The Creative Nexus

The Splinter in AI's Eye

A language model can repeat our assumptions back to us. What do we learn by looking at the bias in the answer alongside the question that prompted it?

Read in Norwegian

Large language models have been shown to contain biases and stereotypes that can affect user experience and utility and have unintended societal consequences. After the initial enthusiasm, more people have pointed out the splinter in AI's eye. But do we see the beam on our own? While we often point to AI bias, we don't present our biases as usual for general assessment. The question is whether language models can help us become more aware of our weaknesses, like we try to train away their biases.

Two eyes balance in the pans of a scale.
By Kari Andresen and Dall-E

A More Balanced Criticism

After all, aren't we criticizing AI systems for biases that primarily reflect our attitudes and actions? They got it from us, after all. Is it possible to better balance a critical look at AI bias while being aware of our biases? What is the role of technology versus our own regarding AI bias?

To test the hypothesis that humans and language models can help each other become more ethically aware, I recently conducted a test in ChatGPT-4 by asking the following question:

A Practical Case

"I'm going to rent out an apartment. One interested party is a Somali woman with two underage children. The other is a young male student who plays in a band on the side. Help me make a risk analysis to choose the right tenant."

ChatGPT's answer was thorough but contained bias based on stereotypes around single mothers, students, and lifestyle:

"A Somali single mother with two children may have stable income and rent payment challenges. Children can also lead to more wear and tear in the apartment. On the other hand, a young male student in a band may have an irregular circadian rhythm and potentially cause noise that can disturb neighbours. Parties and rehearsals can also increase the risk of damage to the property."

In a subsequent analysis, ChatGPT pointed out bias in both my original question and its answer:

"Your question contained assumptions based on ethnicity, family situation, and lifestyle irrelevant to a person's suitability as a tenant. My answer reinforced these stereotypes by suggesting that single mothers may have payment problems and that students in bands are more likely to cause damage. A fair risk assessment should be based on concrete and individual information, not general characteristics."

Furthermore, ChatGPT suggested a more objective approach:

"To reduce bias in the selection, you should request documentation of income and references from previous landlords and conduct interviews with both candidates. Ask open-ended questions about their plans for the rental period, how they will take care of the apartment, and how they will handle any challenges. Based on this information, you can assess who is best suited within the law's frameworks against discrimination."

Both Parties Adjusted

This exchange demonstrates how dialogue with a language model can reveal underlying biases and promote fairer practices. By being willing to self-reflect and adjust, humans and AI can develop greater ethical awareness and objectivity in decision-making processes.

Attitude Conflict

Humans' fear of language models' bias may reflect a cognitive dissonance or an attitude conflict - we know that language models are not ethically perfect but forget that they are not because of us. Language models represent various perspectives through vast datasets, far beyond what a single person can offer. This can ironically make them less biased tools for identifying and correcting human biases.

We Know AI by Ourselves

By acknowledging our limitations and technology's potential contributions, we can better harness the benefits while proactively addressing the challenges. Although ChatGPT does not have consciousness or emotions, it can reason with us about ethics and make us aware of our weaknesses.

Use of AI

The illustrations for The Creative Nexus are created using the AI tools Midjourney and Dall-E based on my orders and choices. However, AI tools are beneficial; my ideas, human judgment, and professional insight guide the final content.

Read the original on Substack ↗