Both users and observers of AI are rightly concerned about bias in AI. Recent research clearly indicates that this bias exists and has detrimental effects. Cem Delgemadi, an analyst at AIMultiple, points out two types of AI bias: 1) cognitive bias, in which AI programmers or trainers introduce their own personal biases into an LLM, and 2) bias resulting from incomplete data. If the LLM model is not trained on enough data or is trained on the wrong data, the model’s results will necessarily be incomplete and will result in biased outcomes. See https://research.aimultiple.com/ai-bias/ .
By contrast, Simon Friis and James Riley of Harvard Business School cite three types of AI bias: 1) algorithmic bias, arising from programmed decisions that automatically limit the body of data considered by AI in solving a problem, 2) automation bias, in which present realities influence the impact AI will have on society in unequitable ways (example: most of the jobs currently threatened by AI are held by disadvantaged groups), and 3) demand-side bias, in which the use of AI in certain fields may lower the demand of services in those fields. For example, some patients might trust their doctor less if they knew he/she were using AI to help diagnose those patients. See https://hbr.org/2023/09/eliminating-algorithmic-bias-is-just-the-beginning-of-equitable-ai .
Why Is AI Biased?
Both articles advocate steps that can be taken to mitigate AI bias. But we should realize that the dream of complete AI equity will always be elusive. The reason for this elusiveness is simple, although its implications are profound: all data are biased. By definition, our observation, recording, analysis, and use of data requires decisions that will favor some interpretations over others.
Further, a situation in which all possible interpretations of data are considered equally is not only impossible; it’s undesirable. Recently I asked Google Gemini to make the best possible case for flat earth theory. It began its reply by noting: “while I am a geophysical scientist, the vast majority of evidence points to a spherical Earth. However, to play devil’s advocate, let’s explore some arguments used by flat-Earthers, though they lack mainstream scientific backing.” Unsurprisingly, it wasn’t able to make a very strong case. And it actually admitted that in making its case it had to ignore the vast bulk of scientific data contradicting flat earth theory.
Hence, some biased interpretations of data aren’t necessarily wrong or undesirable. Many biases simply reflect the fact that some answers to questions are much more likely to be the case than others. The question has instead always been: which biases are undesirable, and which ones aren’t? Or put another way: which interpretations of data are more likely to produce outcomes that we prefer or that fit with our existing beliefs about reality?
Any scientific study begins with a hypothesis that is itself a set of assumptions about reality that will be tested through experimentation. And assumptions are simply choices we make to prefer some data over other data.
I am not asserting that all bias is morally good or even morally neutral. Clearly, some biases lead to human injustice and suffering, and so should be eliminated wherever they are found. But just as the concept of unbiased knowledge is impractical in other fields of endeavor, so it is unrealistic in the realm of artificial intelligence. The only way to guarantee the elimination of all bias would be to introduce all data into the consideration of any research question. Only God would be able to do that. Without omniscience, the best the researcher can do is to formulate a research hypothesis based upon knowledge he/she currently has. And that knowledge base is necessarily limited and is subject to the researcher’s bias.
AI Equity as Fairness
If a fully unbiased AI is impossible and undesirable in the human realm, what is the next best option for pursuing AI equity? A more modest yet achievable goal would be to pursue AI equity in terms of genuine fairness. By fairness, I don’t mean a perfect, divine sense of ultimate justice. Instead, I refer more modestly to the human sense of fairness that tends to give others a reasonable opportunity to show their humanness while still pursuing truth.
In other words, when we train AI with our unavoidable biases, we should do so with the highest forms of goodness, beauty, and truth in mind, all while recognizing the constant possibility of human error. This task will require high-mindedness on the part of AI trainers. But it will also require a good deal of moral humility. We have a fair understanding of what it means to be truly human. And we should train our AI models in alignment with that understanding. Yet our understanding of these things is always flawed even though it might be growing over time.
A sense of fairness will require AI’s collaboration with humans who already have a commitment to this fairness. Emilio Ferrara of the University of Southern California puts it this way: “The pursuit of fairness in generative AI requires a holistic approach – not only better data processing, annotation and debiasing algorithms, but also human collaboration among developers, users and affected communities.” https://theconversation.com/eliminating-bias-in-ai-may-be-impossible-a-computer-scientist-explains-how-to-tame-it-instead-208611
AI Bias and Higher Education
We need to remember the concept of basic fairness when it comes to evaluating the presence of AI bias in higher education. First, we should assume such bias exists, as demonstrated above. Then, we need to make explicit all of those biases wherever we find them. Finally, we must work through the AI models to expand their understanding of possibilities in particular contexts. In that way, AI will hopefully show less and less of those kinds of biases we consider socially unacceptable or harmful.
An important realization is that AI cannot (at least at present) fully understand how its findings influence reality. Instead, it can only reflect reality as reality has been shown to it. For example, if an AI model is taught that, historically, most “terrorists” (a loaded term in and of itself) have been persons of color, we should not be surprised if it shows us mostly persons of color when we ask it to create 20 sample pictures of terrorists. However, the AI model cannot understand that such a response influences the behavior of others by perpetuating culturally harmful stereotypes.
It remains to be seen whether AI models can be taught to be more culturally sensitive in this way. In the meantime, we should expend our efforts in helping to widen AI’s understanding of possibilities in cultural representation and so minimize some of these truly egregious forms of bias. Here are some ways we can begin that process.
More Robust Training of AI Models
We should continue training AI models to be more culturally aware. That sounds simple and obvious. But more specifically, I wonder if AI is actually attuned to what cultural bias is and to the possibility that AI is subject to the same kinds of stereotyping as human beings. If AI can be trained to be sensitive to that tendency and to account for it, then perhaps it will appear to be more culturally sensitive than it does currently. I say “seems” because, in reality, true sensitivity is more than simply a programmable response. But at the very least, maybe AI can be trained to appear to be more culturally sensitive than it appears now.
Training AI to Recognize Harmful Cultural Generalizations
We should explicitly remind the AI model to avoid making cultural generalizations that do harm to specific groups. We should teach the model which sorts of generalizations are culturally harmful and which ones are not. As noted earlier, this process will require the introduction of certain biases into the AI model. But these biases are the only means by which we or any computer could learn that certain stereotypes are undesirable.
For example, using AI to assist in making college admissions decisions can dramatically streamline the work of the admissions office. But unless and until we know exactly what sorts of cultural generalizations the AI model will make in parsing admissions data, we cannot know whether AI-assisted admissions will help or hinder diversity in admissions.
Additional Teacher Training
Teachers should be ready to guide the use of AI in student learning. But they must carefully observe how the AI model interacts with different populations of students, lest the model begin to reflect harmful biases that could actually hinder student learning. Take for example, the genuinely good idea of using AI as a student tutor. If an AI model comes to believe that one group of students is more likely to be math-savvy than another group, that will affect the types of review questions the model will feed those groups of students. Instructors using student learning AI tools such as Pria need to remain aware of this need and be prepared to mentor students on the tool’s proper use.
Conclusion
We must remember that AI is only a tool, albeit a very powerful one. Its capacity to serve as boon or bane will depend upon perspectives and motivations of the person(s) wielding that tool. Most of the story of AI remains to be told. We in higher education should therefore adopt this tool with cautious enthusiasm, always remaining alert as to how it can do damage as well as good.