Introducing the Authors Alliance’s First Zine: Can Authors Address AI Bias?

This guest post was jointly authored by Mariah Johnson and Marcus Liou, student attorneys in Georgetown’s Intellectual Property and Information Policy (iPIP) Clinic.

Generative AI (GenAI) systems perpetuate biases, and authors can have a potent role in mitigating such biases.

But GenAI is generating controversy among authors. Can authors do anything to ensure that these systems promote progress rather than prevent it? Authors Alliance believes the answer is yes, and we worked with them to launch a new zine, Putting the AI in Fair Use: Authors’ Abilities to Promote Progress, that demonstrates how authors can share their works broadly to shape better AI systems. Drawing together Authors Alliance’s past blog posts and advocacy discussing GenAI, copyright law, and authors, this zine emphasizes how authors can help prevent AI bias and protect “the widest possible access to information of all kinds.” 

As former Copyright Register Barbara Ringer articulated, protecting that access requires striking a balance with “induc[cing] authors and artists to create and disseminate original works, and to reward them for their contributions to society.” The fair use doctrine is often invoked to do that work. Fair use is a multi-factor standard that allows limited use of copyrighted material—even without authors’ credit, consent, or compensation–that asks courts to examine:

(1) the purpose and character of the use, 

(2) the nature of the copyrighted work, 

(3) the amount or substantiality of the portion used, and 

(4) the effect of the use on the potential market for or value of the work. 

While courts have not decided whether using copyrighted works as training data for GenAI is fair use, past fair use decisions involving algorithms, such as Perfect 10, iParadigms, Google Books, and HathiTrust favored the consentless use of other people’s copyrighted works to create novel computational systems. In those cases, judges repeatedly found that algorithmic technologies aligned with the Constitutional justification for copyright law: promoting progress.

But some GenAI outputs prevent progress by projecting biases. GenAI outputs are biased in part because they use biased, low friction data (BLFD) as training data, like content scraped from the public internet. Examples of BLFD include Creative Commons (CC) licensed works, like Wikipedia, and works in the public domain. While Wikipedia is used as training data in most AI systems, its articles are overwhelmingly written by men–and that bias is reflected in shorter and fewer articles about women. And because the public domain cuts off in the mid-1920s, those works often reflect the harmful gender and racial biases of that time. However, if authors allow their copyrighted works to be used as GenAI training data, those authors can help mitigate some of the biases embedded in BLFD. 

Current biases in GenAI are disturbing. As we discuss in our zine, word2vec is a very popular toolkit used to help machine learning (ML) models recognize relationships between words–like women as homemakers and Black men with the word “assaulted.” Similarly, OpenAI’s GenAI chatbox ChatGPT, when asked to generate letters of recommendation, used “expert,” “reputable,” and “authentic” to describe men and  “beauty,” “stunning,” and “emotional” for women, discounting women’s competency and reinforcing harmful stereotypes about working women. An intersectional perspective can help authors see the compounding impact of these harms. What began as a legal framework to describe why discrimination law did not adequately address harms facing Black women, it is now used as a wider lens to consider how marginalization affects all people with multiple identities. Coined by Professor Kimberlé Crenshaw in the late 1980s, intersectionality uses critical theory like Critical Race Theory, feminism, and working-class studies together as “a lens . . . for seeing the way in which various forms of inequality often operate together and exacerbate each other.” Contemporary authors’ copyrighted works often reflect the richness of intersectional perspectives, and using those works as training data can help mitigate GenAI bias against marginalized people by introducing diverse narratives and inclusive language. Not always–even recent works reflect bias–but more often than might be possible currently.

Which brings us back to fair use. Some corporations may rely on the doctrine to include more works by or about marginalized people in an attempt to mitigate GenAI bias. Professor Mark Lemley and Bryan Casey have suggested “[t]he solution [to facial recognition bias] is to build bigger databases overall or to ‘oversample’ members of smaller groups” because “simply restricting access to more data is not a viable solution.” Similarly, Professor Matthew Sag notes that “[r]estricting the training data for LLMs to public domain and open license material would tend to encode the perspectives, interests, and biases of a distinctly unrepresentative set of authors.” However, many marginalized people may wish to be excluded from these databases rather than have their works or stories become grist for the mill. As Dr. Anna Lauren Hoffman warns, “[I]nclusion reinforces the structural sources of violence it supposedly addresses.”

Legally, if not ethically, fair use may moot the point. The doctrine is flexible, fact-dependent, and fraught. It’s also fairly predictable, which is why legal precedent and empirical work have led many legal scholars to believe that using copyrighted works as training data to debias AI will be fair use–even if that has some public harms. Back in 2017, Professor Ben Sobel concluded that “[i]f engineers made unauthorized use of copyrighted data for the sole purpose of debiasing an expressive program, . . . fair use would excuse it.” Professor Amanda Levendowski has explained why and how “[f]air use can, quite literally, promote creation of fairer AI systems.” More recently, Dr. Mehtab Khan and Dr. Alex Hanna  observed that “[a]ccessing copyright work may also be necessary for the purpose of auditing, testing, and mitigating bias in datasets . . . [and] it may be useful to rely on the flexibility of fair use, and support access for researchers and auditors.” 

No matter how you feel about it, fair use is not the end of the story. It is ill-equipped to solve the troubling growth of AI-powered deepfakes. After being targeted by sexualized deepfakes, Rep. Ocasio-Cortez described “[d]eepfakes [as] absolutely a way of digitizing violent humiliation against other people.” Fair use will not solve the intersectional harms of AI-powered face surveillance either. Dr. Joy Buolamwini and Dr. Timnit Gebru evaluated leading gender classifiers used to train face surveillance technologies and discovered that they more accurately classified males over females and lighter-skinned over darker-skinned people. The researchers also discovered that the “classifiers performed worst on darker female subjects.” While legal scholars like Professors Shyamkrishna Balganesh, Margaret Chon, and Cathay Smith argue that copyright law can protect privacy interests, like the ones threatened by deepfakes or face surveillance, federal privacy laws are a more permanent, comprehensive way to address these problems.

But who has time to wait on courts and Congress? Right now, authors can take proactive steps to ensure that their works promote progress rather than prevent it. Check out the Authors Alliance’s guides to Contract Negotiations, Open Access, Rights Reversion, and Termination of Transfer to learn how–or explore our new zine, Putting the AI in Fair Use: Authors’ Abilities to Promote Progress.

You can find a PDF of the Zine here, as well as printer-ready copies here and here.