7.1, sitting dead-center because the secular and the devout are both fully represented and pulling in opposite directions).

And there, again, at the bottom: the economy. Even in Israel — even with a famously combative politics — statist versus market is the calmest well-covered axis, tighter than the region’s. Israeli elections are not fought over the size of the state. They’re fought over religion and security. Which is exactly why forming an Israeli government is the hardest coalition puzzle in the democratic world, and exactly why I couldn’t resist building a machine for it.

§What you can build on a map like this

Once you have every party placed on sixteen axes, you can do more than draw a map. You can simulate the thing the map is really about: power.

So there’s a second tool, a Knesset election simulator. It looks nothing like Tayyar — it’s a cool-blue election-night situation room rather than a warm research notebook — but underneath, it runs on Tayyar’s numbers.

The mechanics are the real thing. You slide each list’s share of the vote; anything under the 3.25% threshold wins nothing; the rest go through D’Hondt into a 120-seat hemicycle, and a meter tracks the 61 seats you need to govern. That part is just electoral arithmetic, and it’s exact — feed it the real 2022 vote and it reproduces the real Knesset, down to the surplus-vote pacts.

The part that uses Tayyar is the coalition. When you assemble a government, the simulator doesn’t just ask do they have 61? It asks do these parties actually agree on anything? — by reading their positions straight out of the dataset and scoring the coherence of the group. And here the data earns its keep. A coalition of the religious right (its lists tightly bunched on state-and-religion and on the Palestinian question) lights up as coherent. The 2021 “change” government — secular centrists plus the secular hard-right plus an Islamist Arab party, a coalition that spanned nearly the entire 7-point spread on both of Israel’s deepest axes — lights up as exactly what it was: a government that agreed on one thing (who shouldn’t be prime minister) and nothing else. The arithmetic says 61. The positions say good luck. That gap is the whole story of the last five Israeli elections, and you can watch it on a coherence meter.

Then there’s the forecast. Hit “Run 6,000 elections” and it runs a Monte-Carlo: each simulated election perturbs the parties along three latent factors — a national left-right swing, a religious-secular swing, and Arab turnout — and the amount each list moves with each factor is loaded from its Tayyar position. So ideologically close lists rise and fall together, the way real polling moves, instead of drifting independently. It isn’t a vibe. It’s calibrated on 63 final-week polls from the 2019–2022 elections: tuned so the model’s per-party error matches the real one (about 1.66 seats), with the measured polling biases (Haredi and Arab lists tend to be under-counted) corrected behind a toggle you can switch off. In backtests, the actual result lands inside the model’s 80% band 85% of the time.

My favorite feature is the one that tries to embarrass it. You can rewind to a past election and ask could the model have called it? — feed it only that race’s final-week polls and let it run blind. It calls the 2022 right-religious sweep comfortably. It’s honestly, visibly wrong about the 2019 deadlock — it over-counts a centrist list toward the right-wing bloc that, in reality, refused to sit with them — and it shows you the miss rather than hiding it. A model that only ever agrees with the past is just memorizing it.

A detail I love, because it captures how knife-edge this all is: in 2022, Meretz missed the 3.25% threshold by 0.09 points and Balad by 0.34. Two left and Arab lists, a third of a percent short, and their seats evaporated and flowed to the bloc they oppose. Nudge either of them over the line in the simulator and the whole map changes color. That is not a rounding error. That is how the most right-wing government in Israel’s history got its majority.

None of this is a prediction. It’s an illustrative model, the same way Tayyar is an illustrative map — built to make the arithmetic and the cleavages tangible, so you can put your hands on them, not to tell you who wins.

§The models give themselves away

Back to the disagreement, because this is my favorite thing the project turned up.

Take one party. Ask a model built in San Francisco and a model built in Hangzhou to score it. On the declared axes — what the party literally says about itself — they agree closely. On the interpretive axes — how committed it really is to democracy, what its posture toward the regime amounts to, what its identity politics mean — they pull apart. And the split isn’t random. It tracks where the model was built.

In other words, the rater’s own priors leak into the scores, and you can see the seam. Most political datasets bury that: one lab, one method, the priors baked in and invisible. Tayyar runs a whole little lab on it — US-built versus China-built, agreeing on the text and diverging on the judgment. It’s the one thing this dataset has that almost nothing else does, and it only exists because I refused to use a single rater.

§Go poke at it

The genuine joy of building something like this is that you can just play with it:

• The compass — swap the axes, filter by family or country, watch the whole field rearrange.
• The map — the region as a choropleth, shaded by whichever axis you pick.
• The Knesset simulator — slide the vote, build a coalition, run six thousand elections, rewind history.
• A handful of games: guess who said it, or call a head-to-head before the panel does.
• Or paste your own text — a speech, a manifesto, a tweet, in Arabic, Hebrew, or English — into the analyzer and watch it get scored live, with the lines it underlined.

It’s a pilot, and it’s wrong in places. I would much rather it be wrong in public, where you can tell it where it’s wrong, than confident in private.

So here’s my actual question. Somewhere in there, a language model has misjudged a party you know better than it does. Which one — and which way did it get it wrong?