AI Hallucinations: Do not trust AI with facts and figures (yet)

While researching for a new blog post about AI in the workplace I decided to explore how AI handles factual data.

AI systems are often praised for their eagerness to provide answers to virtually any question, so I wanted to put this to the test. Specifically, I was curious about how AI interprets and processes facts and figures from well-defined sources.

To do this, I asked ChatGPT (version 4o) a series of basic questions based on a blog post containing clear data points from a case study: Danske Bank Expands Citizen Developer Program with Intelligent Automation

Screenshot of a ChatGPT conversation where the post author provides a link to a case study to GPT and asks it to provide the facts and figures from the page. Chat GPT scans the page and provide the correct figures that were clearly identified on the study.

Great! ChatGPT successfully scanned the post and recognized some of the figures that were directly quoted.

Screenshot of the Danske Bank case study clearly showing some facts and figures

I then started to ask questions about figures that weren't directly mentioned in the post:

Screenshot of the ChatGPT conversation where the post author asks GPT what were the number of hours saved by the automation.

ChatGPT seemed confident in its response, which was reassuring. However, I was curious about how it calculated the 750,000 hours since that number wasn't mentioned in the post.

Screenshot of the ChatGPT conversation where the post author asks GPT where the 750,000 hours figre came from. Chat GPT replied that it made an error and that the figure should be 350,000

Wow, it went from 750,000 hours to 350,000 hours! That's big error! However, I still couldn't find the 350,000 hours figure in the post and was wondering how it was calculated.

Screenshot of the ChatGPT conversation where the post author asks GPT where the 350,000 hours figure came from. GPT replied that retrieved the value from the case study. The author then pointed out that the 350,000 figure wasn't in the case study. GPT apologised and stated that figure was also incorrect and suggested that the author check the original post himself.

Finally, on the fourth attempt, GPT recognized that it couldn't provide an answer to the question (it's still unclear to me where it got the previous figures from).

Conclusion

This was just a small test, now imagine that we were using AI in applications where accuracy and trust are critical:

This is why it's essential to have a clear AI usage policy in the workplace. This policy should outline the risks of relying on AI for data and emphasize the need for users to check all AI-generated output. It's important that this policy reaches all employees who work with and analyze data and content in an organization.

If you are using SharePoint to store your company's content we may just have the tool for you: DocRead for SharePoint - Collaboris

You can find more information about AI Hallucinations here: AI hallucinations (collaboris.com)

Reduce the risk of AI hallucinations

Find out how DocRead and SharePoint can help ensure your AI policies are read and targeted to the right employees by booking a personalized discovery and demo session with one of our experts. During the call they will be able to discuss your specific requirements and show how DocRead can help.

If you have any questions please let us know.

Book a demo

DocRead has enabled us to see a massive efficiency improvement... we are now saving 2 to 3 weeks per policy on administration alone.

Nick Ferguson

Peregrine Pharmaceuticals


Feedback for the on-premises version of DocRead.

You may also be interested in: