As of October 2026, AI-designed quantum error correction exists exclusively in simulation. General-purpose models and agents have demonstrated massive theoretical gains: OmniQEC has used frontier LLMs to orchestrate code generation that outperforms bivariate-bicycle benchmarks at 98 and 240 physical qubits OmniQEC: discovering practical quantum error-correcting ... OmniQEC: discovering practical quantum error-correcting ... OmniQEC: discovering practical quantum error-correcting ..., while QC Design’s Meridian reports greater than 10× median logical-error reductions for syndrome-extraction circuits across over 100 design tasks QC Design Reports 10x Logical Error Reduction With ... QC Design Reports 10x Logical Error Reduction With ... Over 10× reduction in logical error rates with Meridian. However, no such scheme has yet been tested on quantum hardware. Recent hardware milestones driven by AI—such as Google’s reinforcement-learning control that cut logical error by 20% Google Uses AI Reinforcement Learning For Quantum ... Towards a quantum computer that learns from its errors or Quantinuum’s AI-decoded color codes Quantinuum Breaks Error-Correction World Record by 10x ... Quantum error correction well below the color code threshold—do not meet the criteria, which strictly exclude calibration, control, and neural decoders.

The fastest path to a qualifying hardware run is likely a device-tailored syndrome-extraction circuit or a small quantum low-density parity-check (qLDPC) code deployed on an existing, flexible architecture. Trapped-ion and neutral-atom platforms offer the all-to-all connectivity required to test new codes or circuits without waiting for next-generation hardware. IonQ has already run nine codes on a single 40-ion device, and neutral atoms have successfully tiled high-rate codes QLDPC Codes, The Complete 2026 Guide To 10x .... Tailoring extraction circuits to a specific device’s noise profile—exactly what Meridian targets—offers a highly plausible route to achieving a 20% advantage over published human baselines on the same chip https://research.google/blog/a-colorful-quantum-future/ Quantinuum Breaks Error-Correction World Record by 10x ....

A successful demonstration requires only a 20% logical error reduction, which is a relatively low bar compared to the 10× to 29× gains currently reported in simulation. Given the intense competition and strong incentives among major players—including Google, IBM, Quantinuum, and Anthropic in partnership with QuEra—there is a massive push to translate these simulated designs into real-world hardware runs. Since the first major simulation results appeared between June and September 2026, the typical 6–15 month timeline for quantum hardware experiments and preprint publication points to a highly active period beginning in mid-to-late 2027. But activity need not immediately yield a qualifying result: compared with related milestones that can be reached in simulation, this one requires physically implementing the design and demonstrating its advantage on a real device. That additional friction makes early 2028 a more appropriate central expectation.

However, translating simulated gains to physical hardware remains a formidable challenge. Many simulated evaluations, such as those by OmniQEC, rely on circuit-level depolarizing noise OmniQEC: discovering practical quantum error-correcting .... In reality, hardware is plagued by leakage, crosstalk, and drift, which often erode theoretical advantages. Furthermore, the resolution criteria are exacting: the AI design must beat the best human-designed scheme on the same device, matched for physical qubit count, and reported within six months. Beating human baselines on heavily optimized platforms is exceptionally difficult; for instance, human-designed color codes on Quantinuum's Helios have reached roughly 5e-6 logical error per cycle and are improving rapidly Quantum error correction well below the color code threshold Quantinuum Breaks Error-Correction World Record by 10x .... Researchers will likely target less-optimized academic hardware or cloud devices to clear this hurdle.

I place the median first qualifying public report at February 28, 2028, reflecting the necessary lead time for device-specific implementation, competitive baseline runs, and preprint preparation, as well as the physical friction of transferring simulated designs to hardware. A quick trapped-ion run of an optimized syndrome circuit could still produce an early result: I assign a 10% probability of occurrence by April 1, 2027, and a 25% probability by August 1, 2027. The chance reaches 75% by December 15, 2028. Conversely, a substantial right tail extends through 2029 and beyond, accounting for the risk that the simulation-to-reality gap proves stubborn, that human baselines remain too strong on capable devices, or that early claims are disqualified due to disputes over whether a general-purpose AI genuinely produced the central idea rather than merely serving as a search tool. I place the 90th percentile at July 31, 2030.