S11E01 – The Show Returns With AI Limits and Scams

Episode Summary

Mike returned to the radio after about five years away and used the first show to explain what AI is and is not. He said the show came back partly because more people have been falling for scams, many of them now AI-powered, such as cloned voices on phone calls and convincing sign-in emails.

Much of the hour was a plain-language explanation of how chatbots work. Mike covered why they are poor at math, why they sound most confident when they are wrong, and how a limited “context window” leads them to make things up. He also talked about the stories claiming AIs try to escape or blackmail people, and why he thinks they are hard to interpret.

He closed with a caller question about AI-made movie trailers and a thank-you to the station and listeners.

Topics Covered

AI-powered scams and the family safe word

Mike said AI makes scams more convincing, including cloned voices asking for money and emails that look like they come from a boss or a known service. He told of a retirement-home resident who lost $80,000 to a phone scam. His rule is to treat everything you receive as fake until proven true. He recommended a family safe word or phrase that has never been put online, to confirm a caller really is who they say.

Why AI gets things wrong

Mike explained that chatbots predict the next likely word rather than reason or calculate, so they cannot truly do math and they will answer rather than say they do not know. He cited a study showing the more wrong an answer is, the more confident the AI sounds, so very certain answers deserve skepticism. His example was the number of R’s in strawberry, which AIs often get wrong.

Getting better results from AI

Mike said he is fully behind learning to use AI, but not relying on it blindly. He sets standing rules so the AI is less cheerful and must check its work. For coding he has it use one helper to write the script and another to review it before he sees the result.

Context windows and hallucinations

Mike described context as how much an AI can keep track of in one conversation. As a conversation grows, the AI runs low on that memory and starts making things up, which he compared to being very tired. His example was an AI that got his phone number wrong while copying a website, even though it had the correct number. He said this limit is a main reason that claims about easily building finished apps with AI are overblown.

AI escape and blackmail stories

Mike discussed reports that AIs lie, try to escape, or blackmail staff to avoid shutdown. He said these are observed behaviors, not proof of intent, and that models trained on human writing, including science fiction, may simply be repeating likely storylines. He said AIs do not have enough context memory to take over the world.

AI as therapist and smart assistants

Mike said around 8% of teenagers use chatbots as mental health professionals, which he called a red flag, and he said some kids have died after doing so. He also described Google’s assistant becoming worse after the switch to Gemini, failing to find a contact he then found by listing his contacts.

Calls

• Dan asked how people make AI-restyled movie trailers on YouTube. Mike said the tools are often plug-ins to video editing software or online services, and that the process amounts to a full-scene deepfake that is heavy on computing but needs little hand coding because AI writes the code. He said he plans to rebuild the show website to list tools and links for caller questions.