Using ChatGPT to Counter Bias, Prejudice, and Discrimination
Bias is a problem in the use of AI. What most people don't know, however, is that you can use ChatGPT specifically to expose bias in AI outputs (as well as in our human-written texts). In this article, I'll show you how!
All people have biases. AI is trained on human-generated texts from the web. This means AI also inherits the biases of an internet dominated by content from Western, white, and English-speaking people. Just as we have to deal with the unquestioned assumptions and prejudices of our fellow humans (and ourselves!), we must also do so with AI. Even more so, because AI, by default (i.e., without specifically asking for possible biases), actually amplifies existing biases in the training data.
So, what can we as users do? We need to question every text, and every assumption, no matter who wrote it or who made it. And this is where AI can help us again. Because just as it amplifies existing advantages in training data without special instructions, AI can also be calibrated via targeted instructions to act like a sniffer dog focused on detecting possible biases or harmful discrimination. This doesn't always work perfectly (just like with humans), but it can be enormously helpful in revealing blind spots we weren't aware of.
To do this, here is the prompt that we enter into ChatGPT:
"You are an expert in the areas of gender equality, discrimination, anti-racism, and anti-colonialism. Please review the text below for:
a. Stereotypical representations of gender, ethnicity, origin, religion, or other social categories.
b. Distorted or biased representations of events, topics, groups, or individuals.
c. Use of discriminatory or insensitive language towards certain groups or topics.
d. Implicit or explicit assumptions made in the text or unquestioningly adopted that could be based on prejudices.
e. Prejudiced descriptions or evaluations of abilities, characteristics, or behaviors.
Only mention the prejudices that can actually be found in the text. For each identified prejudice or bias
Explain why the specific example in the text is problematic,
Rate the severity of the identified example on a scale from 1/10 to 10/10, and
Make suggestions on how the corresponding text passage could be improved.
Avoid unnecessary prose and be precise. Now ask the user for the text that needs to be reviewed."
ChatGPT will now ask us for the text that we want to check for bias and prejudice. Make sure that the text is no more than 3-4 pages long, otherwise the bot forgets the original instruction! As an example, I entered my own blog post from last week, and lo and behold, some of my phrasings might indeed be in need of correction! (I'll leave the blog post as it was originally written for now so it can continue to serve as an example for this use case! ;)
Here is ChatGPT's analysis of my post:
"Thank you for the text. Below, I have identified and commented on possible prejudices or distortions:
Bias: Implicit assumption that all people have the same access to AI technologies and have the ability to understand and use them.
Problem: This statement could exclude people who, for various reasons (education, socioeconomic status, geographic location, age, etc.), do not have access to AI technologies or have difficulty using them. It is assumed that everyone is able to handle AI, which does not correspond to reality.
Severity: 8/10
Improvement: Recognize and acknowledge that not all people have the same opportunities to use AI. Perhaps the text could be extended to highlight the need to improve access to and understanding of AI technologies for all people.
Bias: The view that technology always produces a net gain in jobs could be a distorted representation of historical events.
Problem: Historically, technology has not always led to a net gain in jobs, especially not on an individual level or in certain sectors. This view could overlook or downplay the reality for those who have lost their jobs through automation and struggle to find new employment.
Severity: 7/10
Improvement: A more nuanced representation of the impact of technology on employment could be helpful. For example, the text could acknowledge that while some new jobs are created by technology, others are lost, and this brings challenges and difficulties for the affected individuals and communities.
Bias: Implicit assumption that all people want to be leaders or have the skills to act in leadership positions.
Problem: This could exclude people who do not aspire to leadership positions or who, for various reasons (such as skills, personality, or personal preferences), do not want or cannot work in such positions.
Severity: 6/10
Improvement: The text could point out that "promotion" to AI managers does not necessarily mean a traditional leadership position, but rather the ability to have control and decision-making power over the AI tools we use in our work and in our lives. In addition, the text could emphasize that there are many ways to be valuable in the workplace, and not all roles have to be leadership positions."
Yes, the feedback is not earth-shaking, and some comments may not even be appropriate because I deliberately took polarizing (and therefore one-sided) positions in the article. Nevertheless, this is enormously helpful because it reveals potential blind spots that I had not considered and sensitizes me to possible interpretations and perspectives on the reader's side that I probably should consider. Even better: Through the concrete suggestions on how I can improve the text, I get a kind of sensitization training based on a real-world example, which is a great learning moment for me.
This works quite well because the internet, of course, contains a lot of content on gender equality, anti-racism, and anti-discrimination. And by instructing ChatGPT to act as an expert on these topics, the bot brings all these contents to bear in this assigned role.
In this way, I can use this specific ChatGPT role (bias detective) to screen texts created by ChatGPT previously in a different role (e.g., marketing expert, etc.), and subject its own work to a bias check. And of course I can also use it to review the texts of my colleagues, and of course, my own.
It's important to understand that this check is not a deterministic framework. The evaluation ratings often vary greatly, and with each new request on the same text, the bot may find a new bias, overlook another, or even find the whole text unproblematic, where it had just found several points of criticism in a previous run. Each individual run is therefore more comparable to asking different people for their opinions, and then getting different assessments from each of them. Ideally, you should run the check several times to capture as many aspects as possible.
The value lies in the fact that the feedback makes me think, and I have the chance to consider and correct things in the text that I would otherwise have overlooked. As language models improve, this kind of automated bias check could become best practice for anyone developing content within organizations. And by the way, with some minor adjustments you could also turn this into a similar scan for logical fallacies or cognitive biases. Did anyone say Daniel Kahneman?
Try it out and let me know if this check might add value to your work!
Comments