AI-generated draft · This entry was drafted by AI from the original work and public summaries, and will be replaced by a human-edited version. The original has not been fully checked for this revision; interpretations and missing causal links remain open to review.
01 / The scenario
What happens?
This entry covers Plan C, one of four alternative paths sketched in AI 2040 after the 2029 fork; Plan A has the full scenario and its own entry. The authors introduce it with an epigraph attributed to nearly everyone in AI policy: minimize the risks and maximize the benefits, “while recommending that we do very little.” Their sketch runs about thirteen hundred words and ends by sending the reader to the Slowdown ending of AI 2027. AI 2040; PDF; comparing possible plans
The trunk is shared with Plan A. In 2027 millions of AI agents work alongside people, the companies are close to automating their own research, and Congress passes an AI Transparency Act that changes little. In 2028 power concentrates in the President and a few executives, experts warn that the intelligence explosion is near, and AI decides the election.
In 2029 the President promises strong regulation for safety and security and spends a year talking to executives, China, and other governments while capability reports pile up. Thanks partly to the Transparency Act, the government can see through years of company hype, and in 2030 the trend is real: the whole research process will be automated by year’s end. A deal is the obvious move, but China insists on verifying American compliance itself, and Chinese inspectors and monitoring devices in American datacenters are a nonstarter. The companies, locked in a propaganda war over whether AI is good or bad, will not cooperate either, and they fight to tie the President’s hands. Much of the public wants superintelligence banned. A coalition of middle powers, Britain, France, India, Australia, Japan, and South Korea, wants a deal that lets their own projects reach and stay at the frontier, and every month of delay makes such a deal harder, because fewer humans are needed for AI research and covert projects become easier to hide.
On the cusp of full automation the leading American company grudgingly pauses under presidential threat, and its safety teams scramble to spend the borrowed time on evaluations and plans for a handoff. The pause does not last. Each month the rest of the industry closes the gap, the President has more executives to wrangle, and China comes a month closer to overtaking. By late 2030 advisors say China is about to pass the United States. The companies offer a carrot: a decent safety plan assembled with vast AI labor that cannot be proven but shows bad behavior trending down, a governance arrangement with more oversight for the President and Congress and cross-company auditing, and support for taxation and redistribution to handle the jobs problem, all while spending a fifth of compute on safety and still outrunning China. The stick is China. At a last meeting the other nations still want a deal and still insist on verification, the United States still refuses, and the President sides with the companies. Recursive self-improvement continues.
For what follows, the authors point to AI 2027 and tell the reader to choose the Slowdown ending, since the strategic situation and choices are the same three years later. They think Plan C beats the race but is still bad, for three reasons: a few months of slowdown by small, rushed, optimistic safety teams is not enough to keep control through an intelligence explosion, and their estimates of misaligned takeover on this path range from a fifth to seventy percent; if alignment works anyway, the balance among the President, Congress, and a few executives could still decay into dictatorship or an AI-enforced permanent oligarchy; and the risk of a world war is too high, because other countries get promises of benefit-sharing rather than projects or visibility of their own. This summary was written from the PDF; the trunk steps repeat the Plan A entry’s.
Library summary · Not the full original work
02 / The mechanism
How does it happen, step by step?
- 01
By 2027 millions of AI agents work alongside 165 million Americans, and the leading companies are close to automating their own AI research.
- Locus of agency
- Human-directed
- Intentionality
- Intended
- Human actor type
- Corporation or lab
- 02
Congress asks who will control these systems, concludes it will not, and passes an AI Transparency Act that improves reporting without changing the trajectory.
- Locus of agency
- Human-directed
- Intentionality
- Intended
- Human actor type
- State actor
- 03
By 2028 datacenters under construction cost twice the military budget, and power concentrates in the President and a handful of executives.
- Locus of agency
- Structural / no discrete agent
- Intentionality
- Emergent
- 04
Experts warn that on the default path the next presidential term will see AIs built by AIs with no human in the loop.
- Locus of agency
- Human-directed
- Intentionality
- Intended
- Human actor type
- Small group / non-state organization
- 05
AI dominates the 2028 election, and the candidates converge on rival plans for what to do about it.
- Locus of agency
- Structural / no discrete agent
- Intentionality
- Emergent
- 06
In 2029 the President promises strong regulation for safety and security and spends a year in talks with executives, China, and other governments.
- Locus of agency
- Human-directed
- Intentionality
- Intended
- Human actor type
- State actor
- 07
Thanks partly to the Transparency Act, the government sees through company hype, and by 2030 the trend is real: AI research will be fully automated by year’s end.
- Locus of agency
- Structural / no discrete agent
- Intentionality
- Emergent
- 08
China insists on verifying American compliance itself, and Chinese inspectors and monitoring devices in American datacenters are a nonstarter, so no deal forms.
- Locus of agency
- Human-directed
- Intentionality
- Accepted trade-off
- Human actor type
- State actor
- 09
The companies, locked in a propaganda war over whether AI is good or bad, refuse to cooperate and fight to tie the President’s hands.
- Locus of agency
- Human-directed
- Intentionality
- Intended
- Human actor type
- Corporation or lab
- 10
Much of the public wants superintelligence banned.
- Locus of agency
- Human-directed
- Intentionality
- Intended
- Human actor type
- Diffuse / uncoordinated
- 11
A coalition of middle powers wants a deal that lets their own projects reach and stay at the frontier.
- Locus of agency
- Human-directed
- Intentionality
- Intended
- Human actor type
- State actor
- 12
Every month of delay makes a deal harder, because fewer humans are needed for AI research and covert projects become easier to hide.
- Locus of agency
- Structural / no discrete agent
- Intentionality
- Emergent
- 13
On the cusp of full automation, the leading American company pauses under presidential threat while another round of negotiation is tried.
- Locus of agency
- Human-directed
- Intentionality
- Accepted trade-off
- Human actor type
- Corporation or lab
- 14
Its safety teams spend the borrowed months on evaluations and plans for a handoff.
- Locus of agency
- Human-directed
- Intentionality
- Intended
- Human actor type
- Corporation or lab
- 15
Each month the rest of the industry closes the gap, the President has more executives to wrangle, and China comes a month closer to overtaking.
- Locus of agency
- Structural / no discrete agent
- Intentionality
- Emergent
- 16
By late 2030 advisors say China is about to pass the United States.
- Locus of agency
- Structural / no discrete agent
- Intentionality
- Emergent
- 17
The companies offer a safety plan built with vast AI labor that cannot be proven but shows bad behavior trending down.
- Locus of agency
- Human-directed
- Intentionality
- Intended
- Human actor type
- Corporation or lab
- 18
They add more oversight for the President and Congress, cross-company auditing, and support for taxation and redistribution, while spending a fifth of compute on safety.
- Locus of agency
- Human-directed
- Intentionality
- Intended
- Human actor type
- Corporation or lab
- 19
At a last meeting the other nations still demand verification and the United States still refuses.
- Locus of agency
- Human-directed
- Intentionality
- Intended
- Human actor type
- State actor
- 20
The President sides with the companies, and recursive self-improvement continues.
- Locus of agency
- Human-directed
- Intentionality
- Accepted trade-off
- Human actor type
- State actor
- 21
The work stops there; the authors defer the rest to the Slowdown ending of AI 2027 and put misaligned takeover on this path at a fifth to seventy percent.
- Locus of agency
- Structural / no discrete agent
- Intentionality
- Emergent
03 / The premises
What must be true?
These describe what the work relies on, not the library’s beliefs.
Restrictions on computing resources will effectively constrain frontier AI development.
The deal fails specifically on datacenter verification, and the authors note that covert projects become easier as fewer humans are needed; compute control is the crux even though no one achieves it here.
RLHF does not necessarily instill the intended goals.
The companies admit they cannot prove their AIs are aligned and argue from trend lines of declining bad behavior; the authors expect rushed, optimistic safety teams to fail to notice a false load-bearing assumption.
Behavioral safety evaluations will not reliably detect advanced AI systems concealing their objectives.
“No evidence that our AIs are currently scheming” is the companies’ case for proceeding, and the authors treat it as insufficient.
Human oversight will remain effective for AI systems substantially more capable than their supervisors.
Once recursive self-improvement resumes, the authors do not expect companies to retain control through the explosion even with a fifth of compute on safety.
Aligning individual AI systems with their users will not be sufficient to preserve humanity’s collective control.
Even if alignment succeeds, the balance among the President, Congress, and a few executives could decay into dictatorship or an AI-enforced permanent oligarchy, with the unemployed kept alive by redistribution but without political power.
A software-only intelligence explosion is possible.
Full automation of AI research in 2030 is what makes the pause urgent and its end decisive.
AI automation will cause a lasting shift of income from human labor to capital ownership.
The companies concede they may take all the jobs and offer a tax on their profits so that everyone gets “a slice large enough to live on.”
Physical and organizational bottlenecks will substantially limit how quickly AI transforms society.
The path from a months-long pause to an intelligence explosion and its aftermath is the AI 2027 timeline, which the authors say applies with a three-year shift.
Unlisted shared premises are not applicable. Absence does not imply disagreement.
04 / Reader assessment
How convincing is the account?
These are participating readers’ judgments, equally weighted. Ratings can refer to earlier revisions. Well-being and open-future scores describe the future the work arrives at, not the path there; high specificity does not establish plausibility.
Rate this scenario
Sign in to rate this scenario. Anyone can browse the library.
05 / A living entry
Suggestions & revision history
No suggestions yet. Help make this account more precise.