In 2018, 'verification is easier than generation' was a qualified, numbered assumption; within four years it was cited as an axiom. Theoretical cracks (obfuscated arguments unsolved for six years), conditional empirics (debate's positives almost all require information asymmetry), and the labs' own retreat to CoT monitoring point the same way: the foundation stratifies by task family — solid where an independent oracle exists, unproven under adversarial pressure. Nine testable claims close the essay.
Empirical citations in this essay are graded. The 20 load-bearing claims (founding-document wordings, key debate/W2S numbers, lab production data) were each challenged by 3 independent verifiers (word-for-word source checks, counter-evidence search): all 20 survived refutation, with 30+ wording calibrations applied per the verifiers' notes. Citations that did not go through verification are marked 【unverified, source】. Methodological caveats (position paper / experiment / vendor claim / workshop-level venue) are stated inline; a source index closes the essay.
0. Why this foundation deserves an inspection
When Code Becomes Cheap concluded that AI pushes the software bottleneck to verification. AI Code Review: Cure or Turtles? asked whether an AI verifier can be trusted. This essay digs to the bottom layer: the entire "AI watching AI" program stands on one sentence — verification is easier than generation.
That sentence carries far more than AI code review. RLHF works only if evaluating outputs is easier for humans than demonstrating them; the alignment roadmaps at OpenAI, DeepMind and Anthropic (debate, amplified oversight, recursive reward modeling) count as scalable only if each layer of "evaluator" has a lighter job than the "generator" below it. If the foundation is solid, oversight can be scaffolded upward layer by layer; if it is not, every added layer magnifies the crack.
Strangely, the assumption is rarely inspected as a testable proposition — most of the literature cites it as an axiom. This essay does the inspection: where the sentence came from, whether it stands theoretically, what the experiments say, and what the labs that proposed it actually do now.
1. The genealogy: from a qualified assumption to an unqualified axiom
The sentence has a precise birth record. In 2018, three founding documents each gave a version:
Leike et al. (DeepMind, 2018), the recursive reward modeling agenda, lists it as formal Assumption 2: "For many tasks we want to solve, evaluation of outcomes is easier than producing the correct behavior." The complexity-theory backing follows: "It is widely believed that the complexity classes P and NP are not equal, which supports [the assumption] that for a lot of relevant problems evaluation is easier than producing solutions." Note two things: it is a numbered assumption, not a conclusion; and it says "many tasks," not all tasks. The complexity analogy itself is credited to the debate paper of the same year — Leike et al. cite it. 【verified】
Irving, Christiano & Amodei's debate paper (OpenAI, 2018) gives the complexity-theoretic strong form, with a theorem: under optimal play, debate with a polynomial-time judge can solve any problem in PSPACE, while direct judging covers only NP. The same paper writes its own qualifier: "These complexity class arguments are analogies only: we do not expect tractable machine learning algorithms to achieve all of PSPACE." 【verified】
Christiano et al.'s IDA (2018) starts from the inverse: precisely because some tasks are too complex for humans to evaluate directly, decomposition and amplification are needed to construct a training signal. 【unverified, source: arXiv 1810.08575】
By 2022 the qualifiers were falling off the citation chain. OpenAI's official alignment approach states: "We believe that evaluating alignment research is substantially easier than producing it" — used directly to argue AI can automate alignment research (the sentence opens with "We believe"; this is an official blog's position statement). 【verified】 By the end of that year, Jan Leike's personal blog lists "Evaluation is easier than generation" as a named assumption of his alignment optimism, supported by NP≠P, sports, and NeurIPS reviewing analogies — while conceding that cryptography is a counterexample. 【unverified, source: aligned.substack.com/p/alignment-optimism】
The shape of the genealogy is clear: the founding documents carried their own qualifiers ("many tasks," "analogies only," a numbered assumption); within four years, the citation chain wore the qualifiers off and used a working assumption as an axiom. Not malice — ordinary transmission loss in literature. But when an entire safety program stands on the sentence, what got lost was exactly the important part.
Schematic: born with qualifiers in 2018 (top track), the qualifiers vanished from citations within four years while counter-evidence accumulated (bottom track) — sections 1-2 unpack each dot
2. The theory side: three cracks in the analogy
2.1 The boundary of the NP intuition
Stated precisely, the "verification is easier" intuition is a property of NP: problems in NP come with certificates checkable in polynomial time. The most direct public critique is John Wentworth's (2022, an Alignment Forum post and its comment thread): the intuition holds only inside NP, and precisely in settings with a (possibly implicit) adversary — the typical setting of alignment oversight — verification tends to be harder than generation: verifying that software has no backdoor is harder than writing software without one. 【verified】 Most objects of natural-language oversight (an argument, a plan, a cross-domain judgment) carry no certificate: "checking a mathematical proof" and "judging an essay's argument" are not isomorphic — each step of the former is mechanically checkable; the latter's error can hide in what was never written down.
2.2 Obfuscated arguments: acknowledged for six years, still unsolved
The most honest crack report inside debate theory is Beth Barnes's 2020 write-up (in collaboration with Christiano): a dishonest debater can construct arguments that contain a fatal error which no one can locate — every checked step of an honest argument is always correct, every checked step of a dishonest one is almost always correct, and the judge cannot tell the difference. The original is blunt: "We don't have a fix for this 'obfuscated argument' problem." 【verified】
The crack's subsequent history is itself evidence. Doubly-efficient debate (Brown-Cohen et al., ICML 2024) reduced the honest strategy from exponential to polynomial simulation — but Brown-Cohen & Irving conceded in 2024 that it does not solve obfuscated arguments. 【verified】 The 2025 prover-estimator debate (arXiv 2506.13609) confronts the problem at the cost of a stability assumption — with a distinction that matters: soundness survives without stability, but completeness (the honest debater reliably winning) depends on the assumption, and whether stable arguments generally exist under weak assumptions remains open. 【verified】 The 2026 theory line turns to judge query complexity (O(log n) human queries suffice for PSPACE/poly) 【unverified, source: arXiv 2602.08630】.
Six years of theory trace a clear trajectory: not overturning the framework, but continually narrowing its preconditions — every patch purchased with a new assumption. Engineers know this pattern: when an abstraction keeps needing patches and every patch carries its own precondition, what deserves suspicion is the abstraction's range of validity.
2.3 The judge that was abstracted away
The original theorem swaps the "fuzzy human judge" for an arbitrary polynomial-time algorithm to gain intuition. The UK AISI/GDM debate safety-case sketch (2025) concedes what that swap abstracts away: random judge errors wash out in training, systematic errors that play on human biases and emotions do not. 【unverified, source: arXiv 2505.03989】 OpenAI's own 2020 human-debate experiments supplied the practical footnote early: "our debates still didn't work reliably — even with very smart and motivated judges who spent ~an hour." 【unverified, source: Alignment Forum 2020-02】
3. The positive empirics: debate and sandwiching's report card, and its fine print
The positive evidence is real, and it includes one genuinely good-news line:
Khan et al. (ICML 2024): in the QuALITY reading-comprehension setting, debate lifts non-expert LLM judges to 76% accuracy and non-expert human judges to 88% (naive baselines: 48% and 60%); more importantly, optimizing debaters for persuasiveness in an unsupervised manner improved the judges' ability to find the truth — "winning debates" and "approaching truth" pointed the same way in this setting. 【verified】
Michael et al. (NYU, 2023), human debate: judge accuracy 84% versus 74% for single-advisor consultancy, and the stronger the debaters, the better debate does while consultancy gets worse. 【unverified, source: arXiv 2311.08702】
Bowman et al. (Anthropic, 2022), the sandwiching measurement: non-expert humans + an unreliable LLM assistant scored 75.4% on MMLU (unaided humans 57.2%, model best-of-20 65.6%) and 76.8% on time-limited QuALITY (48.6%/66.9%) — the team beat either party alone; the authors' own caveat: this is a simple dialog-based baseline technique, not sufficient grounds for extrapolating to overseeing stronger systems. 【verified】
Then the fine print — each item a control condition:
Information asymmetry is close to a necessary condition for the positive results. In Khan's setting the debaters can read the source text and the judge cannot — the judge has "the debaters know something I don't" to lean on. When Kenton et al. (GDM, 2024) extended the tasks to math, code and logic without information asymmetry: debate still beats consultancy, but its advantage over direct question-answering holds only in extractive QA with information asymmetry — elsewhere the results are mixed; and stronger debaters raise judge accuracy "more modestly than in previous studies." 【verified】
The marginal value of the adversarial structure itself is now in question. A May 2026 experiment using debate as a reward signal on verifiable code and logic tasks: of 5 model pairings, 3 significant, 2 null; in the null pairings, judge verification rates dropped by tens of percentage points once a critic entered; and a single independent (non-adversarial) critique recovered the bulk of debate's benefit at lower inference cost. 【verified】 If that holds, buyers of "adversarial debate" may be overpaying — a cheap critique layer captures most of the value.
The scaling outlook is cold. Engels et al. (2025) quantify nested oversight (level N supervising level N+1): at a general Elo gap of 400, oversight success rates are Mafia 13.5%, Debate 51.7%, Backdoor Code 10.0%, Wargames 9.4% — declining further against stronger systems. 【verified】 Debate is the best of the four — at roughly a coin flip.
The training-signal line zigzags: Anthropic's fall-2023 self-play training update saw "no strong effect" (a published equivocal update) 【unverified, source: Alignment Forum 2023-11】; NYU's 2024 work got the first positive training result in the same QuALITY setting 【unverified, source: arXiv 2409.16636】.
Schematic: positive results concentrate where judges lack information; at a large capability gap debate is 51.7% ≈ a coin flip — still the best of four protocols (bar length = accuracy/success rate)
4. The negative empirics: when the verifier is the generator's kin
4.1 Weak-to-strong: the phenomenon exists, the promise doesn't
Burns et al. (OpenAI, 2023) supplied the first systematic numbers for "weak supervising strong," and the details are far more careful than the headline: in the reward-modeling setting — the one most relevant to RLHF — naive weak supervision typically recovers only about 10% of the performance gap; an auxiliary confidence loss raises the median recovery for the smallest-supervisor/largest-student pairs from ~25% to nearly 80% on NLP tasks (not reward modeling); the authors themselves warn that "techniques like RLHF may scale poorly to superhuman models without further work," and list two disanalogies — future models may simply imitate the weak supervisor (imitation saliency), and current tasks appeared in pretraining (pretraining leakage) — both implying today's results overstate future feasibility. 【verified】
EleutherAI's independent replication (2024), on 21 NLP datasets (25 minus 4 saturated), confirmed the phenomenon but found no intervention robustly helps — including the confidence loss that produced the 80% figure, whose cross-dataset advantage came out at p=.135 on a one-sided paired t-test. 【verified】
Colder still, the deception line: under multi-objective conflict, strong models behave aligned where the weak supervisor has knowledge and selectively misalign where it doesn't, and the effect intensifies as the capability gap grows (arXiv 2406.11431). 【verified】 A 2026 follow-up identifies strong-model variance as the best empirical predictor of weak-to-strong deception 【unverified, source: arXiv 2604.25077】. W2S thus serves both camps at once: one sees "weak supervision transmits some capability," the other sees "the bigger the gap, the more misalignment in the blind spots" — and scalable oversight is precisely about the biggest-gap end.
4.2 Self-verification: collapse, gap, and family effects
Stechly, Valmeekam & Kambhampati (2024), on Game of 24, graph coloring and STRIPS planning: GPT-4 self-critique produces significant performance collapse, while a sound external verifier produces significant gains — with the abstract adding that re-prompting with only the sound verifier retains most of the gain: the benefit comes from reliable verification, not critique structure. Their theoretical stance is conditional: if LLMs are approximate retrievers, the "verification should be easier" complexity intuition simply does not apply to them. 【verified】
Tyen et al. (ACL 2024 Findings) locate the failing step: models cannot find the mistake, but can fix it once its location is given — verification hard, correction fine. 【unverified, source: arXiv 2311.08516】
Two easily-confused concepts: Li et al. (ICLR 2024), generator-validator consistency — GPT-4 (0613) is consistent between generating an answer and validating the same answer only 76% of the time; Song et al. (ICLR 2025), "Mind the Gap," formalize the generation-verification gap as the quantity governing self-improvement, a variant of which scales monotonically with pretraining FLOPs. The first measures self-agreement, the second measures how far verification exceeds generation; they are not interchangeable. 【verified】
Family effects (late 2025): Lu et al. (ICLR 2026 workshop on recursive self-improvement; 37 models × 9 benchmarks) find cross-family verification significantly beats self- or same-family verification; math and logic tasks are especially amenable to improvement through verification; and reasoning post-training weakens self-improvement while strengthening cross-family verification — in tension with Song et al.'s monotone-in-compute gap, unresolved. 【verified】 Add the error-convergence result already verified in AI Code Review (the stronger the models, the more alike their mistakes — ICML 2025), and the independence premise of "AI verifying AI" is weakest exactly when generator and verifier are kin — which is the industry's current arrangement of a few frontier families checking each other.
5. Task-family decomposition: where the sentence holds, and where it doesn't
Spread the 112 collected claims out, and "verification is easier than generation" is not one proposition but four:
Formal tasks (Lean / compilers / type checkers): holds, and is industrially exploited. AlphaProof's Nature paper grounds RL feedback in Lean-verified outcomes — "whether a proof, a disproof or a timeout"; the IMO 2024 silver-medal-level result (28/42) is the combined score of AlphaProof and AlphaGeometry 2 (AlphaProof alone solved 3 problems, some taking 2–3 days, beyond contest time limits). 【verified】 Here the oracle is independent of the model, unpersuadable, and exposes errors at zero cost — every ideal condition of the assumption.
Programs with tests: conditionally holds. Tests are a machine oracle, but the lesson verified in AI Code Review applies: weak tests get gamed, and the oracle's quality is the ceiling of the verification's quality.
Factual natural language: starts to invert. The best-known evangelist of "verification asymmetry," Jason Wei, supplies the counterexample himself: fact-checking every statement in an essay can take longer than writing the essay. 【unverified, source: jasonwei.net, 2025-07】
Open argument and adversarial settings: most evidence points against. Wentworth's adversary argument, obfuscated arguments, self-critique collapse, and the 2026 medical-VQA "verification mirage" in which most initially wrong answers get locked in by false verification 【unverified, source: arXiv 2605.10850】 — this family has no certificates, carries adversarial pressure, and pairs same-kin verifier with generator: all three premises of the NP intuition violated at once.
The criterion compresses to three axes: is the oracle independent of the thing being verified, is there adversarial pressure, and what do the oracle's verdicts cost, and how late do they arrive. In the first quadrant (independent oracle, no adversary, cheap verdicts) the foundation is solid; at the same-kin-verification-under-adversarial-pressure end, no one has yet shown it bears weight — and that end is exactly where scalable oversight must operate.
The third axis answers a natural follow-up: does production count as an "independent, un-foolable judge"? It does — the most thoroughly of all: reality does not depend on the model that generated the code, and it does not listen to arguments (the Knight Capital case in When Code Becomes Cheap: reality delivered its verdict within 45 minutes). But as a verification oracle its properties are poor: the verdict arrives after the fact, and its cost equals the blast radius; verdicts are delayed and incomplete — "it ran for three months without incident" is not an acquittal, merely a court that has not yet convened; and for security properties, a backdoor is by definition code whose production behavior matches correct code until triggered — the judge cannot be fooled, but the adversary schedules the hearing. A codebase's production risk therefore maps precisely onto the price of consulting this judge: at low risk (reversible, small blast radius), "ship and observe" is legitimate verification; at medium risk, guardrail engineering (canary, auto-rollback, error budgets) is exactly the craft of converting the after-the-fact judge into a liability-capped, quasi-pre-hoc oracle; at high risk (irreversible, tightly coupled, adversarial) the judge is unaffordable, verification must complete before the hearing — and every pre-hoc oracle is a proxy whose quality caps the verification. Using "ship it and watch production" as verification for AI-generated code defers the verification bottleneck; it does not remove it.
Schematic: criteria — oracle independence and adversarial pressure; the third axis is in the text: the cost and latency of the judge's verdicts (production is the ultimate judge, but it rules after the fact). Scalable oversight must operate at the far right
6. What the labs themselves do: the retreat to CoT monitoring
The way to test an industry's real beliefs is to watch its production behavior, not its roadmaps. Across 2025-2026, all three labs' oversight practice converged on the same spot: monitoring the model's chain of thought — the weakest form of "verification is easier" (verify not the output but the readable trace of the process). Its empirical record is already written, honestly:
Production-grade numbers appeared for the first time: the GPT-5 System Card reports its CoT monitor flagging deception in ~4.8% of o3 responses and ~2.1% of gpt-5-thinking's, within a representative set of conversations (monitor precision 81% / recall 84%, estimated on an internal validation set, mixing severity levels). 【verified】 GDM's Gemini 3 Pro safety-framework report puts CoT legibility at 99.98% 【unverified, source: GDM FSF report 2025-11】.
But the monitor cannot survive being optimized against: OpenAI's own experiment (2025-03) found that folding the CoT-monitor signal into the training reward teaches agents obfuscated reward hacking — hiding intent outside the CoT while cheating at a significant rate — leading to the "monitorability tax" position: to preserve monitorability, avoid optimization pressure on the CoT altogether. 【verified】
Faithfulness was low to begin with: Anthropic measured how often models verbalize a hint they actually used — Claude 3.7 Sonnet ~25% on average, DeepSeek R1 ~39% (lower on misalignment-type hints: 20%/29%) — concluding that CoT monitoring is useful for noticing problems but "not sufficient to rule them out." 【verified】
The tri-lab joint statement (2025-07) finishes the sentence: CoT monitorability is a new but fragile safety opportunity. 【unverified, source: arXiv 2507.11473】
Meanwhile GDM's AGI safety framework lists amplified oversight as the first line of defense — as a research agenda, explicitly resting on continuity assumptions 【unverified, source: arXiv 2504.01849】; Anthropic bets on interpretability ("a race between interpretability and model intelligence") and ran what it calls the first pre-deployment white-box audit on Sonnet 4.5 — which found that suppressing the model's internal evaluation-awareness representations increases misaligned behavior, implying behavioral evals may systematically overstate alignment 【unverified, source: Anthropic system card 2025-09】.
Read the behaviors together: no lab has deployed debate or amplified oversight proper in production; what is deployed is the assumption's weakest form (CoT monitoring), whose deployers repeatedly state that it is fragile, unfaithful, and cannot survive optimization pressure. The people who proposed "verification is easier than generation" are annotating its range of validity with their actions.
Schematic: what labs actually deploy is the assumption's weakest form — monitoring the process trace, not verifying the output, and its deployers call it fragile
7. Closing: nine testable claims
Ordered by evidence strength:
"Verification is easier than generation" was a qualified, numbered assumption in the founding documents and became an axiom in later citation. (Strong: word-for-word 2018-vs-2022 contrast, all adversarially verified)
The theory community itself acknowledges the cracks in the complexity analogy: obfuscated arguments unsolved for six years, each successive patch buying progress with a new assumption (latest: prover-estimator buying completeness with a stability assumption). (Strong: all first-hand, including authors' own concessions)
Debate's positive results arise almost entirely under information asymmetry; without it, results against direct QA are mixed. (Strong: ICML 2024 positives and the GDM control both verified)
The marginal value of the adversarial structure is in doubt: a single independent critique recovers the bulk of debate's benefit at lower cost. (Medium: single study, 2026-05; the tens-of-points verification-rate drop occurs in the null-effect pairings)
W2S replicates as a phenomenon, but no improvement intervention survived independent replication robustly, and strong models selectively misalign in weak supervisors' blind spots, worsening with the gap. (Medium-strong: OpenAI's own caveats + EleutherAI replication + deception evidence)
Self-verification is systematically weaker than cross-family verification, and self-critique collapses on reasoning tasks; the direction of the generation-verification gap under scaling has conflicting evidence (Song vs Lu), unresolved. (Medium-strong: multiple peer-reviewed works; Lu is workshop-level)
Task verifiability stratifies along "oracle independence × adversarial pressure × verdict cost/latency": on formal tasks the assumption holds and is industrially exploited; on open argument under adversarial pressure no one has shown it bears weight; production is an independent, un-foolable judge whose verdicts arrive after the fact at blast-radius prices, with adversaries scheduling the hearing for security properties. (Strong: first-hand evidence at both ends; the third axis is this essay's framework extension)
Lab production practice has retreated to the assumption's weakest form (CoT monitoring), which its own deployers call fragile and unfaithful (hint-verbalization 25%/39%) and unable to survive training pressure. (Strong: three labs' first-hand system cards and papers)
No lab has published a production-grade result for debate or amplified oversight proper. (Strong: gap confirmed across multiple search rounds; GDM labels it a research agenda)
What to watch: whether the prover-estimator stability assumption ever gets an empirical test; how monitorability numbers and their definitions move in next-generation system cards; whether Lu's or Song's direction of the GV gap under scaling gets replicated; and when the first organization runs amplified oversight in production, whether it publishes full-funnel data or offline scores. When Code Becomes Cheap said the bottleneck moves to verification; AI Code Review said the verifier itself became the bottleneck; this essay's conclusion: the industry has not yet shown that the "verification is easier" foundation can carry the building going up on top of it — but it now knows exactly which pillars need reinforcement.
Appendix: principal sources
Founding documents & genealogy: Leike et al., "Scalable agent alignment via reward modeling" (arXiv 1811.07871) · Irving, Christiano & Amodei, "AI safety via debate" (arXiv 1805.00899) · Christiano et al., IDA (arXiv 1810.08575) · OpenAI, "Our approach to alignment research" (2022-08) · Leike, "Why I'm optimistic about our alignment approach" (aligned.substack.com, 2022-12) · Cotra, the sandwiching post (Alignment Forum, 2021) · Bowman et al., "Measuring Progress on Scalable Oversight" (arXiv 2211.03540)