Will AI soon surpass all human experts? Discover how new AI models are rapidly advancing on Humanity’s Last Exam, a test designed to push the limits of knowledge.
Artificial intelligence may soon surpass the collective knowledge of human experts, according to predictions from researchers actively involved in its development. These specialists believe that within a year, AI systems could reach a level where they perform better than even the most knowledgeable professionals in various fields.
Tough Benchmark
A major indicator of this progress is Humanity’s Last Exam (HLE), a challenging test created by top technology leaders. The test contains 2,500 carefully chosen questions covering around 100 topics, including rocket science, mythology, and human biology.
Each question requires the level of understanding expected from a PhD candidate, and scoring near-perfectly would effectively make someone a “universal expert.”
Rapid Progress
Just two years ago, early AI systems like ChatGPT, developed by OpenAI, scored only 3 per cent on this test, with competitors from Google and Anthropic performing at a similar level. At that time, researchers noted a clear gap between AI and top human academics.
However, progress has happened swiftly. Google’s Gemini model recently scored 45.9 per cent, a significant increase from 18.8 per cent in a previous attempt, while Anthropic’s Claude has reached 34.2 per cent and is still improving.
Also read: AI Jobs 2026: Will Your 9-to-5 Job Survive the Rise of AI Agents in the Workplace?
Expert Views
Calvin Zhang, research lead at Scale, the company behind the exam, explained that the test was meant to test the limits of human expertise. He highlighted the impressive improvements in AI’s reasoning abilities over recent years.
Kate Olszewska, product manager at Google DeepMind, suggested that if more focus is given to this goal, achieving a perfect score may happen faster than expected.
Global Effort
The exam was developed by Scale in collaboration with the Center for AI Safety and involves contributions from experts in about 50 countries. Around 70,000 questions were submitted following a worldwide call in 2024 offering a $500,000 prize.
After removing easier questions, the list was significantly reduced, keeping only the most difficult ones. Some questions remain confidential to prevent AI systems from learning answers through public discussions.
Beyond Limitsd
If AI eventually achieves a perfect score, researchers say future exams will need to go beyond what is known to humans. This milestone would be similar to past breakthroughs, such as IBM’s Deep Blue defeating chess champion Garry Kasparov in 1997.
Even as AI continues to excel in traditional benchmarks, developers are now focusing on expanding knowledge itself.
Nevertheless, Zhang emphasized that human skills such as creativity, judgment, and hands-on expertise in areas like surgery are likely to remain challenging for AI to replicate.
Source: Daily Mail – Science & Tech
Also read: World’s Smallest QR Code Could Store Data for Centuries