Michael Darius Eastwood · Independent AI alignment researcher
Published
One of the older objections to embedded-substrate constraint on AI compute goes: you cannot put ethics in silicon. Enforcement has to live at the model, or at the deployment, or at the policy layer. Hardware is too dumb, too rigid, too far from the semantics of harm to do useful work. This position was reasonable in 2022 and is empirically dated by 2026. Three references, all in the register, show what the industry now takes seriously.
US patents US 2026/0010411 and US 2026/0010780, filed in July 2025 and granted through the US PTO, describe hardware-level ethics enforcement with a runtime ethics gate performing sub-5 microsecond ethical filtering on inference outputs. The claim is not that the silicon "understands" ethics; the claim is that a hardware path exists to intercept and constrain outputs by policy at latencies compatible with production inference workloads. The patents are issued, not academic; a filing of this content would not have survived examination under a hardware-ethics-impossible presumption inside the US PTO.
The Petrie paper (arXiv:2509.07637, September 2025) works through the engineering of embedded deadman-switch security blocks distributed across an AI accelerator, achieving die-area overhead below one per cent. The economic argument matters as much as the engineering: below-1% overhead is inside the tolerance of a competitive fabrication economics, which is what makes the proposal deployable rather than theoretical. Hardware ethics, on this line, does not need a lithography revolution; it needs an alignment of incentives that the paper claims already exists on paper.
The FlexHEG programme (arXiv:2506.15093), led by Petrie, Aarne, Ammann and Dalrymple and commissioned by the UK ARIA, describes tamper-proof guarantee processors for AI chips. This is not a company's press release; it is a national research agency commissioning work whose premise is that hardware-level guarantees against tampering are a specific, buildable object. The convergence register lists FlexHEG as CONCURRENT with the April 2025 manuscript's Meltdown Triggers argument. The specific overlap is that both take the substrate rather than the model as the locus of enforcement.
Not that the ethical philosophy is solved. That is still hard and this programme's writing on interfaith governance is not a claim to have solved it. What the industry has conceded, in issued patents, arXiv papers, and national research agency commissions, is that embedded-substrate constraint on AI compute is engineerable at production-relevant latencies, at production-relevant overheads, and under production-relevant threat models. The remaining questions are political, economic and regulatory rather than about whether the physics allows a gate. The Chokepoint chapter of the January 2026 book (print 2 Jan, ebook 6 Jan) argued precisely this shift. That argument has aged well.
From the book Infinite Architects: Intelligence, Recursion, and the Creation of Everything by Michael Darius Eastwood.