Artificial superintelligence might be just two years away from reality, according to a new model created by a team of AI researchers and forecasters with backgrounds at organisations such as OpenAI and The Center for AI Policy.

The newly published AI 2027 scenario predicts that within the next two years, tech firms will develop expert-human-level AI models capable of automating AI research, quickly leading to vastly superhuman AIs.

Given the tight timeframes involved, this might sound like the stuff of science fiction, however, the scenario is the first major release from the newly minted AI Futures Project, a nonprofit forecasting the future of AI.

Leading the project is Daniel Kokotajlo, previously a researcher at OpenAI on scenario planning and writer of What 2026 Looks Like, a near-term AI scenario forecast, along with Thomas Larsen, founder of the Center for AI Policy and member of the Machine Intelligence Research Institute.

Combining their own industry experience with expert feedback and scenario planning exercises, the AI Future’s team have developed a detailed two to three-year outline for what the near future of AI will likely hold, including specific technical milestones.

Just one year from now, and assuming no increases in training compute, the scenario predicts that a ‘superhuman coder’ will be created, an AI system that can do the same coding tasks as the best human AI engineer can, but much faster and a lot cheaper.

According to AI Future’s forecast, this will lead to the development of ‘superhuman’ research models by the middle of 2027, which, combined with the coding abilities of AI, will vastly accelerate the path to superintelligence.

Between December 2027 and the end of Q1 2028, the timeline leads to the creation of an artificial superintelligence (ASI), a system that is vastly better than the best human at every cognitive task.

Like every prediction made about the future of technology, the AI Future’s forecast is based on assumptions that not everyone in the AI field agrees with, however, the project provides an exhaustive methodology regarding how these assumptions have been made, including factors such as the cash being spent on AI projects, the trajectory of compute power, research survey data, and previously run simulations.

Added to that, Daniel Kokotajlo – acting as the Project’s executive director – has form in this area, having previously predicted the current AI landscape.

In 2021, he forecast that by 2024, AI companies would be investing more in fine-tuning their models and adding different variations (think ChatGPT 4.o mini or Meta’s Llama 4 Maverick and Scout), than training new or bigger ones.

Recommended reading

• New Nonprofit PRISM Launched To Investigate AI Sentience
• OpenAI Co-Founder’s Safety-focused AI Startup Raises $1 Billion
• UK Shifts AI Focus to Prioritise Crime and Defence

Kokotajlo’s soothsaying also ran to predicting that AI development would be hindered by chip shortages, already being reported, that the frenzy surrounding the tech would begin to fade as early expectations fail to materialise, and that rather than pushing the boundaries of the cutting-edge AI models, chatbots would instead become ubiquitous.

At the heart of AI Future’s efforts is the question of safety. On leaving OpenAI, Kokotajlo called for better transparency and safety practices around AI development, with the 2027 timeline echoing those concerns.

Through the insights gained from its research and wargaming, the project warns that when ASIs begin to be widely deployed, they might develop unintended, adversarial, ‘misaligned’ goals, leading to human disempowerment.

If this new scenario is to be taken seriously, 2027 may be the tipping point that determines humanity’s future.