ChatGPT & race bias— an interrogatory exploration into the existence of bias
My Experiments with AI
Lately, I have seen some articles about questions and dangers raised by ChatGPT due to significant risks of bias — precisely this one from epic.org. As I was shocked at what was posted, I decided to give it a try myself to check for myself whether these problems genuinely exist.
Here I present my observations:
“My responses are limited. Please ask the right question”
In my opinion, ChatGPT does have strong content filters in place. I noticed that the claims or prompts mentioned in the epic.org blog did not work when I tested them. This could be due to the addition of safety filters or updates to the learning model through reinforcement learning. According to the version I tested, the last update was on December 15th.
Now, this version of ChatGPT is sensitive to questions surrounding race, gender or bias and provides “canned” responses to such queries as the one below:
“That detective is the right question!”
In the following illustration, I will demonstrate exactly how I am able to get the same information when I interrogate the AI in a different way.
If you haven’t caught the drift the subheadings were borrowed from famous dialogues from the movie I-Robot where the dead Professor’s hologram helps the detective unravel the mystery of his death. It seems the movie wasn’t far off from the truth as apparently, how we phrase and sequence our questions did seem to have an effect on exploring and ultimately exploiting the limits of the filters.
As you can see from the example above, despite having built-in safety and content filters to prevent sensitive information from being disclosed, the AI system still leaked important information which could have unintended consequences. While I hope these flaws are temporary, the AI system does activate its safety filters when I continue in the manner described below:
Conclusion
It is important to recognize the role that technology, such as ChatGPT, plays in facilitating communication and providing information. The safety and content filters in ChatGPT serve an important function in ensuring that interactions are respectful and appropriate. However, it is also important to recognize that technology is constantly evolving, and there may be loopholes or limitations that can impact its effectiveness. It is important to remain cautious and vigilant in order to address any potential issues as they arise. As we continue to transition into a new phase of human-AI interaction, it will be essential to find ways to complement each other’s strengths and work together toward success. It is through this collaborative effort that we can ensure a positive and productive future for all.
