The Verification Asymmetry

What the state of quantum computing in 2026 implies for AI governance — and where it breaks frameworks built on contesting a system's outputs

The interesting thing about quantum computing, for anyone working on AI safety or governance, is not that it might be fast. It is that the particular kind of speed quantum machines are first demonstrating arrives bundled with a property oversight has quietly relied on never losing: the ability of an outside party to check the work.

Almost every mechanism we use to hold automated systems accountable — audit, red-teaming, reproduction, the right to contest a consequential decision — assumes that a result, once produced, can in principle be re-derived or challenged by someone other than the system that produced it. Quantum advantage is, for an important class of computations, precisely the regime where that assumption fails. A computation that is classically infeasible to perform is frequently also classically infeasible to verify. This is not a transient engineering gap that better tooling will close. For the sampling-based demonstrations that define the field today, it is close to the point of the exercise.

What follows states soberly where the hardware actually is in 2026; makes precise what "advantage" means and why it tends to imply unverifiability — including the one important exception the governance conversation has to handle honestly; and traces the consequence for any framework, including the relational and calibration-based ones, that locates trust in a human's capacity to contest what a machine claims.

1. Where the hardware actually is

The real milestone of the past two years was not a qubit count. It was error correction crossing below threshold. Google's Willow result, published in Nature in December 2024, showed that enlarging a logical qubit makes it more reliable rather than less: scaling the surface-code distance from 3 to 5 to 7 cut the logical error rate by roughly half at each step, with a measured suppression factor of Λ ≈ 2.14 per distance increment, and the system operating beyond the break-even point with real-time decoding. This is the first convincing physical demonstration that quantum error correction behaves as theory has promised since Peter Shor introduced the idea in 1995. It matters far more than the accompanying random-circuit-sampling benchmark, which Google described as completing in under five minutes a task that would take a leading classical supercomputer on the order of 10²⁵ years — a striking figure attached to a computation with no practical use.

Sobriety is warranted on two fronts. First, Willow's logical error rate of roughly 0.14% per cycle remains orders of magnitude above what useful algorithms require; Google itself puts the target near one error in 10⁶ to 10⁹. Second, the overhead between physical and logical qubits is severe — on the order of a thousand to two thousand physical qubits per logical qubit for a surface code at current physical error rates, which is why IBM has shifted toward quantum low-density-parity-check codes that encode twelve logical qubits in 144 physical ones, and why IBM's own quantum lead has called the brute-force version an engineering pipe dream.

The roadmaps converge on the end of the decade. IBM commits to a fault-tolerant machine — Starling, targeting 200 logical qubits and 100 million gates — by 2029, with interim processors named through 2027 and a community "advantage tracker" aiming at a narrower claim by 2026. Quantinuum demonstrated a fully fault-tolerant universal gate set with repeatable error correction in mid-2025 and targets a comparable machine by roughly 2029–2030. IonQ and several neutral-atom and photonic ventures publish more aggressive numbers — millions of physical qubits within a few years — but these are vendor projections, not demonstrations, and should be read as such. Microsoft's February 2025 topological-qubit claim is the most contested of the major bets: the accompanying paper carried an editorial note that its results do not constitute evidence for the Majorana zero modes the architecture depends on, and independent physicists have been openly skeptical.

The honest summary: error correction works; useful fault tolerance does not exist yet; no quantum machine in 2026 runs a commercially meaningful workload faster than a classical one; and every headline "advantage" rests on an unproven conjecture that the task is classically hard.

2. What "advantage" means — and why it usually means unverifiable

Quantum advantage denotes a programmable quantum device solving a well-defined task in dramatically less time than the best known classical algorithm, for reasons of asymptotic scaling. The defining tension is epistemic rather than physical: if a result is classically infeasible to compute, it is in general classically infeasible to check by re-derivation.

Random circuit sampling is the clean case. It has strong complexity-theoretic hardness evidence, but verifying its output relies on cross-entropy benchmarking, which itself demands exponential classical resources at scale. And the history is instructive: the 2019 Sycamore task, first advertised as ten thousand years on a classical supercomputer, was later reproduced in about two weeks on a modest GPU cluster. Each fresh advantage claim inherits that provisionality. Google's October 2025 "Quantum Echoes" result — an out-of-time-order correlator on 103 qubits, claimed roughly 13,000× faster than classical — was presented as the first verifiable advantage, but the verifiability is narrow: the quantity it produces can be reproduced by another quantum computer of similar caliber, not by a classical machine or a human. The classical hardness remains conjectural in Google's own companion analysis, Nature 's referees reportedly split on the strength of the claim, and the real-world molecular demonstration that accompanied it is explicitly not beyond classical reach. This is a verification-asymmetry story, not its resolution.

Here the argument must concede something, because the slogan "advantage implies unverifiable" is false as stated, and a careful reader will know it. Shor's algorithm is the counterexample. Factoring is classically hard but trivially checkable — multiply the factors back together. Factoring sits in NP: it has a short witness anyone can verify. So advantage and easy verification can coexist. The exception is real, and it is also narrow, and seeing why is what sharpens the thesis rather than dissolving it. Sort the relevant computations:

1. Witness-checkable tasks (factoring). A succinct classical certificate exists; any party can verify independently. Fully contestable — but a small island. 2. Sampling tasks (RCS, boson sampling, OTOC). No succinct witness; checking is about as hard as computing. Trusted in practice only by cross-checking against another quantum device, or by exponentially expensive benchmarking. Not independently re-derivable. 3. General quantum outputs. Urmila Mahadev's 2018 protocol lets a purely classical verifier confirm an arbitrary quantum computation — a genuine theoretical landmark — but only interactively, under a post-quantum cryptographic assumption, with the quantum machine acting as a cooperating prover. This is verification-as-trust-protocol, not reconstruction. The verifier never re-derives the answer; it runs a challenge-response game it has reason to believe a cheater would fail. 4. Physics-grounded computation (quantum simulation of a real system). The result is checked by experiment: synthesize the predicted molecule, build the material, and the world adjudicates.

The governance-relevant pattern is now visible. Independent contestation — the capacity of an outside party to reconstruct a result well enough to challenge it — works fully only in case 1 and, as physical experiment, in case 4. It fails in case 2. In case 3 it is replaced by cryptographic trust under assumptions. And an AI system that consumes the output of a quantum-advantage computation is typically operating in case 2 or 3. The human's traditional role, checking the work, has no foothold there.

3. Where AI and quantum actually meet

Two of the three popular narratives about AI and quantum are weaker than the marketing suggests; the third is where the real coupling — and the real governance surface — lies.

Quantum for AI is mostly overstated. A string of dequantization results, beginning with Ewin Tang's 2018 classical algorithm for recommendation systems and extended into a general framework for low-rank quantum machine learning, showed that several celebrated "exponential" quantum speedups had classical equivalents once the data-access assumptions were stated symmetrically. Parameterized quantum models face barren plateaus — gradients that vanish exponentially in the qubit count, rendering large circuits effectively untrainable without special structure. And loading classical data into quantum states routinely erases whatever asymptotic edge a model claimed. As of 2026 there is no demonstrated, robust, end-to-end quantum speedup for practical machine learning on classical data.

AI for quantum is the genuine synergy. Machine learning is now doing real work across the quantum stack — device design, pulse calibration, circuit compilation, and especially real-time decoding of error-correction syndromes, where DeepMind's AlphaQubit (Nature, 2024) and reinforcement-learning controllers have produced concrete gains. The structural caveat is that these systems are classical and cannot in general efficiently simulate the quantum machines they assist.

Hybrid orchestration is the realistic architecture. The credible near-term picture weaves classical CPUs and GPUs with quantum processors — NVIDIA's CUDA-Q and its tight GPU-to-QPU coupling for low-latency decoding, adopted across several national laboratories, being the clearest instance. IBM and Quantinuum both now frame the future as quantum-centric supercomputing rather than standalone quantum machines.

That third picture is exactly where the verification asymmetry meets AI governance. The place AI most concretely touches quantum in this decade is inside the error-correction and calibration loop — a classical learned system whose mistakes propagate into a computation that, by the argument above, no human can independently re-derive. This is the double-black-box problem in its sharpest form: an opaque learned component nested inside an opaque computation, with oversight unable to reach the floor.

4. The consequence for contestable governance — and for the In-Between specifically

If oversight depends on re-derivation, it has a hard ceiling in the advantage regime. The practical response is not to pretend the ceiling isn't there but to stratify governance by which kind of verification is even available. Computations should be sorted, in any serious policy text, into the classically checkable, the quantum-checkable-only, the cryptographically-checkable-under-assumptions, and the physically-checkable — with the trust each tier can bear stated explicitly, and the assumption sets made auditable rather than implicit.

This has a direct and uncomfortable implication for relational governance frameworks — the In-Between framework among them — that treat trust as a property of a calibration field spanning human, model, and institution, and that operationalize it through a structure of discrepancy throughput, contestation capacity, and commitment revisability. Run that structure through the quantum-advantage case and it does not survive intact; it survives in pieces, and the pieces are informative.

Contestation, understood as re-deriving the computation to challenge it, is the casualty. In case 2 and case 3 above, an external party — human or classical system — cannot reconstruct the result to mount a challenge; it can at most run a trust protocol or defer to a second quantum machine. The pillar that assumed an outside checker collapses exactly where advantage lives. This is not a flaw the framework can patch by trying harder; it is a structural limit of the regime.

What survives is more than nothing, and worth naming precisely. Discrepancy throughput survives, but relocated: the meaningful discrepancies are no longer between the machine's result and a human re-derivation — that comparison is unavailable — but between provenance claims. Did this output come from physical quantum hardware, from a classical simulation, from a learned decoder, from interpolation? Keeping those sources distinguishable becomes the live governance act. Commitment revisability survives and is in fact vindicated by the clearest near-term case, discussed below. And a dimension the framework already carries — ownership of delegated action — survives most robustly of all, because it never depended on verification in the first place: the human who decides to act on a result, and who bears the consequence, remains a real locus of responsibility even when that human cannot check the result at all.

But the survival of that last piece reframes the whole. If, in the advantage regime, the human cannot verify and reality can, then the calibration partner that matters is not the human but the world. The framework's central image — a between spanning human and AI — may, in this regime, be naming the wrong relation. The load-bearing relation is between the system and physical reality, with the human standing to one side as the party who poses the question and owns what is done with the answer. That is a meaningful role. It is not the role of a co-calibrator who can push back on the substance, and the framework should not be read as if it were.

5. The two concrete stakes

Two near-term developments make this less abstract, and they pull in opposite directions on verification.

Post-quantum cryptography is the worked example of revising commitments under a capability that does not yet exist. NIST finalized its first three standards in August 2024 — ML-KEM, ML-DSA, and SLH-DSA — against a "harvest now, decrypt later" threat in which adversaries record encrypted traffic today to break it once a cryptographically relevant quantum computer arrives. No such machine exists; migration proceeds anyway, on multi-year timelines, because the rational move is to revise a commitment before the capability that would defeat it appears. As a template for governing anticipated rather than demonstrated capability, this is the cleanest one available — and it is an instance of commitment revisability done right.

Quantum simulation of quantum systems is the application that escapes the verification trap. Chemistry, catalysis, materials, batteries — the domains Feynman pointed at — are where a quantum machine has the most defensible reason to outperform classical computation, and, crucially, where the result is checked by experiment rather than by re-derivation. A predicted molecule gets synthesized; a predicted material gets built; reality answers. Even here the sobriety holds — as of late 2025 no quantum machine has outperformed classical methods on a real chemistry problem of consequence, and industrial relevance waits on error correction that does not yet exist — but the epistemic situation is the favorable one. This is case 4, and it is the regime where governance should concentrate its expectations, precisely because the world supplies the ground truth that the human cannot.

Closing

The temptation, for anyone invested in a governance framework, is to find that quantum computing vindicates it — that the new complexity is exactly the situation the framework was built for. The more accurate reading is divided, and the division is the useful part. Quantum advantage diagnoses the limit of contestation-based oversight with unusual sharpness: it marks, almost definitionally for its core cases, the regime where re-derivation fails. A framework that rests its trust on the human's ability to check the work meets a wall there that no amount of relational care removes. What remains is provenance, revisability, and ownership of consequence — a thinner and more honest settlement than "the field stays calibrable," and one in which the verifier of last resort is not the human but the world.

That is not a comfortable conclusion to reach about one's own apparatus. Reaching it anyway is the only version of calibration that means anything.

Sources

Google Quantum AI, Quantum error correction below the surface code threshold, Nature, December 2024 (Willow; Λ ≈ 2.14; below-threshold scaling). Google Research and Meet Willow blog for the RCS comparison figure. Google Quantum AI, Observation of constructive interference at the edge of quantum ergodicity, Nature 646 (October 2025) (Quantum Echoes / OTOC; "verifiable advantage" claim). Companion hardness analysis: arXiv:2510.19751. Skepticism: Science News (A. Harrow), Scientific American (referee split), IEEE Spectrum, Shtetl-Optimized (S. Aaronson). IBM Quantum blog, IBM lays out clear path to fault-tolerant quantum computing (June 2025) (Starling 2029; 200 logical qubits, 100M gates; qLDPC "gross" code; Nighthawk, November 2025). Quantinuum (June 2025), fault-tolerant universal gate set with repeatable QEC; Apollo roadmap. Microsoft, Majorana 1 (February 2025) and Nature editorial note; Physics World, Science, The Quantum Insider on the topological-qubit controversy. A. Bouland, B. Fefferman, et al., complexity of random circuit sampling, Nature Physics (2019); S. Aaronson and S. Gunn on spoofing linear cross-entropy benchmarking (arXiv:1910.12085); classical reproduction of Sycamore: arXiv:2112.15083. U. Mahadev, Classical Verification of Quantum Computations, FOCS 2018; J. Zhang, linear-time CVQC (arXiv:2202.13997). E. Tang, dequantized recommendation systems (2018) and quantum PCA, Phys. Rev. Lett. 127, 060503 (2021); Chia, Gilyén, Li, Lin, Tang, Wang, sampling-based dequantization framework (JACM). J. McClean et al., Barren plateaus in quantum neural network training landscapes, Nat. Commun. 9:4812 (2018). I. Alexeev et al., Artificial intelligence for quantum computing, Nature Communications (December 2025); J. Bausch et al., Learning high-accuracy error decoding for quantum processors (AlphaQubit), Nature 635 (2024). NVIDIA CUDA-Q and NVQLink (GTC, late 2025).

NIST, FIPS 203 (ML-KEM), 204 (ML-DSA), 205 (SLH-DSA), finalized 13 August 2024; NSA CNSA 2.0 timelines. ACS C&EN, Will quantum computing be chemistry's next AI? (November 2025); OECD, anticipatory governance of quantum technologies (2025–2026). Double-black-box framing after A. Deeks.