Shark Tank star Mark Cuban says AI can’t be trusted without humans

Billionaire investor Mark Cuban warned the corporate world that artificial intelligence's greatest weakness isn't a lack of power; it's a fundamental inability to provide consistent answers.

On their X account, the famous personality from Shark Tank provided reasons why this constraint makes domain expertise more crucial than ever before and dismisses fears about artificial intelligence becoming conscious and conquering the planet.

While regular software functions according to the laws of logic, large language models (LLMs) operate under probabilistic rules. They basically predict the following word with the help of statistics, which leads to varying outcomes even with the same inputs.

"I'm coming to the conclusion that the biggest challenge for enterprise AI, and AI in general, as of now, is that it's still impossible to make sure that everyone gets the same answer to the same question every time," Cuban noted. This inconsistency creates what he calls a "massive liability" for organisations dependent on predictable outcomes.

By using this drawback, Cuban counteracts AI doomers concerned about AI achieving consciousness and ruling humankind. "AI doesn't understand that its output should be consistent, it certainly doesn't understand the real-world consequences of what it's saying. Judgement and the ability to challenge AI output are becoming increasingly necessary and valuable," he argued.

The inconsistency of output demonstrates that there is a lack of awareness of consequences among AI, something that the autonomous superintelligence feared by the critics cannot function without. "Which makes domain knowledge more valuable by the second," Cuban added.

This dependence on humans was confirmed by Google and Uber's CEOs on earnings calls. Sundar Pichai of Google stated that 75% of newly generated code was written by AI, although all of it was reviewed and approved by engineers. Uber's Dara Khosrowshahi mentioned that only 10% of newly generated code was written by AI yet still had to be approved by engineers.