Human-level AI by 2027? Inside a former OpenAI researcher’s forecast

Who wrote this, and why listen?

Aschenbrenner worked on OpenAI’s Superalignment team, which addressed the problem of controlling highly capable AI. Earlier, he researched long-run economic growth at Oxford. That experience offers an informed view of the industry, but does not by itself make a forecast correct.

There is also a financial context: in 2024 he founded Situational Awareness LP, an investment firm focused on AI development and its effects on businesses. Its website states this openly. This is therefore also an investor’s perspective. The arguments deserve scrutiny independently of how persuasive their author sounds.

Leopold Aschenbrenner’s biography ↗

Situational Awareness LP — investment activities ↗

What does ‘human-level AI’ mean?

AGI stands for artificial general intelligence. For this discussion, imagine a system able to learn different intellectual tasks and independently carry them through to a result. This concerns capabilities, not evidence of consciousness, feelings or a human personality.

Consider a game. One assistant writes a jump function. A much more autonomous system could be asked to investigate why jumping feels awkward, propose changes, check the phone controls and avoid breaking saved progress. This is our illustrative example: producing a good individual answer and taking responsibility for the whole result involve different levels of difficulty.

Why does the forecast point to 2027?

Aschenbrenner combined three factors: more computing resources, more efficient algorithms, and the move from chatbots to agents with tools. The last factor includes access to a working environment and the ability to perform a sequence of actions. He extrapolated earlier trends and described an AI researcher by 2027 as plausible.

His text frequently uses OOM, or order of magnitude: a tenfold change. It is a way of describing resource scale, not a multiplier for human intelligence. More computation cannot be translated literally into being the same number of times smarter.

Our main caution about this forecasting approach is that a line on a graph hides the conditions needed to keep it going. Useful data, funding, hardware and training methods that actually produce new capabilities are all required. Even a correct prediction of direction does not guarantee the correct calendar year.

From GPT-4 to AGI — reasoning behind the forecast ↗

The key idea: AI helps build the next AI

The next step in the scenario is automating AI research itself. If a system can propose and test improvements to models, many copies could work on them simultaneously. Better models help produce still better models. This is the author’s explanation for a possible rapid acceleration of progress.

The essay itself acknowledges constraints: experiments require resources, and some work is difficult to automate. Our conclusion is that a million generated hypotheses are not a million discoveries. What matters is how many survive testing and improve a real system.

Imagine an animation studio: extra artists can speed up frame production, while conflicting direction or slow approvals may remain a problem. The analogy does not prove how AI will develop. It shows why speeding up one part of a process does not automatically speed up the entire process by the same amount.

From AGI to Superintelligence — the acceleration mechanism ↗

What independent tests actually say about autonomy

METR measures the difficulty of tasks that AI completes at a given success probability. Its ‘time horizon’ refers to the time a human expert would need, not how long AI works independently without interruption. A 50% horizon means success roughly half the time. These are mainly well-defined software tasks. The measurement page was last updated on 8 September 2026 and is no longer actively maintained.

In an explanatory note, METR researcher Thomas Kwa cautions against applying the results to every profession: the estimates are imprecise and capabilities vary substantially between domains. The note underwent less review than METR’s main research posts. It is useful specifically as an explanation of the metric’s limitations.

For your own project, the practical rule is to evaluate the finished result. If an assistant writes code quickly but you then spend a long time fixing it, generation time alone is a poor measure. Compare total time, quality and the cost of errors on equivalent tasks.

METR — time-horizon measurements and methodology ↗

METR: Thomas Kwa on the metric’s limitations, January 2026 ↗

Even ‘AI makes developers faster’ needs context

In February 2026, METR reported problems with its new developer-productivity experiment: some people did not want to work without AI, and using several agents simultaneously complicated time measurement. The authors suspected newer tools were more helpful but acknowledged weak quantitative evidence for the size of that benefit. Equally, an earlier finding that experienced developers were slowed down in 2025 should not be treated as a final verdict on every current tool.

This is a useful lesson in reading news: ‘a study found’ should always be followed by who was studied, on which tasks, with which models and when. Without that context, it is easy to exaggerate either AI’s usefulness or its inability to help.

METR — productivity experiment update, February 2026 ↗

AI’s future also depends on power plants

The IEA’s 2026 report estimates global data-centre electricity consumption at 485 TWh in 2025 and projects around 950 TWh by 2030. These figures cover all data centres, not just AI. The agency also describes constraints involving grids, equipment and chip supplies. Projected infrastructure growth does not establish an AGI arrival date.

This is why we find the industrial side of the essays important. Behind a convenient chat window lies a physical world of construction, cooling, grid connections and operating costs. Adding another line to a development plan is much easier than bringing a large facility online on schedule.

IEA — Key Questions on Energy and AI, 2026 ↗

Why the author moves from technology to power and safety

Aschenbrenner predicts substantial US government involvement in an AGI project around 2027–2028 and views the technology through US–China competition. This is his geopolitical scenario and position. The essay is not an official announcement of such a project and does not establish that conflict is inevitable.

In another chapter, he asks how to oversee a system whose decisions exceed the overseer’s understanding. That is a more difficult problem than making a chatbot polite. Readers can distinguish the technical risk from the author’s political prescription: agreeing that oversight matters does not require accepting every prediction about governments and the future.

The Project — the author’s government-involvement scenario ↗

Superalignment — oversight of highly capable AI ↗

Three possible paths: our conclusions

Rapid acceleration. Agents become reliable enough to handle much of the research cycle, potentially bringing successive generations of systems closer together. The useful signals would be reproducible improvements and independent checks, not just impressive demonstrations.

Uneven progress. AI becomes very strong at parts of the work but still needs people where requirements are unclear, context is extensive or mistakes are costly. Team composition and task allocation change even while fully autonomous workers remain out of reach in many domains.

Slower adoption. Models improve, but costs, unreliability or infrastructure constrain practical use. Benefits accumulate through specific tools and processes. These three scenarios are our editorial framework for observation, not calculated probabilities or a timetable of promised events.

What developers, designers and game creators can do

Choose one recurring part of your work and test AI on it: finding a bug, exploring interface alternatives, preparing tests or drafting a scene. Record the time spent specifying, correcting and checking the task. That will inform your decisions better than arguing about the date of AGI.

Develop the ability to define a good result. In a game, that includes understandable controls, an engaging mechanic and the player’s experience, alongside working code. In animation, it includes expressive movement, character consistency and artistic direction. This is not a guarantee that any profession will remain unchanged forever; it is a practical way to navigate changing tools.

The central value of Situational Awareness is its question: what happens when AI starts accelerating AI development itself? It deserves serious attention, with each link in the argument checked separately. Preparation can begin now: use the tools, measure outcomes and retain control over consequential decisions. None of that requires certainty about 2027.

Quick answers

Primary sources

• Situational Awareness — original essay series, June 2024
• Leopold Aschenbrenner’s biography
• Situational Awareness LP — investment activities
• From GPT-4 to AGI — reasoning behind the forecast
• From AGI to Superintelligence — the acceleration mechanism
• METR — time-horizon measurements and methodology
• METR: Thomas Kwa on the metric’s limitations, January 2026
• METR — productivity experiment update, February 2026
• IEA — Key Questions on Energy and AI, 2026
• The Project — the author’s government-involvement scenario
• Superalignment — oversight of highly capable AI