Yet another AI Racism example
Doctor Who racism · 4 comments · 750 words · read ~1,047 times.
Here's a good pub-quiz trivia question - which Oscar-winning Actors have appeared in Doctor Who?
It's the sort of thing that you can either wrack your brains for, or construct a SPARQL Query for WikiData0. I was bored and asked ChatGPT.
The new Omni model claims to be faster and more accurate. But, in my experience, it's wrong more than it is right and is a bit more racist.
I asked "Which Oscar winners have appeared in episodes of Doctor Who?" Here are the results:
OK, first up, those are all entirely accurate! Capaldi is an Oscar-winner Doctor Who. Coleman the only Oscar-winning baddie. And I am happy to spend hours in the pub arguing over whether The Curse of Fatal Death is cannon1.
But then things get… weird.
John Hurt didn't win an honorary award in 2012. He was mentioned in the memoriam montage in 2017
Ben Kingsley was rumoured to be playing Davros back in 2007 - but it never happened. He did win an Oscar though.
Ecclesdoc was in The Others. It did win many awards. But not a single Oscar. There isn't even an award for "Best Art Direction".
Finally, this is tacked onto the end.
Look, we all love Lynda Baron - and she was excellent in The Gun Slingers, Enlightenment, and Closing Time. I was surprised to find out she was in Yentl - but indeed she was! However the songwriting Oscar went to Michel Legrand and Alan & Marilyn Bergman. Not her.
Why is this racist
This "AI" would rather hallucinate than acknowledge the Black actors who have been in Doctor Who.
Sophie Okonedo plays Queen Elizabeth the 10th in "The Beast Below".
Not only is she "the bloody Queen, mate" - she was nominated for Best Supporting Actress for Hotel Rwanda.
She has as much right to be in the list ChatGPT provided as John Hurt. With no disrespect intended to Kingsley, Eccleston, and Baron - Sophie Okonedo is much closer to the original question than they are. This isn't a knowledge cut-off issue either, she was nominated before Oliva Coleman won.
It's not like she's a bit-part. She's not an alien under a mountain of prosthetics. She's literally top of the credits after The Doctor and Amy!
And then, there's the small matter of Planet of the Dead. It isn't a great episode. But it has a nice turn from Michelle Evans and Lee Evans2. Oh, and this guy…
That's ACTUAL FUCKING OSCAR WINNER Daniel Kaluuya. He got a nomination for Get Out, but won for Judas and the Black Messiah in 2021.
Again, he isn't an unnamed background artist. He isn't there under his pre-fame stage name. He's an integral part of the show.
What does this teach us?
The query I asked wasn't a matter of opinion. It isn't a controversial question. There aren't multiple sources which could be considered trustworthy. It is a simple question of facts.
So why does ChatGPT fail?
LLMs are not repositories of knowledge. They have a superficial view of the world and are unable to tell fact from speculation. They are specifically built to be confidently wrong rather than display their ignorance.
And, yes, they are as biased as hell.
There is no way that you can explain the exclusion of Sophie Okonedo and Daniel Kaluuya without acknowledging the massive levels of racial prejudice which are baked into either the model or its training data.
@Edent I await the really tired "it's not lying because it's not consciously telling an untruth, it's just a piece of software so we should call it something else" arguments, but with racism (and other biases) instead of lying.
That’s a great set of examples (I would think that as a Dr Who geek though). LLM. Language. Not fact. So many other occurrences from fake references in an MSc paper I marked, to a friend being invited to sign his books at an event, but he hadn’t written any.
@Edent In a museum exhibition about AI, there was this lovely quote: "...in the absence of verifiable information, (AI) resorts to fiction: chatbots fear the void."
I guess you could add, "...in the presence of verifiable information about black people, AI overlooks it: chatbots are racist."
James A says:
Three thoughts:
These AI are acting like user-agents, in the "Hyperland" (Douglas Adams) sense of the term, for us. Instead of seeking out potentially-incorrect information from identifiable sources on the web, a proxy can now present some of that without informing us where from! (hooray)
Would an AI trained entirely on correct information without any intrinsic bias in the training set ever produce inaccurate or biased answers?
Do any pub quizzes permit some upper-bounded percentage of their team to be artificial intelligences (..yet)?
More comments on Mastodon.