05 / EVIDENCE AND SAFETY
How do you judge a quantum-AI claim?
Judge a quantum-AI claim by the exact task, the strength of the comparison and the distance between the measured result and the conclusion. A useful technical advance does not need an AGI prediction attached to it.
What does quantum advantage actually establish?
Quantum advantage establishes a benefit for a specified task under stated comparison conditions; it does not establish superiority across all computation. The relevant benefit might concern runtime, samples or another clearly defined resource.
Google’s random-circuit sampling experiment addressed sampling from a particular class of quantum circuits. Its task was not language modelling, autonomous reasoning or a general industrial workload. Such experiments matter because they test computational capabilities, but their scope should remain visible. A hard-to-simulate output is not automatically a useful input to an AI system. The application requires its own demonstration, including a reason why the sampled information improves the intended task.
What makes a classical baseline credible?
A credible baseline is a well-chosen classical method with adequate tuning, appropriate resources and access to comparable information. Comparing an optimised quantum proposal with a deliberately weak classical method answers very little.
For a quantum kernel, include relevant classical kernels and learning models. For control, include established optimisation techniques. For decoding, compare methods suited to the code and noise model. Match the required output quality before comparing costs. Classical algorithms can improve after a quantum result is published, so a historical comparison is evidence about its stated conditions, not a permanent ranking. Complexity arguments remain useful, but their data-access and precision assumptions must match the application.
What information makes a result reproducible?
A reproducible result exposes enough of the method for another group to repeat the essential comparison and identify where disagreements arise. Code helps, but code without data definitions, settings and resource accounting can leave the main question unanswered.
For learning experiments, look for training and evaluation splits, hyperparameter procedures and uncertainty across runs. For hardware work, look for circuit details, calibration context, measurement counts and treatment of discarded results. For simulations, inspect the noise model and the limits of the classical simulator. Reproducibility is not the same as a guarantee that the conclusion will generalise. It makes the evidence inspectable and gives later work a meaningful starting point.
Why does post-quantum cryptography belong in this discussion?
Post-quantum cryptography belongs here because sufficiently capable quantum computers would threaten important public-key cryptographic assumptions, independently of whether AGI exists. The security question follows from quantum algorithms, not from a theory of general intelligence.
NIST’s post-quantum cryptography project provides standards and supporting material for algorithms intended to resist quantum attacks. ML-KEM concerns key establishment, while ML-DSA and SLH-DSA concern digital signatures. These run on classical systems; they are not quantum key distribution. Organisational planning starts with understanding cryptographic dependencies and data lifetimes. Implementation and migration decisions require qualified security advice and authoritative guidance, not predictions about a particular date for cryptographically relevant hardware.
Does quantum acceleration change AI safety requirements?
Quantum acceleration does not remove the need to evaluate an AI system’s objectives, reliability, access permissions and consequences. It could change the cost or capability of a component, which may change the risks of the overall system.
The NIST AI Risk Management Framework offers a general structure for governing, mapping, measuring and managing AI risks. A quantum-assisted scientific agent still needs controls over tools and experiments, a way to communicate uncertainty and procedures for human intervention. Faster optimisation can pursue a mistaken objective more effectively. Safety evaluation therefore concerns the deployed workflow and its context, not only the processor or the benchmark score.
What would justify a broader AGI conclusion?
A broader AGI conclusion would need evidence of broader capabilities, transfer and dependable behaviour, not just a quantum improvement on one task. The evaluation should test what the system does when its assumptions fail or its environment changes.
Separate a mathematical speed-up, a hardware demonstration, an application benefit and a general-intelligence claim. These are different achievements with different evidence requirements. Ask whether the result survives stronger baselines, includes complete costs and helps on tasks outside the development setting. If the answer is unresolved, the honest conclusion can remain narrow. That does not diminish the work; it identifies which experiment would be informative next and which prediction remains unsupported.
Where can you raise a correction or source question?
You can use the contact page to send a page link, the passage in question and a relevant primary source. Specific technical details make an editorial query easier to assess.