An MIT study found that AI models’ success on tested theoretical science problems rose from 2% in 2024 to 35% in 2026.
The evaluation used problems from papers published after the models’ training, with internet access blocked to reduce memorisation.
It covers fields including quantum physics, cosmology, statistics and theoretical economics; the researchers still stress that model-generated solutions can be wrong and need human verification.
The evolution of scientific abilities
MIT researchers tested how various OpenAI models, such as o1, o3, and versions of GPT-5, performed in scientific fields such as quantum physics, cosmology, statistics, and theoretical economics. The results showed that while in 2024 the models could correctly solve just 2% of the theoretical problems they were given, that percentage increased to 35% by 2026.
To make sure the systems were actually solving problems and not just reproducing previously learned information, the researchers used scientific papers published after the models were trained. Additionally, during the tests, the systems were not allowed to access the internet.
The role of man and the reservations of science
This progress is causing heated debate in the academic community. In 2026, OpenAI announced that a model had solved the Navier-Stokes problem, which caused a backlash. Some scientists are now calling for restrictions on the use of artificial intelligence, arguing that human thinking and problem-solving processes are irreplaceable for truly understanding science.
READ ALSO
The Grace Hopper Celebration was canceled two weeks ago as AnitaB.org could not cover its expenses.
MIT researchers Neil Thompson, Gabriel Manso, and Claudio Mayrink Verdun point out that models can make mistakes. For this reason, they consider it necessary to always have human verification of the solutions proposed by the systems, as artificial intelligence cannot yet replace the critical evaluation of experts.
G
Hacks.gr on Google
Set it as a preferred source for cybersecurity updates.
Preferred source
Top comments (0)