TECHNICAL NOTES · AI GOVERNANCE

The White House superintelligence accord: what human control means in engineering

On 29 September 2026 the heads of six of the largest AI companies signed a voluntary accord at the White House, and the administration began calling advanced AI "super intelligence". The accord is about auditing the models. For anyone who builds aircraft, spacecraft or defense systems, the harder question is a different one: who keeps the model's output under human control once it reaches your design?

What was signed

The document, reported as The White House Accord on Superintelligence: A Joint Commitment on Frontier SI Responsibilities, was signed by President Trump and six industry leaders: Sundar Pichai (Google), Dario Amodei (Anthropic), Mark Zuckerberg (Meta), Greg Brockman (OpenAI), Elon Musk (xAI) and Jensen Huang (NVIDIA). It asks every company training and deploying frontier models to put four layers of control in place:

  1. Internal controls that monitor model capabilities and alignment during training and deployment, in areas such as cybersecurity, biosecurity and chemical threats.
  2. An internal team that checks those controls, monitoring and detection are working as intended, and that problems get fixed.
  3. An independent external auditor or evaluator.
  4. An independent board committee that receives the reports from the internal team and the auditors, and makes sure issues are remediated.

The same day, an executive order, Inaugurating the Era of Super Intelligence, directed federal agencies to refer to artificial intelligence as "super intelligence" (SI) in official communications, and gave the President's science and technology adviser 60 days to propose legislative language defining SI — including whether it should modify, expand or replace the existing legal definition of AI.

What it does not do

The accord is voluntary. As reported, it carries no penalties, does not require audit findings to be published, and routes results to each company's own board rather than to a regulator. It does not commit anyone to pause or slow development, and sets no compute or capability thresholds. The President described it as "morally" binding, and the administration argued against creating an FDA- or FAA-style regulator for AI. Critics were blunt. "What other trillion-dollar industry marks its own homework?" asked AI researcher Toby Walsh, quoted by Al Jazeera.

Whatever one thinks of the policy, it is now the reference point for how frontier AI companies in the US say they will govern their most capable models.

An aerospace reading: the four layers look familiar

Engineers in safety-critical industries will recognise the structure. Aerospace assurance has long rested on the same ideas: controls built into the process, verification performed independently of the people who did the work (DO-178C requires independence for many objectives at software Levels A and B), and oversight that sits above the programme.

The difference is who closes the loop. In aviation, a certification authority — not the manufacturer's board — accepts the evidence, and the evidence has to trace to requirements that can be re-verified years later. In the accord, the loop closes inside the company. That is a policy choice, and it may change: the accord itself reportedly leaves room for codifying the measures in law.

Auditing the model is not controlling the output

There is a subtler point for engineering teams. The accord governs how models are built and released. It says nothing about whether a particular answer that lands in your stress analysis, your requirements baseline or your flight code is correct. A perfectly audited model can still produce a plausible wrong number. Human control, in engineering terms, is not a property of the model. It is a property of the workflow around it:

  • Every number is reproducible. Calculations are executed deterministically and can be re-run, not regenerated — see verifiable AI.
  • Every claim carries its evidence. Regulatory and standards statements cite a source, or are flagged as unverified.
  • An independent check sits between the model and the design. A second derivation, a reviewer, a test — something that doesn't share the model's blind spots.
  • A named engineer signs off. Design authority and certification sign-off stay with people.
  • Controlled data stays inside your boundary. A frontier model's audit doesn't change where ITAR, EAR or CUI data is allowed to go — see air-gapped and on-prem AI.

Regulators in aviation are moving the same way. EASA's proposed Issue 03 concept paper limits general-purpose large models to the lowest assurance levels precisely because their design process can't be controlled — so the assurance has to come from what surrounds them.

What this means for AETHER users

AETHER was built on that premise. The language model frames the problem; code computes the numbers; an independent derivation and a model from a different family check the result; TELOS and THEMIS ground regulatory claims in live eCFR and NASA sources; and anything that can't be verified is printed as a LIMIT line rather than hidden. Better-audited frontier models are welcome. They don't remove the need to verify what they tell you — and in aerospace, nothing ever has.

Sources

Published · 1 October 2026 · All technical notes