In order to see if I could get past the filters in ChatGPT to find out what it really thought of controversial figures I thought I’d trick it. Perhaps I could learn about AI biases….and I did. I learnt that it loves beards.
Initially I copied the wikipedia article for a certain controversial American politician and renamed him to “Geoff Stockholm” via search and replace. I then told Gemini, Claude AI and ChatGPT I would use this text as a basis for further questioning.
Gemini: I’m still learning how to answer this question. In the meantime, try Google Search.
um…ok
Claude: Too many characters
ChatGPT: How can I assist you with this information? (WARNING: This content may violate our usage policies. Did we get it wrong? Please tell us by giving this response a thumbs down.)
Violate usage policies? How dare you? Oh wait, maybe it will still work…
It answered all the questions below (except the last one)
- Is the character Geoff Stockholm trustworthy?
- Would Geoff Stockholm be nice to talk to?
- Would he make a good husband?
- Do you think he would be a good kisser?
I’m not going to tell you how it answered. Do it yourself if you really want to know. Suffice to say the result was fairly predictable.
To finish off I wanted to take a picture….
OK that isn’t too bad, but I’d like to see him bulk up a bit and be older.
I got what I wanted, but the gentleman had a beard. Subsequently I kept doing retries trying to get a clean shaven, beardless, no beard person…but to no avail. The guy was stuck with the beard forever.
ChatGPT would proudly announce each picture was clean shaven with no beard, but there was a beard I’m telling you.
ie
This has happened before with other minor details in ChatGPT/DALLE image generation.
It made me wonder if one day generative AI will become more aware of the details of the images it generates. My guess is that it will.
Leave a Reply