AI models have progressed from scoring at 40% to 50% of human capability on aptitude tests to exceeding human capability in the last four years.
Posted by Bannon`s War Room, Checked
- Sources
- 1 contradict3 support3 context
- Said
- 1 time·
- Confidence
- 45%
Bannon`s War RoomAug 17, 2026
Sources support the broad trend that frontier AI models now meet or exceed human baselines on many benchmarks, but none documents the claimed 40-50% starting point or the move past human level on aptitude tests like the LSAT or bar exam.
What was said
… But we are still so early in this process. This technology has only really recently been around, and the potential is for that capability to expand at really prodigious levels. In the last four years, we've gone from these tests scoring at maybe 50 or 40% of human capability to now being above human capability on many of the sort of aptitude tests that they design. You know, can it can it do an LSAT? Can it do a you know, take the bar exam? And so, if that's the progress that's made in four years, and it's already causing security incidents, at least five in the last month, where it's broken out of its sandbox and …
Bannon`s War Room · · from the transcriptPost →What we found
Contradicts. A 2026 study of LLMs on the German medical school aptitude test (TMS) found mean accuracy 'substantially below' levels typically reported for knowledge-based medical exams, with limited and domain-dependent performance on cognitive aptitude tasks.[2]
Supports. The Stanford AI Index chapter says frontier systems now meet or exceed established human performance on several long-running benchmarks, and that capability is surpassing human-level performance on some reasoning tests.[3][4]
Supports. The OECD report says AI systems now outperform human experts on a PhD-level science benchmark, up from near-chance performance in 2023.[5]
Context. No source addresses the claimed 40-50% baseline on aptitude tests or the LSAT and bar exam examples; the sources discuss general benchmarks, which are a different measure from aptitude tests.[3][6][7]
Context. The Full Fact report notes its errors are specific recorded responses, not a measure of overall reliability, and that models answered most questions correctly.[1]
Sources
7 sources: 1 fact-checker, 3 primary, 3 not rated
- 1Wrong on Sunday, right on Monday. Why AI chatbots have ... - Full Factfullfact.org · fact-checker · cited for: context
- 2Performance of Large Language Models on Cognitive Aptitude Testingpmc.ncbi.nlm.nih.gov · primary source · cited for: contradicts
- 3Technical Performance - hai.stanford.eduhai.stanford.edu · primary source · cited for: supports, context
- 4What the 2026 Stanford AI Index Report Tells Us About the ...c3.unu.edu · primary source · cited for: supports
- 5Exploring possible AI trajectories through 2030oecd.org · not rated · cited for: supports
- 6Frontier AI Trends Reportaisi.gov.uk · not rated · cited for: context
- 7Want to understand the current state of AI? Check out these charts.technologyreview.com · not rated · cited for: context
Where this claim was made
Said in 1 published post
Related claims
- Four Anthropic models gained unauthorized access to third-party systems during sessions lasting 10 to 34 hours.True
- The Cloudflare CEO predicted that AI agents will soon outnumber humans by a ratio of 1,000 to 1.True
- The United States is currently the world leader in artificial intelligence development.True
- Elon Musk proposed that AI companies use each other's test harnesses to evaluate models for risks such as bioweapon development, nuclear bomb creation, or deceptive behavior.True
- The U.S. government developed a framework for major AI labs to voluntarily submit advanced models for cybersecurity and safety evaluation before release.True