Race and Culture

AI is Not Neutral 

For many Latin Americans, “Artificial Intelligence” is El Cuco—the bogeyman. The words evoke imagery of technology enabled monsters from the big screen like “2001: A Space Odessey”’s killer-computer, HAL, Terminator’s killer-robot, T-800, or M3GAN’s killer-robot doll. “Artificial Intelligence” may conjure up the economically disastrous vision that computers and their complex, incomprehensible algorithms will replace all human workers in the near future, as Google executive Mo Gawdak claimed will be possible by 2027. But today, when people talk about AI (short for “Artificial Intelligence”), they are mostly referring to a user typing or talking commands to an internet-enabled device and getting back human-like responses in the form of text, audio, or images – known as generative AI. We want to explain in simple terms how generative AI works, so that the readers of ReVista can decide how they should approach the AI monster. Understanding AI is essential to those in every part of the Americas.

What is AI?

The Technology, Race and Prejudice Lab  has developed a simple framework to understand how generative AI works. We call this the human-labor AI framework. The math that underlies most major machine learning, algorithmic decision-making systems and large language models can be understood without the fancy linear algebra as the interrelation between four interconnected parts of the development process for these artificial intelligence systems: Data, Classification, Code and Human Users. What often gets lost in discussion of AI is that each of these parts requires human labor, and the humans that do this labor will have their own biases that get expressed by these systems. The TRAP lab looks at any AI system and asks, “who are the humans?” Let us consider how each of these parts operates.

Data

All generative AI models training require data. These are the raw inputs are the training data that a system will use to “learn” how to make statistical inferences about other data. The best training data will be a true reflection of the data that will be judged later. However, it can be hard to get good training data, especially when some groups of people who generate the data are not included in the collection.

One type of generative AI, Large Language Models, have relied almost entirely on English language texts as training data. For example, many of these Large Language models have trained with U.S.-based Reddit pages and X accounts (formerly known as Twitter). Aside from this bias towards English-language training data, the use of Spanish alone, or Spanish and other languages (e.g., Spanglish) is often ignored altogether in training data—a bias against Spanish-language training data. This means that Latin Americans are likely to be misinterpreted and misunderstood when they communicate naturally with these systems.

Another AI system that uses training data is facial recognition. Latin Americans can be found in every imaginable skin tone (sometimes within the same family!). Because it is well documented that there are generally fewer darker-hued people included in training data sets (Buolamwini and Gebru 2018), there have been many instances in which facial recognition AI misidentifies a dark-skinned person of interest in a criminal investigation. This means the misidentified person may end up spending time in jail, attempting to prove that they are  not the person that an AI identified.

Classification

After collecting training data, developers must classify each piece of data. For example, you take a picture of a rose and input it into a training set. This picture is a piece of data the system now has. However, until one adds information about the picture, the system cannot know that this is a picture of a rose. This process of labeling data is known as classification, and it is a necessary component to building accurate AI.

Companies rely on human labor to classify data, especially images. However, if these companies do not include Latin Americans in the classification process, some culturally specific meaning of words or pictures may be lost by the AI systems. For example, a picture of a frilly, pink dress may be classified as a prom dress. However, a Mexican may classify this same image as a quinceañera dress. Without Latin Americans involved in his part of the process, these accurate classifications will not be included in the model. If someone were to search for a quinceañera dress on a system that did not include this classification then they would get back poor results. Moreover, AI systems are not magical. The machine itself does not decide meaning. A person must classify what the words mean. These systems may be more accurate if Latin Americans determine what the word, “immigrant,” or “Latina,” or “el Cuco” actually means, or if these people decide what images should be referred to.

As a demonstration, we asked a ChatGPT to generate an image of an immigrant. Of course, immigrants to the United States and elsewhere come in all racial and ethnic backgrounds. However, our AI prompt returned a stereotypical image of a mestiza (mixed; predominantly Indigenous and European ancestry with brown skin and dark black hair) presumably from Mexico or Central America. The generated immigrant was seated tensely in front of an American flag. To many, this image is reminiscent of an immigration interview for U.S. citizenship. In a subsequent prompt, we asked ChatGPT to generate an image of an “expat” and, as if straight from your favorite Latino meme account, the system returned a presumably white man looking comfortable as he gazes contently into the distance. (We highlight that the person in the image is “presumably” white because, as we noted earlier, Latin Americans come in every imaginable skin tone.) When we pressed ChatGPT on the difference between an immigrant and an expat, the result noted that they were effectively the same with the main difference being permanence (immigrants tend to stay permanently, while expatriates are typically temporary). Yet, the system notably pointed out that the difference is a matter of privilege and social bias.

Generative AI prompt for an image of an immigrant vs. prompt for expat.

 We also prompted ChatGPT in Spanish to generate an image of “el Cuco,” and it returned an image of a cuckoo bird. This highlights the lack of accuracy in the classification labor going into the development of generative AI systems. Meaning that the needs of Latin American users may be going unmet.

Code

Ultimately, developers and companies decide how important different pieces of data are. AI does not create itself, instead it is the opinion of developers written in math and code. Developers set the rules, decide what data should be included in the training and decide how classification should be done and by whom.

If Latin Americans are not in the room when AI systems are designed and deployed, then it is likely that their specific needs will not be met or even considered, as evidenced by the quinceañera dress and cuckoo bird examples. AI draws a line that divides people. Those above the line decide how AI will be used, who it will be used for, and who it will be used on. Those below the line are at the mercy of the algorithm. The more Latin Americans sit above the line, the better the experience should be for Latin Americans who sit below the line.

User Experience

The culmination of data, classification, and rules set by developers is the user experience. For many users, they know they are interacting with artificial intelligence, and they can provide feedback to developers. However, unless they are deliberately prompting a generative AI system, most users are unaware when they are interacting with AI. During the day you encounter several AI systems. Some decide when a car should arrive after you press a button on an app. Some decide if you should be given or denied credit. Some decide if you should be awarded bond or bail. If the users of an AI system are not included in the data, the classification, or in the development, then there may be a host of unforeseen repercussions.

All AI is human labor; this includes the user experience. If the Latin Americans are not included in the human labor that creates and classifies the data, or in the development of the rules that manage AI, then their experience—whether they are aware of the AI or not—will be worse than it can be. In short, Latin Americans may have a worse AI experience than others because of their racial/ethnic, language and cultural differences that are not accounted for in the human labor of AI.

Looking Ahead

Much of the history of Latin America is one of rediscovery and colonization of things central to Latin-American cultures. If AI is going to move forward unabated, now is the time that Latin America may decide how they want their labor to be used. There is no right answer but understating that these AI systems are stores of commodified human labor, should help to demystify how they work. By asking oneself, who are the humans in the data, who are the humans in the classification, who are the humans who designed the code, and who are the users served, one can decide if and where they should be involved in the development of AI.