In February 2026, Cortical Labs demonstrated 200,000 living human neurons playing Doom.[1] Not in a research lab under controlled conditions. On a commercial product with a Python API and a price tag around $35,000.[2] An independent developer replicated the integration in roughly a week, compared to the eighteen months the original DishBrain Pong experiment required.[1] Separately, FinalSpark in Switzerland has been selling remote cloud access to brain organoids since 2024, with sixteen organoids available via Python API around the clock.[3]
The tissue was voluntarily donated. What the donors consented to and what the tissue is now doing may not be the same thing. Third-party workloads are being processed on donor-derived human neural tissue today, commercially, through standard programming interfaces. I have not found published documentation from either operator specifying what those donors were told their tissue would be used for, or whether the current computational use falls within the original consent scope.
The evidence is peer-reviewed, independently replicated, and commercially deployed.
In 2022, Kagan et al. at Cortical Labs and Monash University published the DishBrain system in Neuron, showing that human iPSC-derived neural cultures exhibit adaptive, goal-directed learning in a Pong environment within five minutes of play (Kagan et al., Neuron 2022).[4] The cultures learned through stimulation and feedback, in a pattern consistent with biological learning.
A year later, a completely independent group confirmed that the result was not a one-off. Cai et al. at Indiana University published Brainoware in Nature Electronics, demonstrating brain organoid reservoir computing for speech recognition and nonlinear equation prediction (Cai et al., Nature Electronics 2023).[5] Different lab, different platform, same conclusion: biological neural networks perform real computational tasks.
By 2024, FinalSpark had moved from research to commercial infrastructure. Their Neuroplatform, published in Frontiers in Artificial Intelligence, offers sixteen organoids with Python API access and around-the-clock cloud availability. Over a thousand organoids have been processed. Eighteen terabytes of neural activity data collected (Jordan et al., Frontiers in AI 2024).[3] Researchers anywhere in the world can rent time on living human neurons through a web browser.
Then there is the finding that makes the governance question urgent. Gabriel et al. published in Cell Stem Cell in 2021 that iPSC-derived brain organoids spontaneously develop bilaterally symmetric optic vesicles containing light-responsive photoreceptors, with axonal projections connecting to forebrain regions (Gabriel et al., Cell Stem Cell 2021).[6] The reproducibility rate was 66% across five independent donor lines. These substrates developed complex sensory structures that were not present at initialization and were not directed by the operators. The governance question does not depend on whether biological computing scales to data-center size. It depends on whether third-party workloads are being processed on donor-derived tissue.
The Governance Gap
Who governs what these neurons are permitted to compute? I examined the public documentation of both major commercial operators as of March 2026. I did not find published consent documentation tracing tissue to original donors, institutional ethics oversight structures governing computational use, acceptable use policies restricting workload categories, or decommissioning protocols for trained neural cultures.[7] The reader can verify this independently. Both operators' websites, publications, and public documentation are accessible. The absence of these frameworks is not a hidden finding. It is a publicly observable gap.
The HeLa precedent is directly relevant. Cells taken from Henrietta Lacks in 1951 without informed consent were commercially distributed for decades before her family was notified.[8] The parallel is not exact, but the structural pattern is identical: biological material enters commercial use before governance frameworks exist.
Existing biological material governance systems address procurement and storage. UNOS governs organ transplant allocation. The AATB sets standards for tissue banking. ISBER provides biobank management guidelines. None of these address computational runtime enforcement. The Baltimore Declaration (Hartung et al., Frontiers in Science 2023) calls for governance of organoid intelligence but proposes no enforcement mechanism.[9] The research community has identified the need. No one has specified the cryptographic layer that would make governance enforceable rather than aspirational.
Why This Is Different
Biological computing creates a trust boundary that neither AI governance nor biosafety frameworks cover. The substrate is donor-derived human tissue. Unlike silicon, it carries consent obligations that evolve with the substrate's capabilities. A donor who consented to pattern recognition research may not have consented to whatever the substrate develops the ability to do after six months of continuous stimulation. The Gabriel organoid finding demonstrates that these substrates develop capabilities not present at initialization. Consent given at procurement may not cover what the tissue becomes.
This problem compounds because the substrate itself changes. Silicon does not spontaneously develop new structures. Neural cultures do. Governance artifacts that were accurate at deployment become stale as the substrate matures. The policy was correct when established, but the substrate outgrew it. I would call this consent drift, and it is why any governance system that treats the initial consent as permanently valid is architecturally inadequate for a substrate that learns and adapts.
The policy was correct when established. The substrate outgrew it.
And when a biological computer reaches end-of-life, the decommissioning problem has no analog in IT infrastructure governance. An unplugged server is inert. An unpowered neural culture may still contain viable, donor-derived cells with a data-processing history. Decommissioning requires verified destruction of viable human tissue, and the decommissioning process itself requires cryptographic attestation that no viable cells remain. NIST AI RMF, CISA Secure by Design, SLSA, and Sigstore were designed for software and silicon. They were not designed for substrates that carry consent obligations, develop new capabilities autonomously, and require verified biological destruction at end-of-life.
Any governance system adequate for biological computing would need cryptographic attestation at every interface between the biological tissue and the digital system, covering who the tissue came from, what it is permitted to compute, and proof that those boundaries were enforced. And those cryptographic protections would need to outlast the substrate itself. If a neural culture operates for years, records signed today must remain tamper-evident for the culture's entire lifespan. Post-quantum cryptography is not optional for long-lived biological substrates, given the Babbush et al. result demonstrating that elliptic curve schemes face a credible quantum threat timeline.[10]
But coverage and durability are only half the problem. The cryptographic inventory would need to include the biological dependencies themselves, not just software and firmware, but tissue provenance, chemical environment, and consent scope. And because the substrate changes, the governance system would need behavioral monitoring tied to the original consent scope, with mechanisms to pause computation and re-engage the donor when the substrate exceeds its authorized parameters. Static governance for a dynamic substrate is not governance. It is documentation.
We have been working on this problem at Attested Intelligence. Our Attested Governance Artifacts framework extends to biological substrates through the same architecture we built for autonomous AI agent governance, with bio-compute-specific additions for tissue provenance, consent drift monitoring, and verified end-of-life. The full technical framework is published as WP-AIH-2026-001 (v8.0), including reference implementations, a conformance checklist, and an eleven-phase lifecycle from donor consent through tissue deprovisioning. The paper is FRAND-committed and the schemas are open-source under CC-BY 4.0 and Apache 2.0.
The question for the industry is timing. NIST CAISI listening sessions are happening this month. CycloneDX engagement is underway. The window for correct-by-design governance is open now. The retrofit window, as PQC migration in classical IT infrastructure has already demonstrated, costs ten to one hundred times more than building it in from the start.
Two hundred thousand neurons playing Doom. Cloud APIs to brain organoids. No published governance framework.
The question is whether governance will arrive before or after something happens to donor-derived tissue that no one authorized and no one can prove.
References
- Cortical Labs. “Living Human Brain Cells Play DOOM on a CL1.” YouTube, February 25, 2026. CTO/CSO narration. Independent coverage: The Guardian, Tom's Hardware, New Scientist, PC Gamer (Feb-Mar 2026).
- CL1 pricing (~$35,000). cortical.io product page; Top Gear (March 2026).
- Jordan FD et al. “Open and remotely accessible Neuroplatform for wetware computing.” Frontiers in AI 7:1376042, 2024. DOI:10.3389/frai.2024.1376042.
- Kagan BJ et al. “In vitro neurons learn and exhibit sentience when embodied in a simulated game-world.” Neuron 110:3952-69, 2022. DOI:10.1016/j.neuron.2022.09.001.
- Cai H et al. “Brain organoid reservoir computing for artificial intelligence.” Nature Electronics 6:1032-39, 2023. DOI:10.1038/s41928-023-01069-w.
- Gabriel E et al. “Human brain organoids assemble functionally integrated bilateral optic vesicles.” Cell Stem Cell 28(10):1740-57, 2021. DOI:10.1016/j.stem.2021.07.010.
- Brennan J. “Governance Gaps in Commercial Biocomputing: An Audit of Cortical Labs and FinalSpark.” ResearchGate, 2025. Findings independently verifiable from operators' public documentation.
- Skloot R. The Immortal Life of Henrietta Lacks. Crown, 2010.
- Hartung T et al. “The Baltimore Declaration toward the exploration of organoid intelligence.” Frontiers in Science 1:1068159, 2023. DOI:10.3389/fsci.2023.1068159.
- Babbush R et al. Google Quantum AI. “Quantum error correction below the surface code threshold.” March 30, 2026. Demonstrates 256-bit ECDLP breakable in under 9 minutes with fewer than 500,000 physical qubits.