Daniel sent us this one, and it's a layered prompt. He's been thinking about comparative policy analysis — specifically, how countries learn from each other when tackling entrenched social issues. He uses rental and tenancy law in Israel as the example, but the real question is broader. What's the traditional methodology for this kind of cross-country legislative review, what are the common pitfalls, and then — once we understand how it's been done — could agentic AI actually create something like a distributed think tank that does this work faster and cheaper? He's asking us to understand the old way before we get excited about the new way.
I love that framing, because the instinct to say "AI can do this now" without understanding what "this" actually is — that's how you get solutions that solve the wrong problem. Comparative policy analysis has a real name, by the way. In the academic literature, it's called comparative public policy, sometimes cross-national policy transfer, sometimes policy learning. The OECD has an entire directorate devoted to it. They've been doing structured cross-country policy comparisons since the nineteen sixties.
That's a very optimistic name for what's often a exercise in discovering how badly your own country has messed up.
That's the thing. Policy learning can mean learning what not to do. But the formal methodology is genuinely rigorous. There's a framework that a scholar named Richard Rose developed back in the early nineties — he called it "lesson-drawing." He identified something like seven different gradations of how countries borrow policy from each other, ranging from outright copying to pure inspiration where you just use another country's approach as a starting point for thinking differently.
Of course there are.
And the point is, most people imagine this as "let's see what Germany does and do that," but actual policy transfer almost never works that way. The classic pitfall — and this shows up in the literature going back decades — is what they call "institutional naivete." You look at a policy that works in Sweden, you import the text of the law, and you ignore the fact that Sweden has a totally different court system, different cultural norms around landlord-tenant relationships, different tax structures, different everything. The policy isn't the text. The policy is the text plus the institutional ecosystem it lives in.
It's like transplanting a tropical plant into a desert and saying "well, the seed is the same.
And the plant dies, and everyone blames the seed. This is the single most common failure mode in comparative policy. Dolowitz and Marsh wrote the foundational paper on this in nineteen ninety-six — they called it "policy transfer and policy failure." They identified three categories of failure. Uninformed transfer, where the borrowing country didn't understand how the policy actually worked in its home context. Incomplete transfer, where they didn't bring over the crucial supporting elements. And inappropriate transfer, where they brought the whole thing over but it just didn't fit their local conditions.
Daniel's question about how this is done traditionally — walk me through what a parliamentary researcher actually does when they're tasked with this.
Alright, so imagine you're a researcher for a member of Knesset who wants to reform rental law. Step one is usually a literature review — you survey the academic work on tenancy law across jurisdictions. But here's the thing that surprises people: the best comparative policy work almost never starts with "let's look at all the countries." It starts with a typology. You categorize countries by relevant institutional features — common law versus civil law systems, unitary versus federal states, high versus low homeownership rates. Because if you don't control for those structural variables, you're comparing apples and orangutans.
You'd group Israel with other countries that have a similar legal inheritance, similar housing market structure.
Israel's legal system is a hybrid — it's got British mandate common law foundations with Ottoman land law remnants and a very active Supreme Court that does things differently than either. So the naive comparison set would just be "other developed countries," but the rigorous set might be "common law jurisdictions with high urbanization and a housing supply crunch." That narrows things considerably. And that narrowing is actually the hardest intellectual work in comparative policy — deciding what's comparable.
Because if you pick the wrong comparison countries, you can justify basically anything.
Cherry-picking jurisdictions is the original sin of bad comparative policy. You want rent control? Look at Berlin. You want deregulation? Look at Texas. You want something in between? There's a country for that too. The methodology only works if you pre-commit to your selection criteria before you know what the answers are. This is why the OECD's regulatory policy reviews follow such rigid protocols — they use standardized indicators, they compare across a fixed set of dimensions, and they publish the methodology before the findings. It's designed to prevent exactly that kind of motivated reasoning.
The traditional process has built-in safeguards against "I already know what I want, now find me countries that support it.
Built in, but not always followed. And that's the second major pitfall. The first was institutional naivete. The second is confirmation bias in country selection. The third — and this one's more subtle — is what I'd call legal text fetishism.
Legal text fetishism.
The tendency to read the statute and assume that's what happens on the ground. Germany is actually a perfect example. Daniel mentioned Germany as having strong tenant protections, and it absolutely does. German law prohibits eviction without cause — you can't just decide not to renew a lease because you feel like it. Tenants have a legal right to indefinite renewal. But here's where the text-versus-reality gap shows up: the law also allows eviction for "Eigenbedarf" — personal need. If a landlord or their family member needs the unit, they can terminate the lease. And in practice, Eigenbedarf claims have become a major loophole. A study from the German Tenants' Association found that in cities like Berlin and Munich, a significant portion of contested evictions involve disputed Eigenbedarf claims where tenants allege it's pretextual.
You read the German statute and think "no eviction without cause, amazing," but the reality is landlords have found the pressure-release valve.
And this is why the best comparative policy researchers don't just read laws. They interview practitioners. They talk to tenant unions and landlord associations. They look at court dockets to see what's actually being litigated. This is the part that Daniel was getting at when he mentioned getting on a plane — the desk research tells you the formal rules, but the informal norms and enforcement patterns tell you the real rules.
Which is also the part that's hardest to replicate with AI.
We'll get to that. But first, let me give you a concrete example of comparative rental policy done well. In two thousand nineteen, the OECD published a major housing policy review that compared rental regulations across something like thirty countries. They looked at rent control mechanisms, tenant-landlord dispute resolution, eviction procedures, lease duration requirements, deposit rules — the whole package. And rather than ranking countries, they created a taxonomy. They identified clusters of approaches. The German model, which emphasizes tenure security and rent stabilization. The Swedish model, which uses collective bargaining between tenant associations and landlord organizations to set rents. The Swiss model, which treats housing cooperatives as a major third sector between renting and owning. The Anglo-American model, which is more market-oriented with lighter regulation.
Where did Israel fall in that taxonomy?
That's the thing. Israel wasn't in that particular review in depth, but if you map Israel onto those clusters, it's closest to what some researchers call the Mediterranean model — high homeownership, a small and under-regulated private rental sector, and weak tenant protections relative to Northern Europe. Countries like Spain, Italy, and Greece share similar patterns. High homeownership isn't an accident — it's often a rational response to poor rental protections. If renting means you can be evicted at will, you'll do anything to buy.
Which is a knock-on effect that bad rental policy doesn't advertise on the label.
That's the kind of insight that good comparative policy work surfaces. Once you see that homeownership rates and rental protections are inversely correlated across countries, you realize you're not looking at isolated policy choices — you're looking at a system. Change one part, and other parts shift in response. This is why simplistic policy borrowing fails. You can't just drop German eviction protections into the Israeli market without also thinking about what happens to the supply of rental housing, what happens to small landlords, what happens to the construction sector.
Because German landlords have operated under those rules for generations. They've priced the risk in. Israeli landlords would see it as a sudden expropriation.
And sudden regulatory shifts create their own pathologies. If you announce tomorrow that eviction without cause is illegal, you don't just protect tenants — you also cause some number of landlords to exit the rental market entirely, convert units to short-term rentals, or become much more selective about tenants in ways that might be discriminatory. None of which shows up in the text of the German law you're borrowing.
The traditional methodology is: rigorous country selection, deep institutional analysis, on-the-ground interviews, and a systems-thinking approach that looks at knock-on effect. That's the gold standard.
That's the gold standard. And it takes months, sometimes years. A full OECD country review involves multiple research trips, stakeholder consultations, expert panels, draft reports, peer review. The cost can run into the hundreds of thousands of dollars. Which is why most parliamentary researchers can't do the gold standard. They do a literature review, maybe a few video calls with counterparts in other countries, and produce a briefing note in a few weeks. The quality difference between those two products is enormous.
That's the gap Daniel is pointing at. The gold standard is too expensive and slow for most actual policymakers, and the quick version is prone to all the pitfalls you just described. Enter AI agents.
Here's where it gets interesting. Daniel's instinct about AI's potential here is actually well-founded. I've been following some work out of the Brookings Institution and the Tony Blair Institute — both have been experimenting with AI-assisted policy research. The idea isn't that AI replaces the researcher. It's that AI does the part that's currently impossible at scale.
The breadth part. A human researcher can go deep on three or four countries. An AI agent can systematically survey the legislative texts, regulatory frameworks, and published court decisions across fifty countries in an afternoon. Not with perfect accuracy — we'll get to the caveats — but with enough coverage to spot patterns that a human would never find because no human has time to read Estonian tenancy law on the off chance it's relevant.
It's a discovery engine. It tells you where to look.
And this is already happening in adjacent fields. There's a project called Policy Synth that came out of the Harvard Kennedy School — it uses AI to map policy interventions across jurisdictions and identify which ones have evidence of effectiveness. In the healthcare space, there are AI tools that scan clinical guidelines from dozens of countries to identify areas of consensus and divergence. The technology for cross-jurisdictional policy scanning exists. It's just not yet packaged for parliamentary researchers.
Why shouldn't we all be doing this? Daniel's question.
Several things break. First, AI is still bad at the gap between law-on-the-books and law-on-the-ground. It can read the German Civil Code section on Eigenbedarf. It can even find academic papers discussing Eigenbedarf abuse. But it can't go have coffee with a Berlin tenant organizer and learn what actually happens in practice. That tacit knowledge, the unwritten norms, the enforcement patterns — those aren't in the training data.
The AI gives you the formal architecture but misses the plumbing.
And in comparative policy, the plumbing is often the whole story. Singapore's housing policy is famous — something like ninety percent homeownership, the Housing Development Board controls most of the land. But you can't understand why it works without understanding the role of the Central Provident Fund, the ethnic integration policies, the fact that Singapore is a city-state with no rural-urban divide. The AI can list all those facts. It can't integrate them into the kind of causal model that a good policy analyst builds in their head.
Hallucination and sourcing. If you ask an AI to compare eviction laws across fifteen countries, it will produce something that looks authoritative. Some of it will be right. Some of it will be subtly wrong — citing a regulation that was amended three years ago, or conflating federal and state-level rules in a federal system. And if the parliamentary researcher doesn't have the expertise to spot those errors, the AI becomes a very efficient misinformation machine.
Like a research assistant who's read everything and understood about eighty percent of it, but won't tell you which eighty percent.
That's exactly the analogy. And the eighty percent problem is worse in policy than in other domains because policy details are so path-dependent. A single clause in a housing act can completely change the incentive structure. If the AI misses or mischaracterizes that clause, the entire comparison is misleading.
The agentic think tank idea — you've got AI agents doing the scanning, agents doing the synthesis, producing an analytical product. What's the minimum viable version that's actually useful rather than dangerous?
I think the minimum viable version keeps a human in the loop at three specific points. First, human-designed research questions and country selection criteria — the AI shouldn't choose what's comparable, because that's where the motivated reasoning creeps in. Second, human verification of key factual claims — not every claim, but the load-bearing ones that the analysis depends on. And third, human integration of the findings into a causal narrative that accounts for institutional context.
The AI does the grunt work — the literature scanning, the legislative text comparison, the pattern detection across jurisdictions. And the human does the sense-making.
That's the model. And I think it's powerful. A single parliamentary researcher with good AI tools could produce something in a week that currently takes a team three months. Not as good as the OECD gold standard, but dramatically better than the rushed briefing note that most legislators actually get.
Which brings us to the question Daniel embedded in all of this — the meta-question about what we call this practice. Comparative legislative review, cross-national policy analysis, something else?
In the academic literature, the umbrella term is "comparative public policy." But I think what Daniel is describing is slightly different — it's more applied, more action-oriented. Some people call it "policy benchmarking." The World Bank uses "global knowledge transfer." The EU has something called the "Open Method of Coordination" where member states compare policy approaches and identify best practices without binding harmonization. That last one is actually the closest to what Daniel's describing — structured comparison aimed at voluntary improvement, not mandated convergence.
Open Method of Coordination. Sounds like a euphemism for "we compared notes and everyone agreed the Germans do it better.
It kind of is. But it's a euphemism with a methodology behind it. The EU process involves common indicators, national action plans, peer review, and periodic reporting. It's not binding, but it creates soft pressure through transparency. Countries don't want to be at the bottom of the league table.
That's a dynamic that an AI-powered version could actually enhance. If you could generate comparative dashboards across jurisdictions quickly and cheaply, you'd make the league table visible in near real-time.
And that transparency has political effects. When the OECD publishes a ranking of education systems, countries that rank poorly face domestic pressure to reform. The ranking itself becomes a policy actor. An AI system that could generate credible cross-country comparisons on demand would accelerate that dynamic significantly. Whether that's good or bad depends on whether the comparisons are actually credible.
Let's talk about the specific domain Daniel raised. Rental and tenancy law. If you're that parliamentary researcher in Israel, and you've got AI tools, where do you actually look?
Alright, so let me give you a quick tour of the global landscape, because this is where the comparative method gets concrete. Germany, as we mentioned — indefinite leases, rent control through the Mietpreisbremse, the rent brake, which caps rent increases in tight housing markets at ten percent above the local comparable rent. Eviction only for cause, with Eigenbedarf as the contested exception. Sweden takes a completely different approach — there's no statutory rent control in the same way, but rents are collectively negotiated between the Tenants' Union and landlord organizations. It's a corporatist model that keeps rents predictable without heavy-handed government price-setting. The Swiss model is fascinating — they have a massive cooperative housing sector. Something like twenty percent of rental housing in Zurich is owned by cooperatives or non-profits. These aren't government housing — they're member-owned, but they operate on a cost-rent basis rather than for profit. That changes the entire dynamic.
The Anglo-American model?
Much lighter regulation. In most US states, there's no rent control, leases are typically twelve months, and after that the landlord can choose not to renew without cause. The UK is slightly more regulated — Scotland, for example, has recently moved to indefinite tenancies with restricted grounds for eviction. But the overall Anglo approach treats rental housing as a market good first and a social good second. The tenant protections that exist are mostly procedural — notice periods, deposit protection schemes — rather than substantive limits on eviction or rent increases.
Where has Israel looked for models historically?
Israel's rental law is actually quite minimal. There's no comprehensive tenancy act comparable to Germany's. The regulatory framework is mostly in the realm of contract law — the lease agreement governs, and the courts have developed some tenant protections through case law, but there's no statutory right to lease renewal, no rent control on new tenancies, and eviction procedures are relatively straightforward for landlords. The system assumes that renting is transitional — something you do before you buy. The problem is that for a growing number of Israelis, it's not transitional. It's permanent.
Because buying is impossible.
Because buying is impossible. And the legal framework hasn't caught up to that reality. So a comparative review would look at countries that have made that transition — from treating renting as temporary to treating it as a legitimate long-term tenure. Germany made that transition in the post-war period. Sweden made it. The Netherlands made it. Those are the natural comparators.
The AI could pull all of this together — the legislative texts, the court decisions, the statistical outcomes, the academic evaluations.
But here's the part I keep coming back to. The most valuable thing in comparative policy isn't the information. It's the judgment about what's transferable. And that judgment requires a theory of why something works in its home context. The German rent brake works — to the extent it does — because Germany has a well-functioning system of local rent indices, the Mietspiegel, that establish comparable rents. If you don't have that administrative infrastructure, the rent brake is just an arbitrary cap that distorts the market. The AI can tell you the rent brake exists. It can tell you about the Mietspiegel. What it can't do is tell you whether Israel could build a Mietspiegel equivalent, how long it would take, what it would cost, and whether the political coalition exists to fund it.
The AI is a librarian, not a strategist.
A very fast librarian with an encyclopedic memory and occasional confabulation problems. Which is useful. Librarians are useful. But you wouldn't ask a librarian to design your housing policy.
Though you might ask a librarian to tell you what everyone else has tried, so you don't waste time reinventing the wheel.
And that's the sweet spot. The AI eliminates the reinvention problem. It surfaces options that the researcher might not have known existed. Did you know that in Austria, there's a system where tenants can challenge rent levels after the fact and get refunds if the rent exceeds the legal limit? That's a completely different enforcement mechanism than pre-approval of rents. The AI can find that. The researcher might never think to look for it.
Then the researcher evaluates whether that mechanism fits the Israeli institutional context.
The AI expands the option space. The human prunes it. That division of labor is, I think, productive and much more realistic than the "AI think tank" vision where the whole pipeline is automated.
Daniel mentioned the international relations dimension — the fact that sometimes the best policy learning happens face-to-face, getting on a plane, seeing how things work. Is that romanticizing, or is there something irreplaceable about it?
It's not romanticizing. There's a concept in organizational learning called "stickiness" — the idea that certain kinds of knowledge don't transfer well through documents. They require personal interaction, observation, the ability to ask follow-up questions in real time. Policy details can be sticky in exactly this way. You read about the Swedish rent negotiation system and you think you understand it. Then you sit in a room with someone from the Swedish Tenants' Union and you realize you had completely misunderstood the power dynamics, the informal norms, the way the negotiation actually feels.
The vibe of the thing.
The vibe of the thing. And the vibe matters. It matters for implementation. If you import the formal structure of Swedish collective bargaining without the culture of cooperation that makes it work, you get something that looks like Swedish rental law but functions like a bargaining deadlock that leaves everyone worse off.
The face-to-face component is about acquiring the tacit knowledge that makes the explicit knowledge usable.
And that's the part that AI currently can't do at all. An AI agent can't have coffee with a Berlin tenant organizer. It can't tour a Singaporean HDB estate and notice the maintenance quality, the community atmosphere, the things that don't show up in the statistics. The AI operates entirely in the world of explicit, codified knowledge. Policy lives substantially in the world of tacit, experiential knowledge.
Which suggests the optimal model might be AI-augmented researchers who still travel, but travel with much better preparation. They've already mapped the formal landscape, identified the key questions, spotted the apparent contradictions between law and outcome. They arrive knowing exactly what they need to learn.
The travel becomes more efficient, not obsolete. You're not spending the first three days of your research trip figuring out the basics. You've already done that with AI assistance. You're spending your time on the things that only in-person observation can give you.
Let's pull this together. Daniel asked several things. What's this practice called, how's it done traditionally, what are the pitfalls, and could AI agents create a useful mechanism for doing it. You've given us the name — comparative public policy, policy benchmarking, the EU's Open Method of Coordination. The traditional methodology is typology-based country selection, deep institutional analysis, on-the-ground practitioner interviews, and systems-level thinking about knock-on effect. The pitfalls are institutional naivete, cherry-picked comparisons, legal text fetishism, and ignoring tacit knowledge. And the AI potential is real but bounded — it's a discovery and synthesis engine that expands the option space, but the judgment about transferability and the acquisition of tacit knowledge remain irreducibly human.
That's the summary. I'd add one thing about the agentic think tank idea specifically. I think the most interesting version isn't "AI replaces the Brookings Institution." It's "AI enables a distributed network of policy researchers who currently can't afford Brookings-level resources." The parliamentary researcher in a small country, the NGO policy analyst, the local government staffer — these are people who currently operate with severe information constraints. Give them good AI tools for cross-jurisdictional scanning, and you democratize a capability that's currently concentrated in a few well-funded institutions.
It's not about automating the elite think tank. It's about giving think-tank-grade research capabilities to people who will never set foot in one.
And that's where Daniel's idea of agentic think tanks gets exciting. Not as a replacement for human judgment, but as a force multiplier for human researchers who currently can't afford to do this work at all.
The counterargument, though — and I know you've thought about this — is that lowering the cost of comparative policy research also lowers the cost of motivated comparative policy research. If I'm a lobbyist and I can use AI to generate a plausible-looking cross-country comparison that just happens to support my client's preferred policy, I can flood the zone with pseudo-analysis.
That's a real concern. But I think it's less of a concern than it appears, because motivated policy research already exists in abundance. The barrier to producing a biased report was never cost — it was always just having a motivated funder. What AI changes is the cost of producing rigorous research. And rigor is actually harder to fake than bias. You can always find a think tank to produce a report that says what you want. What's scarce is the kind of systematic, methodologically transparent comparison that the OECD does. If AI makes that cheaper, I think the net effect is positive, even if it also makes biased research cheaper.
Because the biased research was already cheap.
The biased research was already cheap.
One more question and then we should wrap. Daniel mentioned secondhand smoking as another domain where this approach applies. Is there something about comparative policy that works better for some issues than others?
Comparative policy works best when the outcomes are measurable and the mechanisms are relatively well-understood. Public health is actually the ideal domain — you can compare smoking rates across countries, control for confounding variables, and isolate the effect of specific policies. The literature on tobacco control is one of the great success stories of comparative policy. We know, with high confidence, that comprehensive smoking bans, high taxes, plain packaging, and public education campaigns work. We know this because dozens of countries have implemented different combinations and we can see the results.
Whereas in housing policy, the causal chains are longer and messier.
A smoking ban affects smoking rates through a fairly direct mechanism. A change in eviction law affects the rental market through multiple channels — landlord behavior, tenant behavior, construction investment, mortgage lending, even family formation patterns. Disentangling cause and effect is enormously difficult. Which means comparative housing policy will always involve more judgment and less certainty than comparative public health policy.
Which is exactly the kind of domain where you want AI doing the information gathering and humans doing the sense-making.
The messier the domain, the more you need the human in the loop. The AI can tell you what happened. It takes a human to have a plausible theory about why.
Now: Hilbert's daily fun fact.
Now: Hilbert's daily fun fact.
Hilbert: In the nineteen twenties, a prominent Austrian geologist named Albrecht Penck proposed that all ice caves were formed by a single prehistoric glaciation event he called the Great Cave Ice Age. He believed cave ice was not replenished annually but was instead fossil ice from this single ancient period. The theory was mainstream for nearly two decades until researchers in the Dachstein caves demonstrated conclusively in nineteen thirty-seven that ice formations showed clear annual growth layers, proving they were dynamic systems that melted and reformed each year.
So close to a good band name, so far from being correct.
Annual growth layers in ice. I appreciate a theory that can be disproven by just looking at the thing.
So to close this out — I think the thread running through Daniel's prompt is about the relationship between information and judgment. We're in this moment where the cost of information is collapsing, and comparative policy is a domain where that collapse could change who gets to participate in serious policy debates. But the thing that isn't collapsing in cost is judgment. Knowing what's comparable, knowing what's transferable, knowing when the formal rules diverge from the real rules — that still takes a human who's done the work. The AI can make that human faster. It can make them better informed. It cannot make them wise.
I think the people who figure out how to combine AI-enabled breadth with human depth — those are the people who are going to produce the best policy analysis in the next decade. Not the people who automate everything, and not the people who refuse to use the tools. The sweet spot is augmented judgment.
Thanks to Hilbert Flumingtop for producing. This has been My Weird Prompts. If you enjoyed this episode, leave us a review wherever you get your podcasts — it helps other people find the show.
We'll be back next week.
