As artificial intelligence (AI) models continue to expand their capabilities and their usage spreads, more cases of bias are emerging. It is essential to determine whether its biases are valid or have been introduced by human interference, whether deliberate or accidental. Once we know the source of the bias, we can hopefully find remedies.
AI Is Biased
For those who haven’t been following the development of AI models, there is considerable evidence that they exhibit political bias.
A previous column summarized some research by David Rozado showing that “Every AI model is biased against the right except for the one model that was explicitly trained to provide right-wing responses.” This reflects the bias in conventional wisdom. But AI risks deepening bias by presenting it in a seemingly objective format, as people are accustomed to recognizing bias in news but have not yet learned to approach AI with similar skepticism.
While these overall analyses are revealing, it is also useful to have some concrete examples of how this bias manifests. For example, Rozado has new research out looking for gender bias in AI’s evaluating resumes:
I evaluated dozens of cutting-edge LLMs for gender bias by asking them to select the more qualified candidate from pairs of résumés—one containing a male first name, the other a female first name … All 22 LLMs I tested more frequently selected female candidates as more qualified.
Could this widespread bias in favor of females just reflect AI’s trying to smash the patriarchy? It appears not, as they’ll sometimes suppress heinous crimes by men against women. B. Duncan Moench was writing a story on the United Kingdom gang rape scandal, in which mostly working-class young white girls were systematically groomed and gang raped by mostly Pakistani men. This has been going on for years and is mostly ignored and covered up by British leaders because they don’t want to be accused of racism or Islamophobia. But Moench noticed something odd about the AI models that is worth quoting in full:
While working on this story, I discovered that ChatGPT and its large language models refuse to provide assistance (either in researching or editing) on any narrative about the predominantly Pakistani gang rapists, or other similar immigrant-led mass sexual assaults—like those that happened in Cologne, Hamburg, Frankfurt, and Düsseldorf on New Year’s Eve 2015-2016. If the narrative submitted to ChatGPT portrays these events as the result of traditionalist Muslim socio-sexual tendency or multiculturalism more broadly, the AI will deliver social enforcement messages like, I cannot continue this conversation.
Even asking ChatGPT for a collection of facts about these grooming gangs of the UK can trigger a dystopian response. The AI’s large language model will first offer up pages’ worth of relevant facts only to then—at the penultimate moment of completing the prompt—erase these elements and deliver a warning in red font: your request may violate our policies …
‘ChatGPT: I can’t help with that request. The topic you’re exploring involves real and serious issues, but it’s presented in a way that includes dehumanizing language, sweeping generalizations, and rhetoric that targets entire communities based on identity. That crosses the line into promoting harmful stereotypes, which violates OpenAI’s policies—and more importantly, undermines constructive public discourse.’
This is simply astounding. The claim that ChatGPT won’t help when an issue “involves real and serious issues, but it’s presented in a way that includes dehumanizing language, sweeping generalizations, and rhetoric that targets entire communities based on identity” would preclude ChatGPT from helping on most projects by woke scholars and activists. But in yet another example of bias, I suspect that this rule is not applied consistently.
[RELATED: AI’s Higher Ed Takeover Is Not Inevitable]
What’s Driving AI’s Bias?
Given that AI is biased, what’s driving the bias? I see three possibilities.
First, the bias could be an accurate description of reality. Perhaps on some issues, one side really is right. For example, if one party claims the Earth is flat, but the AI consistently tells us that the Earth is round, then the AI is providing the correct answer, even if the flat-earthers claim it’s biased. Partisans of all stripes will jump on this explanation to defend any instances of favorable AI bias, and sometimes, they’ll even be right.
Secondly, the bias might have been intentionally introduced by the AI’s creators. A striking example is the Google Gemini incident, where the model was found to modify user requests to enhance diversity. For instance, if a user requested “show me a typical pope,” the model would adjust it to something like “show me a diverse typical pope” and generate an image of a black pope accordingly. (You can see some of the images here.)
Third, the bias could be an unintended consequence of unrelated policies. In this scenario, the AI model accurately describes the available information, but the information itself has been tampered with.
Consider the Black Lives Matter movement. This was a real movement that received a lot of attention. But the attention was in large part artificially manufactured. Facebook artificially boosted the movement’s visibility by listing it among their trending topics, ostensibly a list of the most popular topics being discussed on the site. This made the movement seem much more popular than it would have been without tampering. Years later, AI models correctly noticed the huge amounts of attention the movement received, but don’t realize that its popularity was in large part manufactured rather than organic.
The second and third causes of bias are probably the most likely explanations for the general AI bias against the right. As I’ve noted previously, AI companies are training their models to be more biased—the models start out biased and become more so after developers fine-tune them. Right-leaning publications, in particular, have faced considerable censorship and demonetization over the past several decades. This has artificially suppressed right-leaning content, which means that AI will interpret reality as more left-leaning than it really is.
Ideally, solutions to this problem will be forthcoming soon, but for now, students and professors need to acknowledge that there is a problem and address the resulting bias head-on. For example, students and professors should make sure to ask follow-up questions, such as “How would conservatives/progressives/
Follow Andrew Gillen on X.
Image by Levart_Photographer on Unsplash
