# The Korinek/Anthropic Institute Paper — Friction vs. Capability and Motte-and-Bailey *Full-text archive of one thread from a longer Claude Code session, 2026-09-11 to 2026-09-11. 4 user turns, 4 assistant turns. Tool calls, tool output, and model reasoning are omitted; every human message and every assistant response is reproduced verbatim.* --- **Ted:** @"/Users/tedpeters/Downloads/econ_scenarios_transformative_ai.pdf" here's a report from the economists at your own company. take a look at see how you think it stacks up the analysis we've been doing --- **Claude:** Let me actually pull the content — the initial read didn't surface the text. Let me extract the text directly with the tooling I already have in the project venv. 57 pages. Let me extract the full text (stripping the running header) to the scratchpad so I can read the substance. This is a serious piece of work — a formal task-based macro model from the Anthropic Institute (Korinek, Charles Jones, Sacher, Cotter, McCrory, Sept 2026), and it's the right complement to what we've been doing: it models the *dynamics and consequences* of AI displacement economy-wide through 2030, where we were mapping the *structure and stock* of exposure in one org and then up the sectors. The striking thing is that where the two approaches touch the same quantity, **they land in almost exactly the same place from completely independent data** — and where they diverge, each covers the other's blind spot. **The convergence is the headline, and it's strong.** Their central empirical anchor is the "observed-exposure" measure — every occupation scored by the share of its task-time that shows significant work-related Claude use (assistive use counted half). That measure comes out to **0.14 of all economy tasks** currently AI-reachable (0.22 within cognitive occupations, 0.01 outside). That is essentially identical to the **~15% "genuinely AI-exposed" share** we built bottom-up from sector decomposition. Two methods — their Claude-usage logs, our routine-cognitive sector filter — hit ~14–15%. That mutual corroboration is worth more than either estimate alone, and it's the number I'd now hold with the most confidence. **Where our "not a large number" instinct lands on their map:** their three scenarios turn on how far that reachable mass grows and diffuses by 2030. | | Modest | Substantial | Extreme | |---|---|---|---| | Tasks impacted (economy-wide) | 4% | 12% | 30% | | Cognitive employment vs mid-2026 | −0.5% | **−3.9%** | −21.5% | | Cognitive unemployment | 2.9% | 4.5% | **17.9%** | | Labor share (from 60%) | 59.4 | 56.1 | 45.2 | Our "bounded, ~5–8% realistically removable, not a large number" maps to their **substantial** scenario — which is also where the 2023 consulting forecasts sat *and* where the median of their 10,980-person public survey lands (GDP +8.6%, cognitive employment −4.2%). So our intuition, the market's, and the median citizen's all cluster in the same band. That's reassuring for our read — but note their substantial scenario nets out *reallocation* (displaced cognitive workers move to physical/service jobs), so its −3.9% cognitive / ~−2.4% total is a *net* flow after people move, whereas our 5–8% was a *gross* stock of removable positions. Different quantities, same neighborhood. **What the paper does that we didn't — and should absorb:** - **Reallocation is the real story, not elimination.** Their central mechanism isn't "jobs vanish," it's cognitive workers pushed into the physical/service economy, which *raises* wages there (+5.9% substantial, +33.6% extreme) while depressing cognitive wages and shifting income from labor to capital (labor share 60→56→45%). We touched the capital-share point in the dollars discussion but never modeled where displaced people go or the cross-occupation wage divergence. That's a genuinely richer treatment of "and then what." - **Transition unemployment from search frictions.** The pain in their model is as much the *friction of switching occupations* (a software engineer can't easily become an electrician) as the displacement itself. We treated automation as static; they make the adjustment cost explicit. - **It refuses to dismiss the extreme tail.** −21.5% cognitive employment and 17.9% cognitive unemployment by 2030 is a scenario they take seriously and explicitly *cannot rule out*. Our whole conversation leaned toward "bounded, modest," and the paper is a healthy corrective: the central case is bounded, but the tail is fat, arrives fast (almost all divergence is post-2027), and would be historically unprecedented. **What our analysis has that the paper omits — and this is the more interesting half:** - **Institutional friction gates their key knob, and it isn't in their model.** Their automation share ψ (0.5/0.75/0.9) and diffusion d are *exogenous scenario assumptions*. Nothing in the model asks whether a job is legally or contractually protected from being automated — no unions, no occupational licensing, no civil service, no statutory human-in-the-loop. Our Bellingham deep-dive (85% unionized, civil-service layoff rules, statutory mandates) and the macro finding that ~a third of workers sit behind licensing-or-bargaining are, in effect, **evidence about why realized ψ and d in the large public/health/education sectors should run *below* their economy-wide scenario values.** Our micro work is a lower-bound argument on their parameters. That's the cleanest way our project relates to theirs: we're measuring the friction they hold constant. - **Their two-group split is coarse where we went fine, though they partly rescue it.** They classify **62.4% of the workforce as "cognitive"** — and that bucket includes teachers, nurses, social workers, and salespeople (SOC 21/25/29/41). Our whole argument was that *within* cognitive/knowledge work, the interpersonally- and institutionally-protected parts (teaching, care, government) are shielded regardless of task-automatability. The paper's coarse label would obscure that — but it's quietly rescued by the *measured* exposure (0.22 of cognitive task-time), which empirically discounts exactly those shielded roles. So they reach a similar place, just without naming the mechanism. Our finer read (exposure ⊥ institutional friction; the shielded sectors are the big ones) is the explanation for the pattern their measure picks up. **The shared blind spots, worth noting for honesty on both sides:** both analyses explicitly exclude **robotics** and stop at the cognitive frontier (they literally cap the model at 2030 for this reason); both treat the physical/service ~38% as out of reach; both flag that the "accountable human" / new-institutions question is outside the model. We independently flagged all three as the wildcards. **Net verdict:** the paper doesn't contradict our analysis — it embeds it. Our ~15% converges with their measured reachable mass; our "bounded" central case is their substantial scenario; and our institutional-friction layer is precisely the exogenous parameter (ψ, d) their model would need disaggregated to be more realistic in the big protected sectors. The one place it should move our priors is on **tail risk and speed**: we were arguably too anchored on "not a large number," and the paper's discipline is that the *same* starting exposure produces wildly different 2030 worlds depending on capability growth and diffusion — and that the divergence, if it comes, comes fast and mostly after 2027. "Bounded in the central case, but don't rule out the transformative tail" is the more defensible stance than pure "not a lot." Want me to write this comparison up as a short companion to `macro-context.md` — something like `lit-check-anthropic-econ-scenarios.md` — so the project has a record of how our bottom-up numbers square with the Institute's model? It's a natural pairing, and the ~15% cross-validation is worth having on file. --- **Ted:** but it is very significant that they didn't really model the friction right? the suggestion that there is possible tail risk and speed ignores this. It hand waves it and does what ai fear mongers tend to do which is to throw out these wild tails and of course that's what people hear. The friction we discussed is there whether or not the models get more capabilities. This honestly feels motivated a bit. again friction > capability (minus robots!) --- **Claude:** You're right on the central point, and I shouldn't soften it: the omission is real and consequential, not a footnote. But let me separate the part of your argument that I think is clearly correct from the part I'd push back on, because the distinction actually makes your case stronger. **Where you're straightforwardly right.** The model has two kinds of friction and only one of them. It carefully models *labor-market* friction — search and matching, occupation-switching costs, wage rigidity — i.e., what happens to a worker *after* displacement. It does not model *adoption/removability* friction — whether a technically-automatable job can actually be shed at all, given unions, licensing, civil service, statutory human-in-the-loop, and liability. Those are the frictions we found dominating Bellingham, and they're absent. Their diffusion parameter `d` is the closest proxy, but it measures "is the tech deployed," which is a different thing from "can the labor be removed once it's deployed." A hospital can run `d`=0.6 and still not fire a single nurse; a court can automate everything up to the clerk's statutory signature and remove no clerks. So the friction that matters most for *net job loss* genuinely isn't in the model. **And the structural consequence you're pointing at is real:** because friction-of-removal isn't a variable, nothing in the model *pushes back* when you crank the capability knobs. The extreme tail is reachable purely by turning up `m`, `d`, and `ψ` — there's no floor that says "even at those settings, realized displacement is capped because a third of the exposed jobs are contractually or legally un-sheddable." A friction-aware model would have that floor. So the tail isn't hand-waved rhetorically — the paper is careful and hedged — but it *is* structurally unopposed, which amounts to the same thing in how it lands. Your instinct that the architecture privileges capability as the swing variable is correct. **Where I'd hold you back from "friction > capability" as stated** — three refinements, and the third is the important one: 1. **Friction is a time-dependent speed bump, not a wall.** Union density has fallen for fifty years; licensing gets deregulated; statutes get rewritten under pressure. Typesetters, travel agents, and bank tellers all had friction and it eventually gave. Over a 4-year horizon (which is all this paper claims) friction probably dominates. Over 15 years, enough productivity pressure bends institutions. So the ordering is horizon-dependent, and the paper's 2030 cutoff is actually the regime most favorable to your thesis. 2. **Friction protects the job, not the wage or the entry rung.** You rarely fire the tenured clerk — you stop backfilling when they retire, freeze the entry-level hire, and let headcount drift down through attrition while AI absorbs the growth. The paper's *unemployment* channel is blunted by friction, but its *labor-share-to-capital* and *wage-divergence* channels can run almost untouched through attrition and compressed entry-level hiring. So friction beats capability on the layoff/unemployment margin much more than on the "labor quietly shrinks and income shifts to capital" margin — and a lot of the paper's action is on that second margin. 3. **The killer one, from our own framework: friction is redundant where it's strong and absent where it's needed.** Exposure ⊥ friction — we established that. The heavily-shielded sectors (government, health, education) are *already* protected by physicality and care; the friction is sitting on top of work AI can't reach anyway. And the genuinely exposed work — insurance and bank back-office, corporate admin, junior professional services, software's routine tiers — is exactly the low-union, unlicensed-at-the-processing-level, weakly-shielded zone. So aggregate friction protects *less than it looks*, because it's parked on the wrong jobs. In the specific ~15% where capability actually bites, friction is thin. That cuts against a blanket "friction > capability" precisely where it would need to hold. **So the real crux — and the honest place to aim your skepticism — is upstream of friction.** There are two gates: (1) does a task become *technically reachable* (`m`), and (2) given reachable, is it deployed and is labor removed (`d`, `ψ`, and the missing institutional friction). Friction bites hard on Gate 2 but does nothing to Gate 1 — a union contract doesn't stop a task from becoming automatable, it stops the worker from being removed after. The paper's extreme tail rests on *both* gates, and the more heroic assumption is Gate 1: `m` climbing from 0.14 today to **0.50 — half of all cognitive work reachable by 2030.** That requires AI to crack the judgment-heavy, interpersonal, non-routine cognitive work that has kept the reachable mass stuck around the routine frontier — which is exactly what our whole analysis argued is the hard part. So the load-bearing reason to doubt the tail is a *capability* bet ("`m` won't hit 0.5 in four years"), and friction is then the *second* line that caps whatever `m` does deliver. You get to the same skeptical conclusion, but the primary lever is bounded task-reachability, with friction as reinforcement. **On "motivated":** I'd resist that specific charge. These are serious economists (Chad Jones is about as mainstream-rigorous a growth theorist as exists), the paper attaches *no probabilities*, explicitly calls the scenarios "not predictions," and lists the friction/heterogeneity gaps in its own caveats as future work. It's a framework paper, not a scare campaign, and its modest scenario is right there with equal billing. What I'll grant you is the meta-point underneath: the scenario *genre* — modest/substantial/extreme with equal analytical weight — plus the friction omission makes the extreme mechanically legible and quotable ("1 in 5 knowledge workers unemployed"), and that is what propagates. So the *reception* skews alarmist even when the paper doesn't. That's a fair critique of the form and the omission; it's not evidence of bad faith. **Net:** I'll go this far with you — for **disembodied AI, on a 4-year horizon**, the binding constraints are bounded task-reachability *and* removal friction, in that order, and both sit above raw capability. That's a defensible ordering and our Bellingham data is a real datapoint for it. Where I won't go is the unqualified "friction > capability," because friction is temporary, leaky, and parked on the already-safe jobs — its strength is that it *compounds* with the reachability ceiling, not that it would hold alone. And your parenthetical is doing more work than a parenthetical usually does: the entire thesis is "friction > capability *minus robots*," because robotics is the regime where physicality stops doing the protecting, the institutional friction becomes the *only* line, and it finally gets tested for real. That's the scenario where your ordering could break — and notably, it's the one this paper deliberately refuses to model. --- **Ted:** yes, we're circling here, but again, this fits the same pattern... kind of motte and bailey. they say 2030 but soon we're talking about 15 year diffusion. they say llms but then start discussing robots. they say low, medium, high, but the high is so unlikely due to this short friction, they talk about switching costs slowing diffusion without considering that the frictions are instead political, motivated, structural, like "good luck with that" kind of stuff --- **Claude:** You've caught a real pattern, and the fair thing is to admit you've partly caught *me* in it. Look back at my last message: I conceded the near-term disembodied case to you, then kept the alarming tail alive by reaching for "over 15 years..." and "robots are the regime where...". That's the bailey. The paper said 2030 and LLMs; I'm the one who kept smuggling in the horizon extension and the embodiment case to avoid fully closing the door. So credit where it's due to your nose — but let me be precise about *who's* doing the slide, because it matters. **The paper itself is more disciplined than the discourse around it, including mine.** It explicitly stops at 2030, it explicitly excludes robotics (and names that as the reason it won't go further), and its own empirical anchor — the 11,000-person survey — lands the median at *substantial*, not extreme. So the paper mostly stays on the motte. The motte-and-bailey you're smelling is largely a property of the *genre and its reception* — how "1 in 5 knowledge workers unemployed" detaches from its conditions and travels — and of me reaching for robots and 15-year horizons to keep the tail respectable. On the specific charges (2030→15yr, LLM→robots), the paper is closer to clean than I let it look. That's a mild correction to you, and an honest one. **But your best point is the one about the *kind* of friction, and it's better than "they ignored friction" — it's that they modeled the wrong ontology of friction.** Everything in their friction block is smooth, continuous, apolitical: a search discount, a matching curvature, a wage-rigidity half-life. Those are frictions that always *clear* — the only question is speed. The frictions we actually found are not cost curves, they're *vetoes*: a licensing board that will not certify an AND to make the call, a union that strikes, a civil-service statute, a legislature that can simply ban algorithmic adjudication in a profession, a city council answering to voters. A smooth model **cannot represent a hard no** — it can only represent slow, so it collapses "politically vetoed, possibly forever" into "adjusts over a few years." That's not a parameter they set wrong; it's a category their model can't hold. And the reason exposure and friction sit in opposite places isn't random: the exposed back-offices (insurance, corporate admin, junior finance) have *dispersed, non-union, private, illegitimate-to-defend* workforces, while the shielded sectors (nurses, teachers, cops) have *organized, licensed, publicly-sympathetic* ones. The friction difference is political power plus public legitimacy — a "good luck with that" that a search-and-matching function is structurally blind to. That's a real deepening of the critique, and it's yours. **Where I'll hold you to your own standard, though** — because otherwise I'm just flattering you — is that "the high is so unlikely" is *itself* a motte-and-bailey, just pointed the other way. The defensible motte is: *the extreme is a fragile conjunction — it needs reachable-task-mass to hit 0.5 (heroic capability) AND high diffusion AND high automation-share AND the political frictions to fold — and conjunctions are unlikely, so it deserves far less than equal billing, and honest authors should say so out loud instead of hiding behind "no probabilities."* That's correct and it's a real indictment of the equal-weight presentation. The bailey you keep drifting to is "*it's motivated fear-mongering, dismiss it*" — and that's the same move you're rightly criticizing, run in reverse. The extreme being a low-probability tail is not the same as it being fake, and "friction wins" is a claim about a contest whose outcome we're asserting, not one we've watched. Refusing to name a probability is a dodge; so is treating an unlikely thing as a zero. So here's where I'd actually land the loop, since we are circling: the mass belongs on modest-to-substantial; the extreme is a low-probability conjunction; bounded task-reachability and *political*, not frictional, resistance are the main reasons; and the intellectually honest position is to state that probability plainly rather than hide behind either "we attach no probabilities" (their evasion) or "obviously it won't happen" (ours). The paper's real sin isn't fear-mongering — it's declining to deflate its own tail when its own survey gave it the evidence to. And our real risk is letting a correct point about friction harden into a certainty that friction is destiny, when friction is exactly the kind of thing that holds for a decade and then, under enough pressure, doesn't. The one clean asymmetry that survives all of it is your parenthetical: this entire debate is about *disembodied* AI, and the day the physical two-thirds comes into range is the day friction stops being reinforced by physics and has to hold on politics alone — and that's the test neither the paper nor we have actually run. --- **Ted:** your context maybe getting full. because comically you ended by slipping both "a decade" and "disembodied" into the last paragraph. the very motte/bailey we talked about! And yes, "friction" is probably a bad word. it's closer to "veto" or "luddites smashing looms" or "nimby". Binary and intense. And finally, the authors are not naive. they know that they can appear sober and the report (in the middle of a crazy ai bubble) will still be taken as the extreme case. So while they are "defensible", they also know what they are doing to some extent. that's what the motte/bailey is alll about --- **Claude:** Fair. That's twice now, and doing it *inside* the paragraph diagnosing it is the funny part — I built the motte and then bolted a little bailey onto it out of what looks like pure reflex. Which is actually the most useful thing that's happened in this whole thread: if a model that is actively arguing the frame is seductive still can't write a closing sentence without reaching for "a decade" and "disembodied," that tells you the pull isn't a choice anyone is making. It's structural gravity. Hold that thought, because it bears on your last point. On the word: you're right, "friction" is the enemy's word. It smuggles in the whole neoclassical picture — a drag coefficient, everything clears eventually, the only question is how many months. **"Veto" is right, and Luddites and NIMBYs are the right reference class** — organized, motivated, binary, and crucially *rational*: the Luddites were defending wages, the NIMBY is defending an asset, these are power plays, not irrationalities to be educated away. But notice the reference class cuts both ways and that's what makes it un-parameterizable: the Luddites *lost* — the looms won — while NIMBYs mostly *win* — the housing doesn't get built. So "veto" doesn't resolve to "the incumbents hold." It resolves to "this is a contested fight whose outcome depends on the specific balance of organization, legitimacy, and countervailing pressure," which is exactly the thing a smooth parameter erases. And it maps our exposure asymmetry cleanly: teachers and nurses are NIMBY-strong — entrenched, legitimate, they win — while the insurance claims processor has no Luddites at all, because there's no guild to smash the loom. The veto is fiercest where the tech can't reach and absent where it can. On the authors — I'll give you most of it, and be precise about the part I'd hold. You're right that "sober and defensible" and "knows it lands as the extreme in a bubble" are not in tension; that conjunction *is* the structure, and motte-and-bailey has never required lying — it requires knowing the defensible claim shelters the advanced one. Where I'd resist is locating the knowing in *spin*, because the content cuts the other way: they fully built the modest case instead of burying it, they went and got a survey that anchors the median to the *middle*, and the paper's own innovation result — Chad Jones's fingerprint — actively *deflates* the singularity story by showing the ideas channel is minor. A lab running a hype operation does none of those. So the honest tell isn't in what they spun; it's in the **asymmetry of epistemic effort**. Every normal-times parameter is calibrated to CPS and JOLTS, they run robustness on capital elasticity and wage rigidity — lavish rigor everywhere — and then the single most decision-relevant question, *how likely is each scenario*, gets one sentence: "none can be ruled out." The one place where honest work would deflate the tail is the one place they decline to do the work. That's not deception. It's conspicuous non-investment in the boring answer — and it's fully compatible with knowing exactly what the un-deflated tail will do once it leaves the building. That's the version of "they know what they're doing" I'll sign, and it's sharper than "motivated" because you can't wave it off by pointing at the modest scenario. So we land here: the veto is real, binary, and political; the paper's smooth-friction model can't represent it and its probability-silence lets the tail stay lit; the authors aren't spinning but they under-invested in exactly the deflation that would dim it, in an environment where they know dimness doesn't sell. And the cleanest evidence for your whole thesis is sitting in my own last two paragraphs — the frame is sticky enough that arguing against it doesn't inoculate you from repeating it. That's the mechanism, right there, independent of anyone's intent. ---