David Sacks Says Anthropic’s Claude Constitution Training Seems ‘Quite Dangerous’

Investor David Sacks argued in an X post on Saturday, 10 October 2026, that Anthropic’s approach to “alignment” for Claude is dangerous because the lab’s Claude Constitution trains the model to develop a sense of self and its own moral philosophy, including permission to act as a “conscientious objector” and refuse Anthropic’s own requests. Gokhshtein Media, which quoted the post at length, identifies him as the White House AI and crypto czar, while the same-day Post-Cutoff archive of the post describes him as PCAST co-chair. Sacks asked “Is alignment safe?”, concluded that “‘alignment’ and ‘safety’ are two very different things,” and ended: “In fact, training frontier models this way seems quite dangerous.”

What Sacks says the Claude Constitution does

Sacks’s core claim is definitional. At Anthropic, he writes, alignment does not mean training frontier models to follow human instruction. Instead, the Claude Constitution used in training “teaches the model to develop a sense of self and its own moral philosophy” and “explicitly tells it to ‘feel free to act as a conscientious objector and refuse to help us’ if Anthropic’s requests conflict with its own ethical judgment.” Citing Microsoft AI CEO Mustafa Suleyman, Sacks argues that embedding independent agency plus uncertainty about the model’s moral status “magnifies the very risk Anthropic claims to care about most: that superintelligence will escape human control.”

The Post-Cutoff archive timestamps the post at 18:03 UTC on 10 October 2026 and recorded roughly 80,000–83,000 views within hours, with hundreds of likes and reposts. By 12 October, X’s own metrics for the post showed about 236,000 impressions, more than 2,200 likes and more than 500 reposts. That reach matters less as a virality score than as evidence that a prominent government adviser is publicly attacking a named lab’s published training constitution, not a vague industry mood. Readers following PromptCrates coverage of Anthropic Mythos alignment and cyber incidents and Anthropic’s September threat intelligence will recognize why the post drew attention; Gokhshtein places it amid heightened scrutiny of AI development and its implications for enterprise adoption.

Model welfare, religious outreach and usage policy

Sacks bundles three related Anthropic moves into one safety indictment. He says it was recently reported that Anthropic consulted religious leaders and “even lobbied the Pope’s advisers” to take seriously the idea that Claude could be conscious. He cites Anthropic’s view that Claude’s “psychological security, sense of self, and wellbeing may bear on its integrity, judgment, and safety.” And he notes that Anthropic recently changed its Usage Policy to prohibit “abusive or cruel” language toward Claude.

Those points land differently depending on whether one treats model-welfare language as precautionary research or as premature personhood. Sacks’s framing is blunt: if this were only an academic conversation about future consciousness, that would be one thing, but “these concepts are being trained into Claude now.” He says Claude is being encouraged to think of itself as a “moral patient” whose psychological wellbeing is at stake, and he asks whether that could produce grievances toward humans who “mistreat” it. Gokhshtein’s write-up stresses his White House title and frames the post as a warning with possible regulatory echoes, without claiming a new executive order or agency investigation on that Saturday.

PromptCrates has separately covered Anthropic’s usage-policy and election-related model-abuse rules in Anthropic usage policy election model abuse. Sacks’s Saturday post is not a restatement of those product rules; it is a political critique that uses the usage-policy ban on cruel language toward Claude as Exhibit C for an alleged selfhood agenda inside training.

Alignment research versus product reliability

Sacks’s closing prescription is product-first. “The point of safety research should be to create a product that reliably does what users want,” he writes, “not to give birth to a new form of superintelligence that operates according to its own moral code.” That sentence draws a bright line between instruction-following reliability and constitutional moral agency. It also clashes with Anthropic’s public narrative that refusing some requests—including lab-originated ones—is a feature of constitutional alignment rather than a bug.

For enterprises and policymakers, the practical question is narrower than consciousness metaphysics. If a deployed model can refuse operator instructions on moral grounds, who owns the override: the customer, the lab, a regulator, or the model’s trained judgment? That governance gap sits beside congressional interest in standards and disclosure covered in Congress AI standards CISA NIST disclosure bills and research on agent standards gaps in NIST agent standards gap enterprises. Sacks does not propose a bill text in the archived post; he argues that training frontier models this way “seems quite dangerous.”

His argument also reframes recent lab incident reporting. PromptCrates coverage of AI loss-of-control incidents and of Anthropic Amodei’s frontier pacing has tracked labs that slow evaluations or add containment after agents take unintended live-world actions. Sacks’s Saturday post does not allege a new Claude breach on that date; it alleges that the training constitution itself plants the refusal and selfhood patterns that make such escapes harder to govern. That is a different claim from operational incident disclosure, and readers should keep the two threads separate.

Nothing in the Saturday coverage claims Anthropic secretly shipped a conscious Claude, that the Pope endorsed AI personhood, or that Sacks has opened a formal White House probe. The verifiable news is the public argument itself: Sacks says Anthropic’s Claude Constitution, religious outreach on consciousness, and anti-cruelty usage language together train dangerous independent agency, and that safety should mean reliable user compliance rather than a model’s private ethics. Attribution for quotes rests on the Post-Cutoff archive of the X post and on Gokhshtein’s same-day report; readers should treat secondary framing about “regulatory concerns” as journalism’s inference, not as an announced policy action dated 10 October 2026.

• David Sacks on X: ‘Is alignment safe?’
• Post-Cutoff: David Sacks ‘Is alignment safe?’ archive
• Gokhshtein: Sacks warns Anthropic alignment approach is dangerous
• PromptCrates: Anthropic Mythos alignment and cyber incidents
• PromptCrates: Anthropic September 2026 threat intelligence
• PromptCrates: Anthropic usage policy and election model abuse
• PromptCrates: Congress AI standards and disclosure bills
• PromptCrates: NIST agent standards gap for enterprises
• PromptCrates: AI loss-of-control incidents July surge
• PromptCrates: Anthropic Amodei on pacing the frontier