The AI researcher says a small group with access to advanced models is witnessing capabilities that are increasingly difficult for humans to review and evaluate.
Artificial intelligence researcher and OpenAI co-founder Andrej Karpathy has highlighted a widening gap in global AI adoption, estimating that nearly 75% of the world's population has had little exposure to large language models (LLMs), even as a small group of researchers gains access to increasingly advanced systems.
In a post on X, Karpathy estimated that approximately six billion people worldwide have barely interacted with AI tools. At the other end of the spectrum, around 5,000 individuals working at leading AI laboratories have access to frontier models with capabilities that remain largely unavailable to the wider public.
According to a report by The Times of India, Karpathy described this distribution as a narrowing progression of access and capabilities, with significant differences between casual users, professional developers and researchers working directly with advanced models.
He estimated that between one billion and two billion people use free AI applications for activities such as information searches, writing assistance and everyday queries.
These users typically interact with publicly available chatbots, where the technology functions as an alternative to conventional search engines or productivity tools.
A smaller group of approximately 20 million people, representing around 0.2% of the global population, uses advanced AI models professionally for software development, mathematics and other technical applications.
Karpathy noted that developers in this category are increasingly completing tasks that previously required weeks or months, including building applications and analysing complex software systems.
The capabilities available to researchers within leading AI laboratories extend further. According to Karpathy, advanced systems can coordinate thousands of AI agents over extended periods to undertake large-scale software projects, identify previously unknown security vulnerabilities and assist with mathematical research.
However, the growing complexity and volume of AI-generated work have introduced challenges for human oversight.
"Human review and comprehension are starting to fall behind," Karpathy wrote, referring to the difficulty of evaluating outputs produced by increasingly capable systems.
He identified three factors contributing to the differences in AI adoption and perception: the complexity of tasks assigned to models, variations in their performance across different domains and unequal access to advanced capabilities.
Karpathy observed that AI systems generally perform more effectively in areas where results can be objectively verified, such as programming, compared with tasks that require more subjective evaluation.
Access restrictions also influence how users assess AI capabilities, with free applications, paid subscriptions and internal research systems offering different levels of functionality.
The researcher, who previously worked at Tesla and OpenAI, joined Anthropic in May 2026, where he is involved in research related to model pretraining.
His observations come as AI developers expand the capabilities of autonomous agents and organisations explore their use across software engineering, research and enterprise workflows.
The differences highlighted by Karpathy indicate that public experiences of AI may vary substantially depending on the tools available, the tasks being performed and the level of technical access.
Disclaimer: This article may include information derived from interviews, press releases, public statements, research, company communications and other publicly available or third-party sources. Such material may be summarised, paraphrased or contextualised for journalistic and editorial purposes. All rights in third-party content remain with their respective owners.