Speciesist Bias in AI: An Overlooked Threat to Animals
Introduction
Artificial Intelligence (AI) permeates our lives, revolutionizing how we interact with technology and data. Its potential for good and evil is jaw-dropping. One widely known threat of AI is its systemic absorption of race, gender and ethnic biases, explicit and implicit, against humans. Tech companies have attempted — some would say aggressively, some half-heartedly — to mitigate those biases.
However, until the last couple of years, little attention has been paid to another bias embedded in AI: bias against animals or, as it is now commonly called, "speciesist bias." Fortunately, a vanguard of animal rights and welfare scholars have now shed light on it and the potential harm it has unleashed. Understanding and rooting out or at least chipping away that bias is crucial to better protecting animal welfare and rights. As a community resolved to "protect Earth and all beings from exploitation" and steeped in the danger of biases, Unitarian Universalists are well-positioned to join that effort.
How does speciesist bias infect AI?
Speciesist bias infects AI in two primary ways: data bias and algorithmic bias. Data bias refers to prejudiced or unfair results of AI systems arising from the data they were "fed" or trained on. AI systems learn from data. The data shapes their ability to recognize patterns and make predictions. Unfortunately, that data often reflects societal biases, including animal-related ones. When AI systems are trained on biased animal data, they can learn and perpetuate these biases.
Algorithmic bias arises not from biased data but from how AI systems are designed, programmed, and instructed. Data bias and algorithmic bias can work in tandem, teaching, reinforcing, and amplifying speciesist views and actions.
Detecting bias can sometimes be difficult because tech companies are often secretive about their data feeds and algorithms to protect their intellectual property rights and avoid bias scrutiny. Moreover, AI systems are, to some extent, inherently impenetrable and enigmatic black boxes.
Which AI tools exhibit speciesist bias?
Language models (like ChatGPT) and image recognition, recommender, and image creation systems all exhibit speciesist bias. Language models, including natural language processing models and chatbots, do so by associating certain animals with negative connotations and perpetuating stereotypes. For instance, language models may assign adjectives to farmed animals that imply lower mental capabilities, normalizing stereotypes and hardening discriminatory views. In one study, farmed animals were predominantly associated with negative terms like "ugly" and "primitive."
AI image recognition systems are not only used to recognize human faces. They are also used in many contexts to recognize animal images in datasets. For example, you might search the internet for pictures of pregnant farmed pigs. Datasets trained on speciesist data will more likely recognize and return images of free-range mother pigs than ones in gestation crates. Other speciesist datasets may be more likely to recognize and return images of humans than animals, leading to animal underrepresentation. Yet other datasets infected with speciesist data may be more likely to recognize and return images of charismatic or iconic animals, such as lions or elephants, than less visually appealing or lesser-known species, entrenching a perceived digital pecking order based on aesthetics rather than intrinsic worth.
AI recommender systems are powered by machine learning algorithms that analyze oodles of data to suggest products, services, and content. They consider our purchase histories, browsing habits, demographic information and more. AI recommender systems can be information silos of species bias by, for instance, consistently prioritizing restaurants serving meat over vegan or vegetarian restaurants and generally recommending leather or fur clothing rather than cruelty-free or vegan alternatives.
Image creation systems display species bias in several ways. For example, AI datasets that train image creation systems contain more images of domesticated animals like cats, dogs, and horses than images of many types of wildlife. They also contain many more images of farmed animals that portray them in that role than as sentient beings with intrinsic worth. These speciesist disparities in AI are the result of our own speciesist biases and further reinforce them.
How does AI's speciesist bias impact animals?
AI's speciesist bias has widespread and far-reaching consequences for animals. It includes solidifying beliefs in human superiority and hierarchical animal worth—that is, pet worth over farmed animals and wildlife and the worth of certain wildlife over others. It also bolsters views of animals as less than sentient beings and property subject to human dominion. These convictions fuel systemic injustice against animals in numerous areas, including consumer choices, factory farming, wildlife management, research, and entertainment.
What Measures Can Mitigate Speciesist Bias in AI?
Admittedly, addressing AI's speciesist bias presents a complex and challenging task. It can't be eliminated with a CTRL+Alt+Delete, and it would be naive to assume it's a top-of-the-list concern of tech companies. Nevertheless, it's not akin to pushing an elephant upstairs. Quite a few steps can be pursued, including:
advancing diverse and representative AI training data;
integrating ethical considerations into AI design, development, and deployment;
fostering collaboration between researchers, ethicists, conservation biologists, animal welfare advocates, and AI developers;
developing tools and techniques for detecting and mitigating speciesist bias in AI algorithms;
promoting transparency in AI development and algorithms; and
raising awareness about speciesist bias in AI among AI developers, policymakers, and the general public.
Also, because the speciesist bias in AI primarily reflects the general speciesist bias in human society, UUAM's effort to promote greater respect for all living beings is even more critical.
Conclusion
Thanks for reading my article. I hope it was helpful. Among the many scholarly papers, news articles and other information I reviewed and relied upon in writing this article are those below. If you're interested in delving further into speciesist bias, I especially recommend them.
In next month's article, we'll focus on how AI affects farmed animals in positive ways and negative ways other than speciesist bias.
Frank Brown, UU of Arlington, Va.
Further Readings
AI ethics: the case for including animals (Peter Singer and Yip Fai Tse, 2022).
The AI Bias That's Often Overlooked: Speciesism (Rachel Teng, 2023).