Andrej Karpathy, the former Tesla AI director and OpenAI cofounder, has identified a critical gap in how people understand what artificial intelligence can actually do. His observation about the difference between free-tier and frontier AI models has become a touchstone for technologists trying to make sense of conflicting claims about AI's capabilities.

Why Are People So Confused About What AI Can Actually Do?

The confusion stems from a simple but consequential fact: different versions of the same AI tool perform dramatically differently. Karpathy noted that many people tested free or older AI models, encountered hallucinations and fumbled responses, and concluded that AI is fundamentally unreliable. Meanwhile, others using paid frontier models watched those systems restructure entire codebases or conduct hour-long autonomous research loops.

"It really is simultaneously the case that OpenAI's free and I think slightly orphaned Advanced Voice Mode will fumble the dumbest questions in your Instagram's reels and at the same time, OpenAI's highest-tier and paid Codex model will go off for 1 hour to coherently restructure an entire code base," Karpathy stated.

Andrej Karpathy, Founding Member of OpenAI

Both groups are describing reality accurately. They are simply describing different products. This distinction matters because it creates two symmetrical errors in how people evaluate AI. Some dismiss AI entirely after testing limited versions, missing genuine capability gains. Others trust frontier models too much and stop applying independent verification to their outputs.

How Is This Capability Gap Affecting AI Adoption and Strategy?

The gap Karpathy identified has real consequences for how businesses, researchers, and entrepreneurs approach AI tools. People who distrust AI entirely stop using it as a research layer, potentially missing productivity improvements. People who trust AI too much treat retrieval output as verified research output, which can lead to accepting unverified claims as fact.

This matters especially in high-stakes domains. The World Economic Forum's Global Risks Report 2026, drawing on responses from over 1,400 expert respondents, ranked AI-driven misinformation as the single largest short-term global risk. The mechanism is structural: AI systems retrieve indexed content matching a query but do not verify whether that content is accurate, who published it, or when a domain was registered.

Karpathy's concept of "jaggedness" applies directly here. AI models perform with genius-level capability in verifiable domains, such as code restructuring or mathematical problem-solving, while performing with what he describes as child-level naivety in unverifiable domains, such as reputation claims or medical assertions.

Steps to Evaluate AI Claims and Avoid Misinformation

• Verify Source Authority: Check whether a named accuser exists, when a domain was registered, and whether any institutional body has validated the claim before treating AI output as fact.
• Distinguish Between Retrieval and Research: Understand that when an AI writes "multiple sources suggest X," it has retrieved multiple indexed pages containing the word X, not spoken to those sources the way a journalist would.
• Apply Independent Verification: Use a structured verification framework for any AI-surfaced information in high-stakes domains, including business due diligence, medical claims, financial assertions, and reputation allegations.

The framework for evaluating claims has become more critical as AI tools proliferate. A fabricated allegation and a sourced rebuttal look identical to a retrieval system: both are indexed text containing the same keywords. The AI presents them side by side as "mixed reports" and calls it balance, without checking whether either claim is accurate.

Karpathy's influence extends beyond his direct commentary. He has become a trusted voice for technical founders seeking deep understanding of how AI systems actually work. His YouTube channel, which focuses on neural networks, large language models, and AI engineering fundamentals, is listed among the top resources for technical founders seeking advanced AI knowledge.

The practical implication is clear: in 2026, the winners will not be those who simply understand AI, but those who apply it strategically while maintaining healthy skepticism about its limitations. Karpathy's distinction between capability tiers and verification gaps provides a framework for that balanced approach. His recent recognition as a thought leader in this space, including receiving Nvidia's first DGX Station hand-delivered by CEO Jensen Huang for his OpenClaw agent work, reflects his continued influence on how the industry thinks about AI development and deployment.