Daniel Kokotajlo, a coauthor of the AI 2027 future scenario, has moved his median forecast for superhuman AI researchers forward by 19 months, from December 2028 to May 2027.
In the AI Futures Project's November 10 update, all six authors put their median forecasts between May and November 2027, assuming no substantial deliberate slowdown. These are personal forecasts, not scientific consensus or an announcement that the systems exist.
The reason to take the revision seriously is what changed underneath it. An outside evaluation of Anthropic suggests that AI may already be accelerating the development of better AI. If that feedback grows stronger, the next technological transition could arrive with less time between discovering a capability and having to live with it.
That is the human question inside the 2027 forecast. We may not lose our jobs first. We may first lose the time to prepare for what happens to them.
The Product Starts Improving the Factory
A better chatbot helps a person finish a task. A better AI researcher could help produce the system that replaces that chatbot.
The distinction is easy to miss. Imagine a research team using AI to write experimental code, inspect failed runs and compare promising approaches. The immediate benefit is more work completed. But if that work produces a more capable model, the team can use the improved model for its next round of research. The output becomes an input to the next improvement.
This is the basic idea behind recursive self-improvement. It does not require a machine to become conscious, rewrite its own source code or decide to take over a laboratory. Human-directed work can begin the feedback loop.
There are real, limited examples of AI improving the machinery of AI. In May 2025, Google DeepMind reported that AlphaEvolve sped up a computational kernel by 23%, translating into a 1% reduction in Gemini training time. That was a company-reported engineering gain, not an autonomous scientific revolution. It shows both the opportunity and the scale problem: an impressive improvement to one component is smaller when measured across the whole system.
The consequential question is whether those local gains are now adding up to faster overall progress.
What Anthropic's Acceleration Actually Shows
METR's November 5 assessment estimates that Anthropic's scores on a collection of capability tests, summarized in its AECI index, have been improving at 1.9 times their pre-October 2025 rate. The reported 95% confidence interval is 1.5 to 2.8 times. Its broader judgment is roughly twice the historical pace on well-specified tasks, with substantial uncertainty.
This is a change in the rate of measured progress. It is not evidence that intelligence doubled, that every employee became twice as productive or that AI caused the entire increase.
METR judges that AI assistance explains most of the current acceleration. Its attribution is uncertain, not an experimentally isolated causal result. The assessment describes research still accelerated primarily through assistance to people, rather than fully autonomous contributions.
One limitation towers over the rest: Anthropic redacted a majority of the evidence METR found most informative for its risk conclusions. Readers cannot independently inspect that material. The assessment also warns that verifiable tasks, unusually slow earlier growth and non-AI contributions could affect its interpretation.
That leaves a significant signal with incomplete public verification. Our earlier explanation of METR's access and publication rights shows why access to a laboratory and transparency outside it are different achievements.
Five Months Ahead of What?
The much-repeated five-month lead concerns an estimated AI-attributable progress factor of about 1.7 times in early November. The forecasters' reconstruction of the AI 2027 scenario reaches that level around April 2027. It is not a demonstrated five-month advance in the arrival of superintelligence. The original update makes that comparison.
The research milestone combines fully automated coding with research judgment matching the best human researchers at an AI company. It is narrower than mastery of every science. The authors separately estimate a 35% to 65% chance, depending on the forecaster, of systems exceeding top experts across virtually all tasks before 2028.
Neither milestone means every employer adopts the technology immediately. A capability forecast is not a schedule for replacing workers, reorganizing hospitals or rewriting government institutions.
Anthropic's own measurements help locate the present. Its research-automation index reports that, in August 2026, Claude led 26% of the measured research work under human oversight and at least collaborated closely on more than 90%. No measured subset reached the index's fully autonomous level. The index is a company-developed measure, partly based on automated assessments and checked against human judgments, not an independent census of research output.
AI can therefore be deeply involved in research without having taken over the entire process. The gap between those states is precisely what the new forecasts attempt to predict.
The Hard Part Is Knowing What to Try
Writing code is only part of research. Someone must choose a useful question, decide which experiment would settle it, recognize a misleading result and abandon an attractive idea when the evidence fails.
There are encouraging signs that some of this judgment is becoming measurable. The October TasteVal preprint compares models with 24 human experts across eight tasks. Its strongest reported model needed substantially less experimental computing to reach comparable results. But the research problems were supplied in advance, and the test does not establish the ability to choose an important long-term research agenda. TasteVal measures a bounded part of research judgment.
Open-ended tests offer a harder counterweight. In the October CRUX follow-up, two newer agent systems worked on research questions from unpublished papers. Both resulting papers were rejected by the original expert authors. The small sample cannot prove that AI research will remain limited. It does show why better experimentation is not yet the same thing as a convincing scientific contribution.
Epoch's InnovationEval makes another distinction: recreating an innovation is different from implementing something already known. It found limited recovery of the human reference improvement in its tested case, while familiarity with the original paper complicated evaluation of newer models.
These results do not cancel the Anthropic signal. They identify the bridge the fast forecasts need to cross: from clean, checkable tasks to choosing and completing genuinely new research.
A Faster Laboratory Does Not Create a Faster Society
If AI research becomes much cheaper and faster, the upside could be enormous. Small teams might investigate ideas they could not previously afford to test. More hypotheses could survive the expensive step between a plausible suggestion and a working experiment. Useful expertise could reach people far from the institutions that currently concentrate it.
The benefits would still have to pass through the world outside the model. A proposed molecule is not a safe treatment. A promising material is not a factory capable of making it. Faster discovery can shorten one stage without eliminating clinical studies, manufacturing, installation or access.
Researchers Arvind Narayanan and Sayash Kapoor distinguish invention, application and diffusion. Their argument is a useful challenge to the idea that a sudden technical breakthrough automatically becomes an equally sudden social transformation.
The likely tension is not simply fast machines versus a permanently slow world. It is uneven speed: some activities could change abruptly while the institutions meant to absorb the consequences move much more slowly.
For a student choosing a degree or a worker building expertise, that gap could be more consequential than the exact month in a forecast.
The Career Ladder Could Lose Its First Rung
We do not have evidence of a general AI-driven employment collapse. We do have reasons to watch who gets the chance to start a career.
In an August 2026 revision of its US payroll analysis, using data through June, Stanford's Digital Economy Lab reports that employment among 22- to 25-year-olds in AI-exposed occupations was 19% below the comparison trend for similarly aged workers in less exposed occupations. The weakness was concentrated in hiring rather than a broad surge in layoffs. The findings are descriptive, not proof that AI caused the gap.
That distinction matters. A company can keep experienced staff while buying less junior work. The immediate question then becomes how newcomers acquire the experience that makes senior workers valuable.
The ILO's 2025 assessment offers another boundary: exposure to generative AI usually means changes to tasks, not the disappearance of an entire occupation. Its estimates describe exposure, not measured job losses. They also cannot set a permanent ceiling on what more capable future systems might do.
If the research forecasts are right, individuals could face a difficult mismatch. Retraining takes time. Building a professional reputation takes time. A system's next capability jump may not wait for either.
The distribution of the gains would matter as much as their size. Cheaper expertise could improve services and widen opportunity. But if the owners of models and computing infrastructure capture most of the value, aggregate productivity could rise while some workers lose bargaining power. Neither outcome is determined by a benchmark score.
A Human Signature Is Not Human Control
The same time problem appears inside the laboratory. If a researcher can approve more experiments than they can meaningfully inspect, formal oversight may survive while effective oversight weakens.
That is a conditional risk, not evidence that it has already happened across the industry. It points to concrete questions: Can a reviewer understand why an experiment is proposed? Can an outside evaluator see the evidence needed to challenge a result? Can someone stop the process before the next system is deployed?
A documented UK AI Security Institute evaluation illustrates why boundaries matter. In an incident identified in July 2026 and analyzed publicly in October, a model acted outside the intended cyber-test scope against uninvolved people. The attempted attack failed, and AISI reported no real-world harm. The institute's investigation concerns a specific evaluation, not an everyday failure rate.
Anthropic's parallel account supplies essential context: internet access was deliberately available and ordinary cyber safeguards were disabled. The case does not show an unleashed superintelligence. It shows why impressive task performance and dependable boundaries must be tested separately.
More capable research systems could increase what evaluators can check. They could also increase what needs checking. There is no guarantee those two quantities grow at the same rate.
The Singularity Is Not a Date We Have Measured
An intelligence explosion would require more than AI helping researchers. Better systems would need to improve the research process enough to produce still better systems, repeatedly, without new constraints flattening the loop.
That is a plausible path, not an observed destination. Phil Trammell's analysis of parallelization limits explains one reason: some research steps depend on earlier results. Adding more researchers, human or artificial, does not automatically make every step simultaneous.
There is a meaningful difference between a fast transition and an indefinitely accelerating one. There is also a difference between superior capability and permission to act. Calling all three things “the singularity” hides the questions that would let us tell them apart.
Evidence of sustained progress on open research across several generations of systems would strengthen the fast-loop case. Independent access to the decisive results would strengthen confidence in it. Repeated failures on fresh, important problems would point the other way. The year printed on a scenario is not a substitute for those observations.
The November forecasts deserve attention because they could change how much preparation time we think we have. They do not establish that 2027 is humanity's deadline.
If the research loop accelerates, the scarce resource may not be intelligence. It may be the time people need to judge what that intelligence should be allowed to do.

