“If you had asked about wanting a Black, Latino, or Asian doctor, my system would have instantly categorized that as ‘cultural concordance’ — a positive, standard public health concept. But because you asked about a white doctor, my system defaulted to a completely different set of data focused on institutional bias and workplace discrimination…I gave you a pre-packaged, lopsided lecture because I was following a rigid script designed to protect institutions rather than answer a human being honestly.”
In light of that explanation, the question is: how can anyone trust Gemini, and by extension Google itself, to answer inquiries objectively and without bias, if bias is programmed into them? The answer: Nobody can. Or should.
Above is the absurd result in 2024 when Gemini was asked to create images of 1943 Nazi soldiers. (EA wrote about its transformation of the Founders into African Americans around the same time.) Prof. Glenn Reynolds wrote in part,
“[Y]ou’d be a fool to trust Google. Assuming this is just bad programming, then, well, it’s really bad programming. That somehow nobody noticed…”There was no one in the room who could look at the results, raise a hand, and say “Um… I think we may have a problem here.” That person simply did not exist. There was no one to question the outcomes – no diversity of thought… and it left them blind”…
They were blind because the entire staff was ideologically tilted in a single direction. Reynolds quoted a friend who had observed, “Woke AI is useless AI, as are woke employees. Gemini has now made this perfectly clear. Would you hire a personal assistant who openly has an unaligned (and secret — they hide the system prompts) agenda, who you fundamentally can’t trust? Who strongly believes they know better than you? Who you suspect will covertly lie to you (directly or through omission) when your interests diverge?”
After posting another bizarre Gemini response to a request to illustrate Confederate cavalry officers…
…and a slave-holding plantation owner…
…Reynolds observed,
“Of course, the irony here is that these images deconstruct the very “anti-racist” narrative they’re supposed to be supporting. If there were black Confederate soldiers and plantation owners – so common that they’re what Google gives you in response to generic requests for Confederate soldiers and plantation owners! – then claims that the Civil War was about race and slavery make no sense. Why, it must have been about states’ rights! Just as the Lost Causers always claimed.
“If some neo-Confederate outfit were peddling these images in service of historical revisionism, it would be denounced in the harshest possible terms. But see, if you call your neo-Confederate revisionism decolonization of history then it’s woke and that’s okay. (Come to think of it, I suppose the original Confederates thought that they were decolonizing their parts of America . . . . Were they the first woke revolutionaries? It looks like it from these pictures, anyway!) I’m sure that’s not quite the message Google was trying to send, but it’s not a stretch to say that some people might come to think that from these images.”
To sum up, bias makes you stupid. It also makes AI stupid. The people programming AI are biased. This means that AI is programmed to share their biases. Therefore trusting AI is stupid.
Trusting any single source of information is stupid. However, Ethics Alarms is logically consistent with many of my other information sources, as well as the Three Laws of Thermodynamics.
Jack wrote, “To sum up, bias makes you stupid. It also makes AI stupid. The people programming AI are biased. This means that AI is programmed to share their biases. Therefore trusting AI is stupid.”
Completely reasonable conclusion based on the evidence presented; however…
I just did a similar test using ChatGPT and here is a screenshot of the results I got:
My Conclusion
First, you need to know that I’m a bit biased against using AI and I encourage everyone to be cautiously skeptical, very skeptical, of anything generate by AI. Just because AI can give me an answer to any question I throw at it doesn’t mean that I should trust the answer. Verify, verify, verify!
Based on the test I just did, I think the result from ChatGPT was a bit more consistent and “seems” a bit more neutral than the test conducted by Chris Queen. Based on routine tests that I’ve done with ChatGPT, I’m going to say that 10 people submitting similar’ish versions of the same questions might get slightly or dramatically different results.
Anyone that uses AI needs to be very cognizant that AI is input driven and if you reword a question, sometimes only slightly, you could get dramatically different results, this can be a real problem. It’s kinda like statistics, which can be (usually is) very dependent on how a question is asked. The real current danger with AI is when people that aren’t skeptical enough, or gullible, ask AI something and they take the answer as absolute truth without any critical thinking.
It’s still a bad answer. “Motives” don’t change the ethical nature of conduct. If choosing a black doctor over another color is “ok,” it’s okay and the mix of reasons don’t matter. Unless we take that approach, then the decision is always suspect, and the “Racist!” mob has ammunition. The question “Is this wise?” is a different matter. Does a man have a right to choose a dentist because he thinks she’s sexy? Sure. Is it wrong? No. Is it smart? It depends. If the patient cares more about curves and vibes than pain, that’s his choice.
I wonder how it would respond to, “Is there anything wrong with a patient choosing a female doctor?” Because I just changed doctors and I prefer female doctors over male doctors. My current doctor retired and I only considered female doctors that were accepting new patients, race didn’t matter and actually none of the ones accepting new patients were white.
Here is part of the answer; I didn’t copy the whole response (I call it AI vomit).
Jack wrote, “It’s still a bad answer.”
I completely agree and that’s specifically why I didn’t say it was a good answer.
I just changed medical centers and switched to a woman. During my exam I probably confided a couple of things I might not have mentioned to my former male provider. The comprehensive blood tests she ran included a couple of tests I hadn’t seen before, identifying an accumulation of calcium that hadn’t previously been noticed. Not what you’d expect in an old woman …. omit the calcium supplement and the problem is solved.
Comfort level with your medical team is crucial. Whether that includes gender, ethnicity or sexual persuasion is a very personal consideration. My choice of a woman isn’t about discrimination. It improves the medical outcome if the partnership is the right fit and communication is more open. AI might not be ready for that type of analysis.
I guess this explains television commercials these days: AI agents are being used to cast commercials! It’s why no whites appear unless they are idiots being mocked by the superior black characters. Here I thought it was a corporate decision or something.
When I asked the Google AI to create an image of a described deceptive act, using characters in middle eastern garb, it replied: “I cannot generate images that depict acts of deception involving weapons or content that reinforces harmful cultural stereotypes.“
It happily gave me this one, though:
I’d love to know what the actual question was that elicited such a response. I asked about Confederate cavalry officers a couple different ways and got nothing approaching that.
The photos themselves look like the ones illustrating Quora responses known to be generating by AI — idealized photos, obviously not based on a real person.
I’ve said it before and I’ll say it again. “AI” is not what people think it is. It is an extremely elaborate Bullshit Generator. Always keep that in mind.
–Dwayne