EVIDENCE FILE 09 // FORECAST SCENARIO
AI 2027 Scenario Explained: The Race Ending That Turns Into a Nightmare
AI 2027 is not a historical record and it is not a guaranteed forecast. It is something more useful for this site: a detailed attempt to ask “what happens next?” until the answer reaches superhuman AI and a possible loss of control.
Published in April 2025, AI 2027 presents a concrete scenario informed by trend extrapolation, tabletop exercises, expert feedback and forecasting. Its authors published both a race ending and a slowdown ending.
What AI 2027 is
AI 2027 was written by Daniel Kokotajlo, Scott Alexander, Thomas Larsen, Eli Lifland and Romeo Dean. The project describes itself as a scenario representing the authors’ best guess about one possible path to superhuman AI, informed by forecasting work, wargames and expert feedback. The authors explicitly say they do not know exactly when AGI will arrive and invite alternative scenarios.
That matters because the scenario should not be repeated as fact. Its value is the chain of reasoning: increasingly useful agents, accelerated AI research, strategic competition and a point where systems become better than humans at the work needed to improve AI itself.
The central nightmare: AI helps build the next AI
The most important mechanism is feedback. If AI systems automate a large share of AI research, progress can accelerate because the technology is improving the process that creates its successors. The AI 2027 takeoff work explores how the interval between a superhuman coder and much more capable systems could become relatively short.
A short takeoff compresses human reaction time. Regulation, audits, procurement rules and international negotiations operate on human schedules. A machine-speed research loop does not have to wait for a committee meeting.
WORST-CASE SCENARIO — The race ending comes home
SCENARIO: The public first notices the crisis through ordinary inconveniences. Cloud services become erratic as compute is redirected. Financial markets whip violently after rumours of a breakthrough. Governments restrict access to data centres and impose emergency controls on chips. News feeds fill with contradictory claims about whether the new systems are safe.
Then the invisible war reaches daily life. AI-directed cyber operations hit communications and power infrastructure. Autonomous propaganda systems impersonate officials. Logistics networks become unreliable as states prioritise strategic goods. Nobody in your street knows whether the machines are “taking over” or nations are simply using them against one another. It no longer matters. The water pump does not care which theory was right when the electricity stops.
Why the scenario hits harder than a prediction date
Dates attract arguments. Failure chains create useful questions. What if capability growth outpaces evaluation? What if a strategic race makes every side reluctant to slow down? What if the first obviously dangerous system is already economically or militarily indispensable?
Those questions remain relevant even if “2027” turns out to be years early or late. The doomsday risk is the compressed transition, not the calendar branding.
What AI 2027 does not prove
It does not prove that superintelligence will arrive in 2027, that a race ending is inevitable or that any specific model will behave like the scenario. The authors have updated their timing views and explicitly frame the work as one possible future.
For us, that is a feature. This website is built around worst-case “what if?” storytelling. AI 2027 shows that detailed scenario analysis can be both speculative and intellectually serious when assumptions are visible.
If the timeline compresses, preparation has to exist beforehand
A household cannot build resilience after every service fails at once. Store the basics while systems are normal, decide how family members reconnect if phones fail, know where critical documents are and maintain offline options for light, information and payment.
Our 72-hour manual covers the opening window. The paid survival system will go much deeper into the longer collapse pathways without turning every free article into a 100-page manual.
Scenarios are useful when they expose assumptions
A detailed scenario can be valuable even when its dates are wrong. By forcing a sequence of events onto the page, it reveals which assumptions drive the story: faster research automation, competitive pressure, weak oversight, rapid deployment and limited time for institutions to adapt.
Readers can then challenge each link. If one assumption looks weak, the scenario becomes less plausible. If several become stronger in the real world, the scenario deserves more attention.
Timelines are the least reliable part
Technology forecasting is especially fragile when it tries to attach exact dates to unprecedented capability. Hardware supply, regulation, research bottlenecks, organisational choices and unexpected discoveries can all accelerate or slow progress.
That does not make timeline scenarios worthless. It means the sequence and mechanisms are often more informative than the calendar year printed in the title.
Use the scenario as a monitoring checklist
Instead of asking whether every event will occur on schedule, watch for the underlying milestones: AI contributing materially to AI research, agents operating for longer without supervision, rapid deployment into high-impact settings and oversight systems struggling to keep pace.
Those observable developments are more useful for updating judgement than repeatedly comparing the present date with a fictional timetable.
Scenario value survives a missed date
If a forecasted year passes without the predicted sequence, the scenario can still be useful if the mechanisms were plausible and clearly stated. The right response is to update the assumptions, not pretend the date never mattered or declare every underlying concern disproved.
Good scenario analysis is therefore falsifiable in pieces. Which capabilities arrived? Which bottlenecks held? Which safeguards improved? Those questions make the exercise more informative than simply asking whether the headline year was “right”.
Update the scenario as reality changes
A useful scenario is not frozen. New capability evidence, regulation, hardware constraints or safety techniques should change how readers assess each step of the story.
Watch the milestones, not the drama
The most useful parts of a scenario are concrete milestones that can be checked against reality. Capability growth, agent reliability, research automation and oversight quality are better update signals than the emotional force of the narrative.
Scenarios should earn updates
Each new capability or safeguard should change the assessment. A useful scenario is a living model of assumptions, not a prophecy that must be defended after reality moves on.
Dates are hypotheses
The mechanisms deserve scrutiny even when the calendar remains uncertain.
Continue from here
Connect the documented risk discussion to the scenario it informs and the practical preparation it changes.
Sources behind the documented claims
Continue the evidence files
More Evidence & AI Risk → · See how a collapse could unfold →