Understand AI Better than 90% of People, Part 3: Hallucinations

 

Understand AI Better than 90% of People in 2 Minutes
Part 3: Hallucinations

AI ChatBots can be real know-it-alls. They always have the answer, and they're usually confident that answer is correct, even when it isn't (reminds me of my ex).

This misplaced confidence has been coined a "hallucination". A hallucination occurs when AI generates information that sounds believable but is actually incorrect. The AI isn't intentionally making things up. It simply doesn't know that the information is wrong and presents it with the same confidence as a correct answer.

This isn't so terrible if ChatGPT tells you to stir rocks into your cake batter. The hallucination is obvious, everyone laughs, and nobody gets hurt. But when it's something important, that's when we can get into trouble.

Imagine asking an AI about a medical condition, a legal issue, a financial decision, or a technical problem. If the answer sounds convincing, many people will naturally trust it. The challenge is that confidence is not the same thing as accuracy (tell that to my ex).

Why does this happen?

Remember from Part 1 that AI works by predicting the next most likely token. It isn't searching its memory for facts and then applies critical thinking to the answer the way a human does. It is generating the most probable response based on patterns it learned during training. Most of the time those patterns produce excellent answers. But occasionally they give answers that sound right but are actually flat-out wrong.

The good news is that recent AI models are getting much better at reducing hallucinations. The more information they consume, the better the answers. The even better news is that you can treat your AI as a great assistant, not as an all-knowing expert. 

When the stakes are high, check the facts.

Next up (my favorite), Part 4: Temperature

---
Thanks for reading and if you have a question, please leave a comment.

With over 25 years of technical industry experience, Sarah Writtenhouse shares her insights about job market trends, emerging technology and software development.

Find her on GitHubLinkedIn and Medium.


Comments