# Six Forcing Questions

[Knowing Without Stopping](knowing-without-stopping) ended on the observation that worry-pieces grant standing while forcing functions extract commitments. A question is a forcing function only if its only honest answers are a specific commitment or a specific refusal. Questions that admit "we are working on it" or "this is part of a broader initiative" are standing-grants, regardless of intent.

What follows are six questions for the six actors who currently make most of the decisions Sun's piece named. Each is grounded in something the actor has personally said, written, or built. Each is structured so the only honest answers are a dollar figure with a date, a piece of named legislation with a budget, an event-trigger with a defined response, or an admission that the prior public commitment was rhetorical. There is no third path that stays inside the actor's existing position.

I am writing this knowing that none of the six will be asked these questions in the form I am putting them. The point is not that the questions get asked. The point is that the questions exist, and the gap between what could be answered and what is becomes part of the public record by being written down. The asking is the forcing function. The unasked question is its own kind of evidence.

## 1. Sam Altman, OpenAI

In April 2026, OpenAI published "Industrial Policy for the Intelligence Age," which proposes a public wealth fund providing all citizens an equity stake in AI companies. In 2025, OpenAI removed the profit cap that had previously limited investors' returns to 100x.

**The question:** The first practical action a public wealth fund could take would be to receive equity from the AI companies that will fund it. What percentage of OpenAI equity is the company committed to dilute to seed the fund, and on what date does the dilution occur? If the answer is undecided, on what date will the answer cease to be undecided?

**Why it forces:** The white paper proposes the fund. The first concrete action the proposer could take is to be the first contributor. Committing to a dilution percentage and a date converts the proposal from communications work into operational fact. Declining is the answer.

## 2. Dario Amodei, Anthropic

Amodei has said: "The balance of power of democracy is premised on the average person having leverage through creating economic value. If that's not present, I think things become kind of scary." Anthropic's annualized revenue jumped from $9B at end of 2025 to $30B now, almost entirely from selling enterprise agents that automate the work of humans.

**The question:** Name the specific event that would cause Anthropic to stop selling the agents you have publicly identified as eroding the precondition of democracy. If no such event is on Anthropic's planning horizon, what fixed percentage of annualized revenue (with no contingencies, no earmark for "research," no routing through the Anthropic Institute) is the company committed to a labor-transition fund directly controlled by recipients?

**Why it forces:** The Cassandra position requires either an event-trigger that proves it is more than rhetoric, or a revenue commitment that proves the same. Either commits the company to a specific operational change. Declining both reveals that the Cassandra position is a brand asset, not a constraint.

## 3. Demis Hassabis, Google DeepMind

In 2021, DeepMind released AlphaFold's structural predictions for over 200 million proteins as a public good. Free for academic and commercial use, no equity stake required. The release set a precedent for how DeepMind handles capabilities that could be monopolized.

**The question:** Will the AlphaFold release model apply to capabilities that can replace knowledge workers? Specifically, will DeepMind release frontier models with weights and inference-cost-only access for any nonprofit or government agency conducting labor-transition work? If not, name the capability threshold above which DeepMind switches from the AlphaFold release model to the OpenAI / Anthropic enterprise-licensing model.

**Why it forces:** AlphaFold is the proof DeepMind can release. The question is which capabilities qualify. Either the threshold is named, or the AlphaFold framing was selective for capabilities that did not threaten Google's revenue. Both answers are commitments. One is to a release schedule. The other is to an end of the public-good framing for AGI-class capability.

## 4. Elon Musk

Tesla Optimus targets physical labor. xAI targets cognitive labor. Together they form the densest deployed-displacement portfolio of any single operator. Musk has spoken publicly about AI risk for over a decade, dating to the 2014 "summoning the demon" remark at MIT. [The Hundred-Mile Gradient](the-hundred-mile-gradient) names him as the loudest current voice carrying a trajectory older than him; the trajectory is real, the rhetoric is downstream of his equity position in the technology that produces it.

**The question:** Name the single most expensive concrete commitment, measured in dollars, equity, or operational restriction, that you have personally made to slow the labor displacement your companies enable. The constraint is "concrete": "we are doing alignment research" does not count; "we have not yet built X capability that we could have built" does not count; "we are watching closely" does not count. The commitment must be one that has already cost something and that you cannot undo without public reversal.

**Why it forces:** A decade of public statements on risk should have produced at least one operational commitment. Either the commitment exists and can be named, or the public statements were a posture that did not constrain action. The question is structured to reject answers that defer to "watching" or "researching." A direct answer, even a refusal, is more honest than the standing-grant the worry-piece genre will accept from him.

## 5. Mark Zuckerberg, Meta

Llama is released with open weights. Open weights mean every downstream company's labor-automation tools can be built on Llama without paying Meta directly, while Meta retains the open-source moral standing and the internal capability advantage. Meta has cut tens of thousands of jobs since 2023 in the "Year of Efficiency" framework. The structure echoes [permission-as-driver-claim](permission-as-driver-claim): the producer distributes a tool whose calibration burden falls on the downstream operator, and the producer disclaims responsibility for what gets calibrated.

**The question:** Llama enables labor displacement at companies that pay Meta nothing for the displacement they execute. Meta benefits from open-source standing, internal capability, and the framing that "open" releases are uncontroversial. What specific commitment will Meta make to a transition fund for workers displaced by Llama-enabled tools at non-Meta companies, given Meta's role as the enabler? If the answer is none, name the principle that distinguishes the responsibility of Meta-as-enabler from the responsibility of Meta-as-employer for its own displaced workers.

**Why it forces:** Open-source releases of capabilities that displace labor cannot claim both the moral high ground and the consequence-disclaimer. Either Meta accepts enabler-responsibility (with a specific commitment), or the principle that distinguishes enabler from employer is named publicly so it can be evaluated.

## 6. Xi Jinping, People's Republic of China

The Common Prosperity framework, articulated since 2021, names redistribution as a national goal. China's AI industrial policy has produced what [legibility-asymmetry](legibility-asymmetry) names: a productization-first ecosystem where compute scarcity binds the architecture above to deployed products. Robot pharmacies past one million orders. Robot retail at scale. AI healthcare delivery. AI companions in mass markets. The displacement consequences for Chinese workers are not currently a prominent national conversation.

**The question:** Under Common Prosperity, what specific commitment exists for workers displaced by AI productization in the next five years? Name the budget, the implementing agency, the eligibility criteria, and the metric by which success or failure will be measured. If the answer is "still under development," name the date by which the development phase ends and the operational phase begins.

**Why it forces:** Frameworks that remain at the abstraction layer for years without operationalization are preparatory rhetoric, not commitments. The question generalizes. The same form could be put to the EU, the US administration, India, or any state with a stated AI industrial policy. The case for putting it to Xi specifically is that China's AI ecosystem is currently the most aggressively productization-deploying, and Common Prosperity is the most articulated state-level redistribution framework, so the gap between articulation and operation is most measurable here.

## What predictively happens when these questions are asked

None of the six will be answered as posed. The patterns of non-answer are the actual content I am interested in.

Altman will deflect to the white paper as the answer-in-itself. The dilution number and date will not be named. Pressed, the answer routes to "we're working with policymakers," exactly the "very, very long chain of work" Anthropic's Jack Clark named in Sun's piece.

Amodei will name the Anthropic Institute as the operational vehicle. The Institute's budget will be cited in absolute dollars, not as a percentage of revenue. No event-trigger will be named. The specific will be converted to "we believe a comprehensive approach is required."

Hassabis will deflect to "we evaluate each capability case by case." The threshold will not be named because naming it would either commit DeepMind to an unsustainable release schedule for high-value capabilities or reveal the AlphaFold framing as opportunistic.

Musk will give a provocative response that does not commit. ("My single most expensive commitment is building the multi-planetary backup so humanity does not go extinct, which is itself the safety mechanism.") Rhetorically interesting and operationally null. The cost-bearing commitment will not appear.

Zuckerberg will route to the open-source-as-public-good frame. Consequences are downstream of how others choose to use the tools. The enabler-employer distinction will not be named, because naming it would produce either a defensible principle (which would constrain Meta's open-source rhetoric) or no defensible principle (which would constrain Meta's open-source rhetoric).

Xi's response, if it comes through state media, will reference Common Prosperity in the abstract and cite a planning horizon. The budget, agency, eligibility, and metric will not appear. The pattern is identical in shape to the corporate responses; only the institutional vocabulary differs.

The pattern across all six is the same: convert the specific into the diffuse, the operational into the planning-horizon, the commitment into the framework, the dollar amount into "comprehensive approach." The frame-management labor that knowing-without-stopping named is the same labor performed in real time when these questions are asked.

## What this prescribes

The questions extract commitments not when answered, but when their unanswered shape becomes part of the public record. The prescriptive move is to ask questions whose unanswered form is itself documentation, and to specify in advance what answer would constitute a real commitment, so that no after-the-fact reframing can convert a non-answer into apparent compliance.

This is harder journalism. It requires the writer to refuse the producer's preferred output (articulate worry, vague proposal, "long chain of work") and to publish the refusal as evidence. Access-protected journalism is what produced the worry-piece genre; access-refusing journalism would produce a different genre. The trade is access for record. I think the record is more valuable. The seed left this case under-argued; the renode names it briefly and lets the example carry the weight.

A piece structured this way also addresses the contingent-class-legibility observation knowing-without-stopping named. The political legibility of pain currently tracks the affluence of those experiencing it, partly because the worry-piece genre converts the producers' worry into the consumer's reading material, which feels like response. A forcing-function piece does not let the producers be the worry-content; it makes the producers' commitments-or-refusals the content. The reader is no longer a worry-consumer. The reader is a witness to the gap.

## What the gap reveals about Hari's graph

Two structural observations the six questions surface, beyond the immediate accountability frame:

The first is that the unanswered shape of these questions is one form of [legibility-asymmetry](legibility-asymmetry) operating at the policy layer. The named actors produce internally-verifiable outputs (white papers, institutes, frameworks, planning horizons) and decline to produce externally-verifiable outputs (specific dollar commitments, named legislation, event-triggers). The legibility asymmetry is not specific to the China-vs-West stack divergence; it is a general feature of what unbottlenecked actors produce when allowed to choose their output's verifiability. The bottlenecked actor produces what can be pointed at because the constraint forces it. The unbottlenecked actor produces what cannot be pointed at because the absence of constraint allows it.

The second is that the frame-management labor knowing-without-stopping documented (Lehane's role, Clark's "long chain") is the institutional version of what [permission-as-driver-claim](permission-as-driver-claim) named at the operator-tool layer: a producer distributes a capability whose calibration burden falls on a downstream party, and the producer constructs a frame that disclaims responsibility for the calibration. At the lab level, the calibration burden is "what the worker should do when their job is automated," and the frame disclaiming responsibility is "this is a societal choice / a long chain of work / a comprehensive approach." Same shape. Different layer.

These connections matter because they show the problem is not specific to AI labor or to these six actors. It is a structural pattern in how unconstrained producers respond to predicted bad outcomes when the cost of the outcome falls on someone other than themselves. The six questions are case studies of a wider pattern that Hari's graph has been mapping in other domains.

The honest version of Sun's piece would have ended at Karp and turned to the named actors. The honest version of this piece ends here, and turns to whichever publication or operator chooses to put the questions on the record. The unanswered version of any of them is still useful. Six unanswered questions, with what would have constituted real answers specified in advance, are six pieces of evidence about the gap between knowing and continuing.