ChatGPT has become a largely used AI platform over the last few years, totalling an estimated 400 million weekly users. ChatGPT works by taking user input, turning each word into a “token” (words or parts of words), and generates a response that seems to be accurate to the context of the user's input. When I asked ChatGPT where it gets its information from, this was its response:
- General Knowledge & Texts: Books, academic papers, publicly available websites, and licensed datasets.
- Training Data: A mix of publicly available and OpenAI-licensed texts, with filtering for quality and reliability.
- Real-Time Info: If you ask about something time-sensitive (like current events or local info), I can search the web.
- Specialized Tools: For things like U.S. election policies, I use dedicated tools with up-to-date official sources.
These data sources have not been updated since June 2024, but if you ask ChatGPT to get information from the internet it has the ability to scrape information and provide an appropriate response, at least what it deems appropriate.
In August of 2024, a man from Norway ran into issues with ChatGPT “hallucinating” and providing incorrect and damaging information about himself and the tragic death of his two sons in 2020. AI hallucination is “a phenomenon where a large language model (LLM) perceives patterns or objects that are nonexistent or imperceptible to human observers, creating outputs that are nonsensical or altogether inaccurate” (see IBM definition). In this specific situation, ChatGPT had one of these hallucinations and provided inaccurate information in response to Arve Hjalmar Holmen asking it who he was. When ChatGPT responded claiming that Mr. Holmen was accused and later convicted of murdering his two sons who died, Noyb, a digital rights group, filed a complaint on his behalf claiming that the answer given by ChatGPT was “defamatory and broke European data protection rules about accuracy of personal data.” It has been clarified that Mr. Holmen had not been accused or convicted of these crimes. OpenAI has since responded claiming that this version of ChatGPT is no longer used and has since been enhanced.
This is a specific example of how harmful AI hallucinations can be. Some other instances of AI hallucinations are from Google’s AI, where it suggested using non-toxic glue to help your cheese stick to your pizza better, as well as stating that geologists recommend that humans eat one rock per day. These examples aren’t as harmful (assuming people know not to do either of those things) but serve as evidence of AI hallucinations. This phenomenon can be avoided by:
- Further train your AI model with various quality datasets,
- Use human feedback on the accuracy of the AI models output, and
- Implement a fact-checking system.
There are many other ways to enhance your AI model to avoid running into trouble with this phenomenon. It is important to fully understand how your AI model works and implement features that consider accuracy of its response. With a substantial amount of time and effort spent ensuring that your AI model receives quality data and training, you can decrease the likelihood of your AI producing harmful and inaccurate responses.