AI is an increasingly omnipresent element of society, but its use comes with a number of ethical and legal issues. Some of the concerns about AI are bias and fairness, accuracy and misinformation, data privacy, labor and automation, and safety and accountability.
In order to ensure that AI systems are used ethically, many AI ethics experts recommend that AI applications meet certain standards. Exactly which standards are used varies, but some commonly included AI ethics standards are:
In common usage, "bias" is simply any preference for or against something or someone—any subjective opinion that isn't responsive to facts. In machine learning, "bias" refers to a tendency for an algorithm to return results that are incorrect in some systematic way. For instance, a recommender system for a streaming site might be biased toward recommending movies that were produced in-house.
Bias in algorithms can come from the structure the algorithm, or it can come from the data that the data is given. No matter what, algorithmic bias always originates from the bias and assumptions of the people who created the algorithm.
Algorithmic bias can have tangible and negative effects on the real world. For instance, an algorithm that pre-scans resumes for hiring that is biased against traditionally female names can further an existing hiring or wage gap. Algorithms used in the criminal justice system can be biased to suggest harsher punishments or increased surveillance for people of color.
Machine learning algorithms are predictive, and many of them are very good at the tasks they're designed for, but none of them are perfect. Algorithms that have been designed for content moderation (for instance, detecting explicit imagery or offensive text on social media) often have a lot of both false negatives and false positives. That means that offensive or dangerous material stays up, while useful or benign material is taken down. Inaccuracies in medical diagnostic algorithms, security algorithms, or criminal justice algorithms can have far more significant and negative effects. Although humans are also often a source of inaccurate information, it's important to bear in mind that AI algorithms are not infallible, and that their results need to be checked.
Generative AI like ChatGPT introduces another way that AI can propagate inaccuracy. Generative AI algorithms predict text, not facts. When an algorithm like ChatGPT writes something, it has no way of knowing whether what it's written is accurate--it only knows that it sounds like something that a human might write. Therefore, ChatGPT often fluently and convincingly writes totally inaccurate statements. This makes the use of generative AI algorithms for purposes like search engines very troublesome.
The entire purpose of AI is to replicate tasks that humans can perform. This means that when AI is introduced into a field, it can take jobs that were previously performed by humans. In some cases, that's good! It would be great if we could replace jobs that are dangerous and unpleasant with AI. For instance, content moderation on social media is notoriously traumatizing for human moderators. If it could be reliably replaced by AI, even partially, that would be a huge benefit. But our economy requires human employment. Without some method of replacing income from jobs taken by AI, we all struggle as a society.
AI is also increasingly being used to try to replace human labor in jobs that people both enjoy doing, and in jobs where people enjoy interacting with another person. The most high-profile example of this right now is the increasing use of AI-generated artwork, which takes jobs away from working artists, but it is present throughout the creative industries, even in roles that you might not think of as creative. For instance, human book translators are increasingly being replaced by AI translators, or asked to work at much cheaper rates to "fix" the translation that an AI has created. In addition to the labor concern here, there is also an issue of accuracy and quality; AI-produced translations may be cheaper and quicker to produce, but they often fail to capture the full nuance of a text.
Creative jobs also raise another concern about the use of AI: copyright. Generative AI is trained on large datasets of text or images that have been publicly posted. Much of this writing and artwork was created by artists who did not consent to having their work used to train an AI--especially not an AI that is now being used to replace their labor. The legal status of AI training data is still in flux.
The most famous safety concern involving AI is the decision-making behind self-driving cars. Because AI predictions are not always accurate, there is the potential for a car run by AI to mistake a pedestrian for something non-living, and cause harm or even cost someone's life. But this is true whenever an AI runs something that could potentially be dangerous, whether that be heavy machinery or a medical diagnostic algorithm.
Humans, of course, are equally capable of making mistakes--just look at the number of fatal car accidents every year. But humans can also be held accountable, which an algorithm cannot. Moreover, the algorithms used for decision-making are increasingly "black box" deep learning algorithms, in which even the programmers do not understand the factors that lead the algorithm to make a decision. So when a black box AI algorithm does make a dangerous mistake, it can be difficult to understand where the mistake came from or how to correct it.
Machine learning relies on large collections of data, and collecting that data in ways compliant with traditional data ethics would be onerous. Therefore, many "big data" sets are collected without the clear permission of the people the data are about and by. Many sets are scraped from publicly posted videos, images, or text. But posting a video on the internet is not necessarily implied consent for including that video in a machine learning data set, especially one whose purpose the creator may not know about or agree with. Even data for which permission is technically given (for instance, when users on a social media site click "agree" on the terms and conditions that allows the site to collect user data) it's not always clear that the people giving permission have been thoroughly informed of what they're agreeing to.
Some of the purposes of machine learning algorithms might be considered invasive of privacy. For instance, facial recognition algorithms can trace where people are, or identify them from photographs--even if they never gave permission for those photographs to be shared.
The increasingly common presence of AI in day-to-day life has heightened the need for transparency in its use: people should be aware of when they are interacting with artificial intelligence, who created the AI they're using, and for what purpose.
Advances in generative AI have made transparency a particular concern. Recent versions of software like ChatGPT can create text in response to a prompt that is indistinguishable from human-produced writing. In academia, this creates concerns over academic integrity in assignments, and is leading to a reevaluation of the types of writing assigned to students. In journalism, some online outlets have already begun publishing articles generated by AI. Given the issues with accuracy in generative AI, a lack of transparency in its use in journalism leads to lower confidence that what we're reading is correct.