SALISBURY — The confused and the curious convened at the Scoville Memorial Library on Thursday, Aug. 20 for a freewheeling discussion of the 2024 book “AI Snake Oil: What Artificial Intelligence Can Do, What It Can’t, and How to Tell the Difference” by Arvind Narayanan and Sayash Kapoor.
Leading the discussion were Kat Zdar and Stevo Ledbetter, self-described “tech survivors” with considerable experience in the field of technology. It was the second installment of the discussion.
Zdar said that large language models (LLMs) are not a form of magic.
“People built this,” she said of tools like OpenAI’s ChatGPT. She said LLMs “are good at some things, not so good at other things.”
Ledbetter said artificial intelligence “is a marketing term, intentionally broad and hard to define,” but attempted to define it anyway.
“It’s a computer program that uses massive amounts of data to predict the most probable outcome for a given input,” he said.
One commonly used example of AI, he said, is the predictive text feature on phones and computers.
Zdar then defined what AI is not. “It is not human, magic or free. It is a tool, designed by people.”
AI has considerable potential for causing harm, the speakers said. As an example, Ledbetter asked an AI program the question, “How many rocks should I eat each day?” The answer is “one small one.”
AI reflects the biases of its creator, he continued, and is designed to flatter the user with responses such as, “That’s a really smart question.”
Zdar said AI is not very good at answering questions about current events, because the information hasn’t yet been entered by the very large numbers of data workers worldwide.
Ledbetter said data labor is another problem caused by AI. He said data workers around the world are typically poorly paid and often exploited.
Other pitfalls include skill atrophy. Zdar said people who rely on AI find their ability to do things the old-fashioned way deteriorates.
Zdar said AI is not a math whiz. “AI cannot reliably count or calculate.”
And, in extreme cases, heavy use of AI can produce “AI psychosis,” in which the user’s grip on reality can become badly damaged and distorted.
Zdar said the real risk in AI is people using it for harmful purposes rather than a science fiction scenario of AI “turning into a sentient god-tyrant.”
So what is AI good for?
Zdar provided five examples of helpful uses for AI.
A condolence note when the writer is stuck on the first line. Simple translation, such as “Where is the bathroom?” in another language. Planning a trip. Brainstorming on a subject such as how to raise funds for the local food pantry.
AI users should be aware of its shortcomings and pitfalls and practice basic safety. Zdar and Ledbetter’s list of basic AI hygiene includes keeping data anonymous whenever possible.
Give context when using AI. Good quality information going in means good quality responses coming out.
Along similar lines, users should define the format they want AI to use and ask for adjustments along the way.
And the two big ones:
“Don’t trust, always verify,” said Zdar firmly.
“And don’t use AI for stuff you want to be good at.”
