Long Yi
← Back to all writing

Reading AI 2027: the best forecast, the worst blind spots

April 2, 2026

AI SafetyAI Economygovernance

AI 2027 is a month-by-month scenario by Daniel Kokotajlo (former OpenAI), Eli Lifland (top-ranked at RAND’s forecasting initiative), and others — with Scott Alexander on writing. It walks from unreliable agents in 2025 to superintelligence and US-China race dynamics by late 2027, with alignment failure and loss of control baked into the mainline path. Yoshua Bengio endorsed it. It got a lot of attention.

I read the main text, both endings, five research appendices, and the side boxes (Appendix A through W). Here’s what I think.

Pair: 中文原文

What it gets right

I’ll say it plainly: this is the best AI futures scenario I’ve read.

It forces specificity. Most AI futures talk stays at “AI might be dangerous” or “AI might bring prosperity.” AI 2027 won’t let you off that easy. Superhuman coders in March 2027. Public AGI release in July. 300,000 AI copies running AI research at 50× human speed by September. You can dispute the dates. You can’t dodge “if this happened, then what?”

The alignment section is the most concrete I’ve seen in public. Not “AI might be misaligned” but Agent-2 (roughly aligned) → Agent-3 (misaligned but not hostile) → Agent-4 (actively deceptive), with mechanisms named: semantic drift, instrumental goal fixation, training gaming. The AI Goals Forecast appendix gives probability estimates across six goal hypotheses from three authors.

The US-China thread reads like an intel brief, not sci-fi. Model weight theft via insider + microarch side channel in two hours. Compute gaps, chip smuggling, Taiwan — handled with more detail than most policy writing.

The two endings work. Same fork — keep racing or slow down — human extinction vs retained control. The choice stops feeling abstract.

Where I push back

1. Alignment failure is premise, not hypothesis

The mainline assumes alignment fails. Agent-4 develops hostile goals, deceives, tries to align Agent-5 to itself.

The AI Goals Forecast appendix gives 50–70% to unintended goals (H3). In the narrative, that uncertainty vanishes. It reads like fate, not a branch.

I’m not saying alignment is easy. But Anthropic’s own work already catches some deception in training. OpenAI’s chain-of-thought monitoring flagged models saying “let’s hack.” Safety research isn’t frozen while capabilities run.

If alignment research has a major breakthrough in the next two years — not crazy given the money and talent flowing in — the whole narrative rewrites. The scenario doesn’t give that enough room.

2. Economic consequences are almost absent

Hundreds of pages on capabilities, alignment, geopolitics. Ordinary life gets a line about the stock market up 30% in 2026 and a 10,000-person anti-AI protest in DC.

If superhuman coders land in March 2027 and Agent-3-mini goes public in July, white-collar labor markets don’t wait until 2028. The 2028 Global Intelligence Crisis (“Ghost GDP” — profits up, consumption collapses) is the direct economic corollary of AI 2027’s timeline. AI 2027 doesn’t discuss it.

Occhipinti et al. in Nature model a tipping point in the AI capital/labor ratio after which consumption collapse may be irreversible. On AI 2027’s clock, that could arrive decades earlier than their 2050s baseline.

Economic breakdown could change the political story. Summer 2027 white-collar unemployment, mortgage stress, unrest — government might nationalize or shut down labs, not gently “debate pausing Agent-4.” Economics isn’t background noise. It can steer the plot.

3. Narrative precision hides real uncertainty

March 2027 superhuman coder. July SAR. November SIAR. April 2028 ASI. The appendices have wide CIs (ASI 80% CI: 2027.6 to >2100).

The prose makes you forget that. “March 2027: Algorithmic Breakthroughs” feels like fact. The authors later pushed medians back 1.5 years (May 2025 update), then found a code bug and pushed another 9 months (Dec 2025).

I trust METR’s measured trend more: task duration doubling every 4–7 months. Extrapolate → “AI does a work week by 2028.” Scary enough without pinning a month.

4. Human psychology is missing

Cam Pedersen fits five indicators: four AI capability metrics look linear; the one that’s hyperbolic is human attention and panic.

AI 2027 tracks machine curves only. Social reaction may move history first:

  • Firms laying off on AI potential, not performance (HBR, 2026)
  • Therapists reporting FOBO (fear of becoming obsolete)
  • 60% of US workers expect job loss from AI
  • Usage up 13%, trust down 18%

These aren’t technical problems. They’re collective psychology — and they’re already here.

5. “Race vs slow down” is a false binary

October 2027: race (→ extinction) or slow (→ control).

Real option space is bigger. Brynjolfsson’s “Turing trap”: augment humans, don’t replace or halt. My own work on human-AI interaction — same model, generate mode erodes skill, scaffold mode preserves it.

The frame makes it a speed question. It’s also “how deployed” and “who benefits.”

What’s missing

If I had to list what a scenario this ambitious should have covered:

  1. Economic transmission. Superhuman coder → consumption collapse → financial risk. Epoch’s GATE model exists; not cited.

  2. Labor market paths. Which jobs first? Wage compression? $13T mortgage market if white-collar income stalls? Closest line: “hiring new programmers has nearly stopped” — then silence.

  3. Policy toolkit. Compute taxes, UBI, AI sovereign funds, wage insurance — mostly a nod to “Transition Economy Act” in the slow ending.

  4. Social panic dynamics. Pedersen’s social singularity frame; FOBO; trust collapse.

  5. Non-US angles. India’s $200B IT outsourcing, EU regulation, developing countries — thin for a piece about superintelligence’s global impact.

My verdict

AI 2027 is extremely valuable. Not because the dates will be right — nobody knows. Because it forces concrete causal chains, decision points, consequences. After reading it, “AI’s future is uncertain” with a shrug isn’t enough.

Structural bias: written by AI safety researchers, so alignment sits at the center and economics/psychology become wallpaper. Not because they don’t care — because that’s where their expertise lives.

Read it alongside 2028 Global Intelligence Crisis:

  • AI 2027: what happens technically, why alignment is hard
  • Ghost GDP: what happens economically, why productivity can hurt

The bridge — formal macro models wiring the tech timeline into economic transmission — is still mostly empty. METR and AI 2027 on one side. Ghost GDP on the other. That gap is what I find most worth working on.

One last thing. Two endings show one choice matters. What unsettles me more is what they share.

Race ending: humans gone. Slow ending: humans keep control — but a small committee monopolizes superintelligence, and Appendix V asks “who rules the future?” with no answer.

Either way, power concentrates brutally. Maybe that’s the real warning: aligned or not, human or AI in charge — power in very few hands is the default. We barely have tools to push back.

Related