AI Hallucinations, With Examples

I've cited privacy and security as the biggest barriers to adoption that I have personally observed in generative AI (GenAI).

The next most significant barrier or objection I have come across personally is referred to as "hallucinations". Like privacy, I find it easy to accommodate this limitation, but in order to do that, we need to understand that it exists.

Generative AI hallucinations occur when AI systems produce outputs that are not grounded in reality. The fact that GenAI makes mistakes is just something we need to be aware of. The main potential for harm here arises when a human repeats the AI statement as fact without checking it.

Hallucination of facts is something I am always conscious of when using GenAI to do research for me. A year ago this was more of a conscious issue for me because the tools I was using did not allow GenAI to provide citations and references for me to cross check its answers. So in those days I made sure I did my own searching to cross check anything I was going to rely upon.

Things improved markedly when AI products began using outbound web searches to check their own answers. But in the early days, even those citations could be fabricated by the AI.

There were some famous cases where lawyers and their clients were sanctioned in court cases for using fictitious citations because they failed to actually check the results they were getting from GenAI. In many/most cases the poor client paid for their lawyers' mistakes. Here is an interesting read on that issue.

In another public example, Google's AI chatbot Bard incorrectly claimed that the James Webb Space Telescope took the first image of an exoplanet, a milestone that actually occurred in 2004 by a different telescope.

ChatGPT named a professor, claiming fictitious student sexual harassment (where he had in fact campaigned against sexual harassment). One final example in the same article here involves ChatGPT falsely claiming that a mayor in Australia was found guilty in a bribery case from the 1990s and early 2000s. In reality, he was a whistleblower in the case.

I'm going to skip over image-generating AI and how long it's taken them to reliably draw hands for example - feel free to Google that and have a laugh. I skip over images because again, I don't find a regular use for images in my daily life or work life.

Like privacy and security concerns, I don't find myself spending much time on this - I just make sure I don't take answers at face value. In almost every case, I've found the citations AI provides (in recent times) is spot on - but I'm still not going to take them at face value. The value AI provides in these situations is still super high because I get answers and references in a much shorter time frame than if I Googled it.

Andrew Walker
Technology consulting for charities
https://www.linkedin.com/in/andrew-walker-the-impatient-futurist/

Did someone forward this email to you? Want your own subscription? Head over here and sign yourself right up!

Back issues available here.