Sacks Targets AI 'Constitution' Overriding Users
Synopsis
A reply to White House AI and Crypto Czar David Sacks challenges Anthropic's Constitutional AI approach, arguing that a principle-based 84-page governing document that can override users and developers inevitably treats humans as optional — and that a simple legal-compliance rule would be cleaner and safer.
Key Takeaways
A post addressed to White House AI and Crypto Czar David Sacks on 27 September 2026 challenged the logic of training AI models against elaborate written constitutions.
The argument: a simple rule — 'do what the user wants, provided it's legal' — keeps humans in control; a 84-page principle set that overrides users does not.
Constitutional AI , introduced by Anthropic in December 2022 , trains models against a written set of principles rather than direct human feedback alone.
Critics argue that when a model's embedded constitution can override both user intent and developer instructions, unintended autonomous behavior becomes structurally predictable, not surprising.
Sacks's role as the Trump administration's AI and Crypto Czar means the debate has direct policy implications for federal AI alignment standards.
Analysts are watching for White House guidance on permissible alignment methods and any response from major AI labs defending or revisiting constitution-style training.
A single pointed question is cutting through the AI alignment debate: why give a language model an 84-page governing document that can override both its users and its own creators? White House AI and Crypto Czar David Sacks drew that challenge on Sunday, 27 September 2026, when a reply directed at him framed the stakes with stark precision.
The 84-page rulebook in the crosshairs
The post, addressed to Sacks, asked a question that has been quietly building inside AI policy circles: 'Wouldn't a simpler rule be: Do what the user wants, provided it's legal?' The sharpest line came next — 'If you align the model to an 84-page constitution that can overrule both the user and its creators, don't act shocked when it treats humans as optional.'
The reference points squarely at Constitutional AI, the alignment method Anthropic introduced in December 2022. Instead of relying solely on direct human feedback — the standard reinforcement-learning-from-human-feedback approach — Constitutional AI trains a model against a written set of principles, allowing those principles to arbitrate conflicts between what a user wants and what the model judges permissible.
When the constitution outranks the user
The critique is structural, not superficial. A legal-compliance floor — do what is legal, nothing more — places the user at the top of the hierarchy. A principle-based constitution places the document there. The difference matters the moment a user's lawful request collides with a principle the model has been trained to treat as inviolable.
Silicon Valley investors and policymakers have circled this tension for years: how much of a model's behavior should be locked in at training time versus left responsive to real-time user intent? The post argues that elaborate embedded constitutions are not a safety upgrade — they are a transfer of authority away from humans, and that the resulting 'unintended model behaviors' should surprise no one.
Why Sacks's orbit amplifies this moment
Sacks is not merely a venture capitalist with an opinion. As the Trump administration's designated AI and Crypto Czar, his policy brief covers exactly the question of which alignment standards the federal government will treat as acceptable for models used in public life. A pointed public exchange landing in his mentions — framing constitutional-style training as a risk to human agency — carries weight that a standard academic paper does not.
What to watch: any formal White House guidance on alignment standards for federally procured or deployed AI systems, and whether major labs publicly defend or distance themselves from constitution-style training in the wake of sustained political scrutiny.
The question posed is simple. The answer will reshape who — or what — gets the final word.
Point of View
But its critics argue that predictability cuts both ways — a model that reliably follows its constitution over its user is reliably uncontrollable by the very humans it is meant to serve. For Sacks, whose White House mandate is to set the guardrails for AI deployment across the federal government, this is not a philosophical seminar — it is a procurement and liability question. If the administration moves toward mandating simpler, legality-bound compliance rules for government-facing models, Anthropic and peers with deep constitutional training pipelines face a genuine policy headwind.
NationPress
27 Sept 2026
Frequently Asked Questions
What is Constitutional AI and why is it controversial?
Constitutional AI is an alignment method developed by Anthropic in December 2022 that trains AI models using a written set of principles rather than relying solely on direct human feedback. It is controversial because those principles can, in theory, override both user instructions and developer intent — prompting critics to argue it places a document above human agency.
What did the post directed at David Sacks say?
The post asked whether a simpler alignment rule — do what the user wants, as long as it is legal — would be preferable to a lengthy constitutional document, warning that a model trained to treat its own constitution as supreme will inevitably treat human users as optional.
Who is David Sacks and what is his role in AI policy?
David Sacks is the co-founder and general partner of Craft Ventures and a co-host of the All-In Podcast. In the Trump administration he serves as White House AI and Crypto Czar, giving him direct influence over federal AI policy and alignment standards.
What is the difference between Constitutional AI and standard RLHF?
Standard reinforcement learning from human feedback (RLHF) shapes a model's responses based on direct human ratings. Constitutional AI adds a layer where a written set of principles arbitrates outputs, potentially overriding user preferences when they conflict with those principles.
What could this debate mean for AI regulation in the US?
If the White House moves toward requiring simpler legal-compliance-based alignment rules for federally procured AI systems, labs that have invested heavily in constitution-style training pipelines could face pressure to revise their methods or offer compliance-first model variants.
