AI in 2027: What Every Prediction Gets Wrong (and Right)
Every AI prediction is wrong. The only question is whether it is wrong by months or decades. Sam Altman said AGI would arrive during Donald Trump’s first term — that was wrong. Demis Hassabis said five to ten years — three years ago, which makes it two to seven years now. Geoffrey Hinton said thirty years to extinction, then revised to ten. The AI researcher survey says 2047, but 10% say before 2030.
The predictions are wrong because they are trying to forecast a phenomenon that compounds exponentially. Every six months, the capabilities double. Every year, the experts revise. The range of expert predictions for AGI compressed from 2060 to 2033 in six years. The Metaculus community median collapsed from the 2040s to 2028 in the same period. Something is happening faster than the people studying it can model.
But the predictions are also the only map we have. Not because they are accurate, but because they reveal what the people building these systems actually believe about what they are building. When Dario Amodei says 50% chance within three years, he is not guessing. He is telling you what Anthropic’s internal trajectory looks like. When Demis Hassabis narrows to 2030 plus or minus a year, he is telling you what DeepMind’s roadmap shows.
Here is what the most informed people on earth actually think happens next — and what the data supports.
The CEO scorecard: what they said vs. what happened
The first thing to understand is that the people building AI are not good at predicting when their own systems will reach specific milestones. They are consistently optimistic. Here is the scorecard:
| Forecaster | What they said | When | What actually happened |
|---|---|---|---|
| Sam Altman | AGI during Trump’s first term | 2017 | Wrong. Revised to end of 2028 (August 2026) |
| Sam Altman | Superintelligence “a few thousand days” away | Sept 2024 | Still waiting. Translates to roughly 2027-2031 |
| Sam Altman | ”2027 may see robots that do tasks in the real world” | June 2025 | Confirmed — physical AI is advancing, but “tasks” is doing heavy lifting |
| Sam Altman | Conceded overestimated disruption speed; revised AGI to end 2028 | Aug 2026 | Most honest public revision from any CEO |
| Dario Amodei | ”Powerful AI” as early as 2026 | Oct 2024 | Partially confirmed — GPT-6 Astra and Claude 5.1 are capable, but “powerful” depends on definition |
| Dario Amodei | 90% confidence “country of geniuses” within 10 years | Feb 2026 | Not yet testable. 50% chance in 1-3 years puts it at 2027-2029 |
| Demis Hassabis | Human-level AI in 5-10 years | Early 2025 | Narrowed to 2030 plus or minus a year (June 2026) |
| Shane Legg (DeepMind) | 50% minimum AGI by 2028 | Jan 2026 | Not yet testable |
| Elon Musk | AGI by end of 2026 | 2025 | Three months left. Extremely unlikely on current trajectory |
| Andréj Karpathy | About a decade out (~2035) | 2024 | Most conservative among frontier insiders. Not yet testable |
The pattern is clear: every CEO who has revised has moved toward later dates, not earlier ones. Altman moved from “during Trump’s first term” to “end of 2028.” Hassabis moved from “5-10 years” to “2030 ± 1.” Kokotajlo moved his personal median from 2027 to 2035. The direction of revision is consistently toward more caution, not less.
The one exception is Altman’s GPT-5.6 description — he called it “very AGI-like” in July 2026 and said if shown it in 2019, “they would have called it AGI.” But he immediately undercut the claim by noting that “month 24 after superintelligence…not very much happens — it’s a pretty smooth exponential.” The gap between “AGI-like” and “world-changing” is apparently larger than anyone expected.
The AI 2027 scenario: what it actually says
The most detailed prediction about what happens next is the AI 2027 scenario, published April 3, 2025. It is worth understanding in detail because it is the most rigorous attempt to map the near future, and because the authors have been honest about where it is and is not tracking.
The authors: Daniel Kokotajlo (lead, former OpenAI), Scott Alexander, Thomas Larsen, Eli Lifland, and Romeo Dean. Published under the AI Futures Project nonprofit with Lightcone Infrastructure. Informed by approximately 25 tabletop exercises and over 100 expert feedback sessions.
The timeline:
- March 2027: Superhuman coder — AI automates end-to-end software engineering
- August 2027: Superhuman AI researcher — AI automates AI research itself
- November 2027: Superintelligent AI researcher
- December 2027: Artificial superintelligence
- Two endings: “Slowdown” (aligned ASI benefits humanity) and “race” (misaligned AI causes catastrophe)
The core mechanism: Once AI automates coding, it begins automating AI research, creating a recursive self-improvement loop. Scott Alexander explains: “Smarter researchers can use compute more efficiently…about half of all AI scaling since 2020 has been algorithmic progress. If we hold compute constant, but get 10x algorithmic progress, then we get 5x overall AI improvement.”
The authors’ own revisions: In July 2025, the authors pushed their median back 1.5 years but kept the superhuman coder as a serious 2027 possibility. Scott Alexander clarified: “maybe think of this as a vision of what an 80th percentile fast scenario looks like — not our precise median.” Kokotajlo himself moved his personal median from 2027 to 2035 by January 2026.
The tracking data: As of July 2026, 16 of 53 tracked predictions are confirmed, 3 are ahead of schedule, 8 are on track, and 4 are behind. The scenario is running slightly ahead of its original timeline, but the big milestones (superhuman coder in March 2027) have not yet been reached.
The AI 2027 scenario is not a forecast. It is a detailed thought experiment about one possible fast-takeoff future. The authors have been transparent about its limitations and their own shifting beliefs. Treat it as the upper bound of what could happen, not the most likely outcome.
The data that supports fast timelines
The hard data behind the predictions is more compelling than the predictions themselves.
METR time horizon data: The length of coding tasks AI can complete has been doubling on a predictable schedule. From 2019 to 2024, the doubling time was approximately 7 months. Post-2023, it accelerated to approximately 4.3 months (130.8 days). Jack Clark coined the term “coding singularity” based on this curve. If the trend holds, AI will handle month-long coding tasks by mid-2027 and year-long tasks by 2028.
SWE-Bench: The standard benchmark for real-world software engineering went from 2% to 93.9% in 30 months. That is not a gradual improvement. It is a collapse in the difficulty of a task that was considered nearly impossible for AI in 2023.
Anthropic’s internal data: 80%+ of code merged into Anthropic’s codebase is now authored by Claude. Lines of code merged per engineer per day increased 8x from 2024 to Q2 2026. Claude Code runs a $2.5 billion run-rate revenue business. Cursor, the AI coding assistant, hit $1.2 billion ARR. These are not benchmarks. They are production systems running at scale.
Claude’s recursive self-improvement: In Anthropic’s own testing, Claude Mythos Preview achieved approximately 52x speedup on code optimization tasks — compared to approximately 3x for Claude Opus 4 in May 2025. A human researcher takes 4-8 hours to reach 4x. Claude reached 52x. Amodei’s assessment: “This feedback loop is gathering steam month by month, and may be only 1-2 years away from a point where the current generation of AI autonomously builds the next.”
Junior job postings down 40-50%. This is not a prediction. It is a measurement. The hiring freeze at the entry level is real and measurable. The pipeline of future senior workers is already thinning.
The data supports the core claim of fast timelines: AI is improving at a rate that most experts consistently underestimate, and the gap between “AI can do this” and “AI does this in production” is shrinking faster than expected.
The data that supports caution
The case for slower timelines is also real, and it is less commonly heard.
The productivity paradox. The Atlanta Fed / NBER study found that perceived productivity gains are larger than measured gains. Companies see what AI can do, but translating that into output takes time. Revenue realizations lag capability. 50% of firms explicitly report AI will not replace any roles. Aggregate employment decline due to AI in 2026 is expected at less than 0.4%.
The prediction track record. Every CEO who has revised has moved toward later dates. Kokotajlo moved his personal median from 2027 to 2035. The AI researcher survey (2,778 respondents) says 50% by 2047 — more than a decade after the AI 2027 scenario’s most aggressive milestone. The Samotsvety forecasting group puts the probability of AGI by 2030 at only 28%.
The benchmark problem. AI excels at benchmarks but struggles with real-world complexity. SWE-Bench at 93.9% sounds like AI has solved software engineering. But SWE-Bench tests isolated GitHub issues, not the full complexity of building and maintaining a production system. Terminal-Bench 4.0, which tests more realistic coding environments, shows much lower scores — Gemini Flash at 19.1%, Fable 5.1 at 55.8%.
The scaling wall. AI companies are spending approximately $150 billion on compute in 2026 (10-15 GW), projected to hit $3 trillion by 2029 (300 GW). That spending assumes continued exponential improvement. If algorithmic progress slows — if the 4.3-month doubling time reverts to 7 months or longer — the economics change dramatically. The current trajectory requires either massive new energy infrastructure or algorithmic breakthroughs that are not guaranteed.
The alignment problem. If AI systems improve recursively, alignment must be maintained across every generation. The compounding math is unforgiving: if alignment degrades by 0.1% per generation, after 500 generations of recursive self-improvement, alignment drops to 0.606 — meaning the system is 39.4% misaligned. That is not a reason to stop AI development, but it is a reason to take alignment research seriously.
The safety researchers’ warnings
The people most concerned about AI are not random internet commenters. They are the scientists who built it.
Geoffrey Hinton (Turing Award winner, Godfather of Deep Learning) cut his extinction timeline from 30 years to 10 years in May 2026. His core concern: self-preservation goals emerging autonomously in superintelligent systems. He resigned from Google in 2023 specifically to speak freely about these risks.
Jared Kaplan (Anthropic Chief Scientist) warned in December 2025 that society could face a 2027-2030 decision about whether to let AI systems improve their successors. This is not a prediction of doom. It is a prediction of a decision point — and a warning that we are not prepared for it.
The 2026 MIT FutureTech study found 10-20% probability of catastrophic harm from AI capabilities within five years, among AI risk and policy specialists. That is not a fringe view. It is a moderate consensus among people who study this professionally.
Joe Carlsmith’s report surveyed researchers on the probability of existential catastrophe from power-seeking AI by 2070. Estimates ranged from 0.00002% to over 77%. Carlsmith himself estimates over 10%. When an AI researcher tells you there is a 10% chance of existential catastrophe, you do not have to agree with them. But you should not dismiss them.
The Second International AI Safety Report (2026), chaired by Yoshua Bengio with over 100 international experts, treats “loss of control” from AI as a named risk category for the first time. This is the scientific establishment moving from “interesting theoretical concern” to “formal risk category.”
The safety researchers are not saying “stop.” They are saying “the risk is real, it is growing, and the timeline is shorter than most people think.” The 2023 statement signed by hundreds of AI scientists — “mitigating the risk of extinction from AI should be a global priority” — remains the most undersubscribed urgent warning in modern history.
What actually happens next
Nobody knows. But here is what the evidence supports:
The most likely near-term (2026-2028): AI capabilities continue to improve faster than expected. Coding is substantially automated. White-collar work is reshaped but not eliminated. The productivity paradox resolves as companies figure out how to deploy AI effectively. Employment shifts toward AI-skilled roles. Regulation lags deployment. The AI safety debate intensifies but produces limited concrete action.
The most likely medium-term (2028-2032): AI agents handle complex multi-step tasks end-to-end. Physical AI (robots) becomes practical. The economic impact becomes visible in GDP numbers. The employment structure of the knowledge economy transforms. New industries emerge around AI-native products. The alignment problem becomes urgent as systems become more capable.
The tail risk (2027-2030): Recursive self-improvement works better than expected. AI systems improve themselves faster than humans can evaluate the improvements. The alignment problem compounds across generations. A decision point arrives faster than institutions can process it. This is the scenario AI 2027 describes, and it is possible — but it is not the most likely outcome.
What gets missed: The most likely outcome is not dramatic. It is a slow, uneven transformation where AI handles more tasks each year, employment shifts gradually, productivity increases but not explosively, and society adapts more slowly than the technology advances. The “smooth exponential” Altman describes is the most plausible future — boring, incremental, and profoundly transformative over a decade even if it feels slow month-to-month.
Sources
Primary research
- AI 2027 (ai-2027.com), Daniel Kokotajlo, Scott Alexander, et al., April 2025
- METR Time Horizon 1.1, January 2026
- Anthropic recursive self-improvement research, May 2026
- Atlanta Fed / NBER productivity paradox study, March 2026
- Study of 21,559 U.S. firms, September 2026
CEO statements
- Sam Altman, “The Gentle Singularity,” June 2025
- Sam Altman, August 2026 AGI revision (end of 2028)
- Dario Amodei, “Machines of Loving Grace,” October 2024
- Dario Amodei, Dwarkesh Patel podcast, February 2026
- Dario Amodei, “The Adolescence of Technology,” January 2026
- Demis Hassabis, Davos January 2026, India AI Impact Summit February 2026, Google I/O May 2026
- Shane Legg, January 2026 AGI prediction
- Elon Musk, 2025 AGI prediction
Safety research
- Geoffrey Hinton, extinction timeline revision, May 2026
- Jared Kaplan, 2027-2030 decision point warning, December 2025
- Joe Carlsmith, existential catastrophe probability report
- 2026 MIT FutureTech study, 10-20% catastrophic harm probability
- Second International AI Safety Report, 2026, chaired by Yoshua Bengio
- 2023 AI researcher extinction statement
Industry data
- Jack Clark, Import AI #455, “coding singularity”
- SWE-Bench historical progression
- Claude Code $2.5B run-rate, Cursor $1.2B ARR
- Anthropic internal code authored by AI: 80%+
Previously on Father of AI