AI Gives Predictable Answers, But With Unpredictable Results

Associate Professor Michael Pound: LLMs Are Probabilistic - and That Poses Risks Anna Delaney (annamadeline) • June 19, 2025

Generative artificial intelligence doesn't think. It predicts. Each output is the next most likely word, not a reasoned decision. Michael Pound, associate professor at the University of Nottingham, warns that when large language models generate malicious content, it's not due to confusion or logic flaws - it's simply probability at work.

See Also: Enterprise Browser Transforms App Delivery and Compliance

Pound said this core misunderstanding has led to poor implementation choices. Developers often assume repeatable outputs, but the systems behave differently each time.

"It's got no concept of true or false or right or wrong," Pound said. "It just says, 'I think these are some likely words.'"

In this video interview with Information Security Media Group at Infosecurity Europe 2025, Pound also discussed:

• Applying strict input controls and output filtering ...