Hello, and welcome to "Ethics of AI." I'm here to share my perspective on the ethics of using AI within libraries. No facts or opinions shared during this presentation represent my employer.

I would like to note, I am not the first librarian to explore the ethics of AI through the lens of the ALA Code of Ethics. Part way through the drafting this presentation I realized it pretty closely mirrored the structure of Violet Fox's "A Librarian Against AI or I Think AI Should Leave" zine which provides a good, and often entertaining, overview of various ethical concerns about AI.¹

Before we begin in ernest, I'd like to define artificial intelligence, or AI, what it is and what it isn't.

AI is, above all else, a marketing term.² Since the mid-1950s, AI has been applied to a wide variety of technologies and fields of study, such as machine learning, natural language processing, neural networks, image recognition, transformers, deep learning, rules-base systems, machine translation, search engines, production systems, robotics, expert systems, recommendation engines, video games, information filtering systems, and autonomous vehicles. Some of these technologies are closely related, some have very little in common with others. What they do all have in common is they are attempts, to paraphrase John McCarthy, who coined the term artificial intelligence, to make machines simulate intelligence.³

From its first use to the present day, AI has been used to make big promises but few deliveries. McCarthy coined the term in a proposal for a workshop where he expected that he and nine of his colleagues could make "a significant advance" in making machines simulate intelligence in only two months. They didn't.⁴

This initial workshop set the pattern for AI ever since. Throughout the 1960s and early 1970s, artificial intelligence research garnered significant funding. Researchers made grandiose statements, but failed to deliver promised results.⁵ The failure to deliver on promises led to reduced federal research funding in the mid-1970s.⁶ In the wake of this first boom and bust came an even larger one, where AI would see commercial breakthroughs in the early 1980s only to have research funding and commercial interest largely disappear by the later part of the decade.⁷ We are now in the midst of another boom.

The failures of AI in the past does not mean that no useful technologies came out of such research, just that what has emerged has usually been more modest than what marketers promised or come far later. Some truly useful technological advances may come out of this current boom, but they almost certainly won't live up to the promises AI vendors are making.

An important aspect of information literacy is being able to spot misinformation and promises that are too good to be true. It's my hope that a small side effect of this talk will be to provide some tools to allow those who are not deeply immersed in tech to more accurately evaluate the claims made around AI.

While I outlined a broad, and incomplete overview of various technologies marketed as AI, this presentation will focus on one particular kind: generative AI.

I've chosen to focus on generative AI because that is the technology being most heavily marketed toward consumers, businesses, and academia. Generative AI has become so ubiquitous over the last few years that it's what most people mean when they say "AI." I will likewise treat "AI" and "generative AI" as interchangeable in this talk. I will be focusing primarily on text-based generative AI programs, also known as large language models or LLMs, software like ChatGPT, Gemini, and CoPilot.

Generative AI is a label used for various programs that are able to output text, images, audio, or video, based on large amounts of human-created text, images, audio, or video. This includes programs like ChatGPT, Microsoft CoPilot, DALL-E, Stable Diffusion, MusicLM, Sora, and others. It's also the technology behind the creations of deepfakes, seemingly real audio, images, and video, that have been used to interfere in elections, spread misinformation, and generate nonconsensual pornagraphic images and videos of adults and children.⁸

On the surface, generative AI programs can seem impressive and even human. This effect is enhanced by the massive sums of money that the AI industry has spent on advertising in the past few years to promote the idea that AI is a near-magical solution to all our problems.⁹ But much of what makes AI programs seem so impressive is due to the massive amount of human labor behind them.¹⁰ As AI researcher Adio-Adet Dinika put it, "AI isn’t magic; it’s a pyramid scheme of human labor."¹¹ From the vast body of creative work that is used to build their input data, to the software engineers that design them, to the countless people who work to label data and rate responses before a program is released to the public, the ability of AI programs to simulate intelligence relies entirely on human labor and actual human intelligence. Without massive amounts of human input, these programs wouldn't even function. What we get when ChatGPT responds to a question isn't so much the output of a sophisticated algorithm as it is the mangling of an almost immeasurable amount of human labor. In the end, despite the promises of corporate propaganda, generative AI programs are little more than plagiarism machines and statistical word spitters.¹²

For the discussion of ethics, I'll be leaving aside deepfakes. Since their uses range from the merely plagiaristic to the personally and socially destructive, I believe their ethical problems need no further explication. Instead I'll be exploring how more apparently benign uses of generative AI, like information seeking, productivity boosting, and personal amusement, relate to the ALA Code of Ethics.

The ALA Code of Ethic's first point states:

1. We provide the highest level of service to all library users through appropriate and usefully organized resources; equitable service policies; equitable access; and accurate, unbiased, and courteous responses to all requests.¹³

So can we, as library workers, use generative AI to provide the highest level of service and provide our communities with accurate and unbiased information? I imagine most of you have heard about AI hallucinations. "Hallucinations" is the misleading term used by AI researchers to describe inaccurate results, what anyone in any other field would simply call errors. The use of hallucination is misleading because it implies the error was caused by a malfunctioning mind when no mind exists, only statistics. It also inaccurately suggests that the error is an aberration rather than a result of an AI program working as designed.¹⁴

No generative AI program is designed to or capable of providing only factually accurate responses. They can't be. They have no capacity to understand or reason.¹⁵ They're designed to provide responses that appear correct. Sometimes they do provide factual responses. The process by which they do so is identical to the process they use to provide fictional ones. The accuracy or inaccuracy of a response is happenstance, a byproduct of how responses are generated. Moreover, much of the meaning perceived in AI output comes from users reading meaning into statistically generated text.¹⁶

That generative AI programs have no means to understand user input or the content of their output, does not in and of itself mean their output is not usable. Perhaps the designers of these programs have crafted statistical systems so sophisticated that accuracy in AI output is a mathematical certainty. They didn't and it isn't. Generative AI programs frequently produce inaccurate answers, often presenting their misinformation in confident language, while often failing to cite their sources. When researchers Klaudia Jaźwińska and Aisvarya Chandrasekar compared OpenAI’s ChatGPT Search, Perplexity, Perplexity Pro, DeepSeek Search, Microsoft’s Copilot, xAI’s Grok-2 and Grok-3 (beta), and Google’s Gemini they found that each service frequently provided incorrect answers with ChatGPT, Deepseek Search, Copilot, both versions of Grok Search, and Gemini providing far more incorrect answers than correct ones.¹⁷ Another study found that when asked about federal court cases, OpenAI's ChatGPT 4's output was incorrect 58 percent of the time and Facebook's Llama 2 was incorrect 88 percent of the time.¹⁸ Inaccurate responses are so common, it's hard to keep up with all of them. Whether it's lawyers including fabricated quotes and cases in court documents¹⁹ or AI programs providing dangerous medical advice, even in clinical settings,²⁰ or Google telling people to put glue on pizza and to eat rocks,²¹ news of AI's untrustworthy output is nearly unavoidable.

Even apparently simple tasks, like summarizing documents can be exceedingly difficult for generative AI programs. A test of ChatGPT, Microsoft Copilot, Google Gemini, and Perplexity, found that 91 percent of responses had some issue: slightly more than half of the answers had significant issues, 19 percent of them had factual errors and 13 percent had manufactured or mangled quotations.²² In evaluating different AI programs for this purpose, the Australian Securities and Investments Commission found that, of the three programs tested, the highest performing one only managed to score 60 percent on their test criteria, what would earn any school child a D-. When they tested this program against human-written summaries it scored worse across all five criteria tested, even after it had been optimized for the task.²³ The problem is bad enough that even OpenAI, the company responsible for ChatGPT, published a study that found incorrect responses are inevitable given current implementation practices.²⁴ This finding echoes a study published a few months prior that characterized factually incorrect output as an "innate limitation" of large language models.²⁵

In the library context, AI inaccuracy is, unsurprisingly, just as much of a problem as it is in the broader world. Ingrid Reiche found in a study of two different open source applications for creating digital image metadata that neither tool was capable of producing accurate metadata on its own, necessitating human intervention.²⁶ This echoes a study from the University of Utah that found providing accurate, detailed, input data that included proper nouns dramatically reduced AI accuracy. That is, the high-quality, human-produced metadata from their digital collections made their AI program worse. The kind of metadata that enhances discoverability and provides useful information to researchers is something their AI program couldn't handle. They used a script to strip proper nouns from the input data and fed this genericized data into their program. The only evidence they presented to demonstrate their model functioned at all was a single set of keywords for a single image. Of the eight keywords provided, only one accurately described anything in the image. According to their abstract, this one-for-eight hit rate for a single image only cost $34,700 to develop.²⁷ Sai Deng explored using AI programs to generate MARC catalog records and demonstrated that AI programs from Google, OpenAI, and Microsoft all produced numerous errors.²⁸ Ex Libris's own documentation for the "AI assistant" built into the Alma Metadata Editor, notes that the "assistant" program returns "inaccurate or generic data," doesn't "follow cataloging standards," doesn't reliably "return all requested metadata fields," fails to accurately process "some languages," doesn't "return correct subjects for certain authority vocabularies," and cannot always process images.²⁹

This inability to produce accurate results is bad enough even proponents of using AI in libraries note that AI chatbots have trouble handling all but the most basic of inquiries and that ChatGPT's output is often "inaccurate to the point of complete fabrication."³⁰ If even its boosters openly admit how wildly inaccurate generative AI is, it hardly seems like something that would benefit institutions that are trusted to be repositories of knowledge. Instead using AI to provide information would lower the quality of our service and prevent us from providing accurate, unbiased information.³¹

Point 2 in the Code of Ethics states:

2. We uphold the principles of intellectual freedom and resist all efforts to censor library resources.³²

Being able to learn about and use generative AI fits well within a person's right to intellectual freedom and within the ALA Code of Ethics.

That said, AI does have some potential to harm individuals' intellectual freedom. Catherine Smith notes that the use of AI in description and discovery can potential contribute to a false information landscape for users.³³ She further notes that how many AI programs function is unknown or unclear due to vendor secrecy or the way their workings are often portrayed as nearly magical and too complicated for most people to understand.³⁴ This obscurity can make it hard to evaluate how such programs may or may not impact the information environments they build for the communities we serve.

3. We protect each library user's right to privacy and confidentiality with respect to information sought or received and resources consulted, borrowed, acquired or transmitted.³⁵

When it comes to privacy, long a cornerstone of library service, AI has a checkered history. AI vendors appear to have little to no interest in protecting user privacy.³⁶ In 2024, Microsoft began to secretly collect LinkedIn user data without prior consent.³⁷ Microsoft has also been developing programs like Copilot Recall and Copilot Vision that capture sensitive information including credit card numbers and social security numbers and share them with Microsoft.³⁸ In May 2025, generative AI vendor Luka was fined 5 million euros by Italy for illegally collecting user data via its Replika program.³⁹ Meta's AI app made it easy for users to accidentally share their private usage of the chatbot to their public Facebook timelines.⁴⁰ This year, ChatGPT has begun monitoring peoples chats in order to facilitate turning chats over to law enforcement.⁴¹ Many users of ChatGPT and Grok have had seen their prompts become part of the public internet and showing up in search results.⁴² This is just a small sampling of the privacy issues that have come to light in the last few years.⁴³

The general refusal of AI programs to respect user privacy is viewed by some who promote its use in library services as a benefit. In an article promoting the use of AI in reference services, Md. Ashikuzzaman cites being able to collect and analyze user data as one of the primary benefits of employing an AI chatbot.⁴⁴ Fatouh and Hamam likewise see ignoring patron privacy in favor of conducting "in-depth analysis of user interactions with chatbots" to "gain meaningful insights into patron behavior, preferences, and information needs" as a benefit of using chatbots in libraries.⁴⁵

Privacy concerns can be mitigated by only using AI programs that are vetted for proven privacy protections, especially those that can be run by a library itself and don't rely on vendors uninterested in protecting privacy.

4. We respect intellectual property rights and advocate balance between the interests of information users and rights holders.⁴⁶

The different generative AI programs all have some features in common. One, they require an enormous amount of input data. In the cases of text-based programs like ChatGPT, that means a huge volume of words written by people: news articles, books, Wikipedia, and so on. For image-based programs it means a vast corpus of human-made art and photographs. Two, they use statistics to generate output, using elements pulled directly from their input data that are then reassembled as output appropriate for the provided prompt. At their core, generative AI programs are plagiarism engines that use statistics to recombine human input into computer output.⁴⁷ Often the plagiarism is a pastiche of inputs, but other times it is unattributed verbatim or near verbatim quotes from sources used in the input data.⁴⁸ The New York Times, in the lawsuit they filed against Microsoft and OpenAI thoroughly documents GPT-4, a program used by both companies, reproducing verbatim passages of Times stories.⁴⁹

If all the text and images used as input data for various AI programs were legally acquired, used with permission, and its creators were fairly compensated, then the fact that the output of generative AI is entirely a regurgitation of its input data might seem like a minor complaint. Instead, the input data is a combination of freely available and usable sources and wholesale copyright and licensing violations.⁵⁰

To populate the AI responses in its search results Google takes content from publishers without their permission and chose to give them only limited options to opt out according to an internal document.⁵¹ Danielle Coffey, CEO of News/Media Alliance characterizes this practice as "theft."⁵²

AI companies refusal to respect existing intellectual property laws has resulted in lawsuits being filed against OpenAI, Anthropic, Facebook/Meta, Ross Intelligence, Stability AI, Udio, ElevenLabs, Perplexity AI, Suno, MiniMax, Microsoft, GitHub, and Cohere.⁵³ On September 25, Anthropic received preliminary approval for 1.5 billion dollar settlement due to its use of pirated books as input data for its AI program Claude.⁵⁴

Given commercial AI's problem with plagiarism and piracy, thorough vetting is necessary to verify that any such products used by libraries are built using only legally-acquired input data. Given the previously addressed problems with accuracy, there's no way to ensure proper citations from existing commercial AI programs. Their use in libraries then, results in a tacit endorsement of plagiarism.

If libraries build their own AI programs using only input data they have a right to, they can avoid most of the intellectual property issues inherent in commercial AI. However, the problem of plagiarizing prior work, often without proper attribution, remains.

5. We treat co-workers and other colleagues with respect, fairness, and good faith, and advocate conditions of employment that safeguard the rights and welfare of all employees of our institutions.⁵⁵

One of the purposes of AI is to replace human workers.⁵⁶ This usage is hitting young workers hard as many employers have replaced entry-level positions with AI.⁵⁷ Companies, including Klarna, UPS, DuoLingo, Intuit, and Cisco, have pivoted to AI while laying of tens-of-thousands workers.⁵⁸ AI is also being used to replace thousands of Information Technology (IT) and Human Resources (HR) workers at companies including Microsoft, IBM, and Walmart.⁵⁹ This summer, Amazon CEO Andy Jassy announced the company's plan to replace workers with AI⁶⁰ The federal government is also exploring replacing workers with generative AI.⁶¹

Libraries are not immune to this approach. Corey Halaychik encourages us to "celebrate" AI replacing library jobs. He encourages using AI for technical services and cataloging, where its impact will be invisible to the public, while avoiding its use in public-facing services because that "looks better." Although he makes grandiose claims about what AI can accomplish, Halaychik presents no evidence to support those claims.⁶²

Md. Ashikuzzaman suggests using chatbots for reference work would be a "cost-saving measure" by reducing "the need for additional personnel to handle reference services."⁶³ Of course to make the added expense of an additional software system and the IT costs to maintain it and integrate it with library services cheaper than not adding those expenses requires cutting staff or, at best, not rehiring vacated positions.

Carlo Iacono envisions a future for libraries so different that many traditional library jobs, and implicitly the workers that fill them, will go by the wayside. Where traditional jobs remain, they are greatly reduced in his vision. Gone are reference and instruction staff. In their place are so called "AI Literacy Specialists," whose only purpose is to guide other's usage of AI.⁶⁴ Of course, it's possible in this imagining that libraries could retrain all their existing workers into new roles, assuming, of course, that that makes budgetary sense.

When AI is not touted as a way to replace workers, it's promoted as a means of improving worker efficiency and freeing them from drudgery. It hasn't yet been shown to do either, but it can do the opposite. Despite widespread adoption,⁶⁵ generative AI has yet to demonstrate measurable productivity boosts.⁶⁶ A report from MIT put it starkly, "Despite $30–40 billion in enterprise investment into GenAI, this report uncovers a surprising result in that 95% of organizations are getting zero return."⁶⁷

A study from the Harvard Business Review posits that one reason for the failure may be that many workers use AI to generate low-quality work, or "workslop." This workslop often requires coworkers to work harder to rewrite it or even understand its meaning and can reduce productivity. Workslop reduces other's opinion of coworkers that use it and reduces their desire to collaborate with those coworkers in the future.⁶⁸ Even when it's not being used to produce workslop, the assistive approach to using AI in workplaces can result in skilled workers being replaced by lower-paid workers fulfilling the same duties while babysitting generative AI programs.⁶⁹

Within the library world, the picture appears much the same. Even boosters of using AI in libraries often demonstrate its troubling shortcomings. In a paper supportive of using ChatGPT for library metadata, Jenny Bodenhamer found that its attempts to generate call numbers and subject headings were riddled with errors, although she found its ability to extract keywords satisfactory.⁷⁰ Halaychik argues that using AI for public-facing services could "negatively impact" public services and should be avoided.⁷¹ Since their output can't be trusted, AI programs that generate metadata would require human workers to verify and correct its output to ensure accuracy. Such a workflow reduces a skilled worker to a babysitter for a misbehaving machine while still requiring specialized knowledge to ensure correct call numbers and subject headings.

Beyond the direct negative effects AI can have on workers, the constant barrage of messaging telling us AI will take our jobs or will transform our worklives, privacy concerns around AI, the use of AI to replace human interaction, the rise of AI-generated misinformation, and concerns about bias all contribute to increased anxiety among workers, further degrading workplace morale.⁷²

Instead of improving working conditions, generative AI more often demoralizes and devalues human workers while adding to their drudgery and stress.

Beyond our own institutions, artificial intelligence companies rely on exploitative labor practices to generate and prepare data for input into AI programs and to rate and refine the output of their AI programs. How AI companies treat their workers may seem beyond the scope of the ALA Code of Ethics, but if we bring these technologies into our workplaces, then the people that built them become our colleagues. Our labor, which is fully within the scope of the Code of Ethics, becomes part of their labor. If you think I'm venturing a little far afield, I hope you'll indulge me all the same. I will get to one tech giant whose labor practices include extracting unpaid labor out of everyone in this room.

Generative AI programs require enormous amounts of input data, which usually needs to be labeled by people so the statistical models at the core of these programs have values to weight different pieces of input by. This work is often done by people paid well below poverty rates. Some rate image labels through Amazon's Mechanical Turk. Others are refugees, prisoners, and others in positions of precarity around the world. All of these people's work is absolutely vital to the functioning of the AI industry. They are the foundation upon which the entire generative AI industry is built upon. Yet, despite the crucial role they play they are paid appallingly low wages while often having to endure horrid work environments.⁷³ When these workers seek better pay or work conditions, they risk losing their jobs, as has happened on at least two occasions to workers contracted to work on Google's AI products.⁷⁴

Many companies, including those working in other areas of AI, also rely on fully uncompensated labor by people who are often unaware that they are working to label data for AI or otherwise generating data for AI-related use. This includes children, prisoners, and refugees.⁷⁵ It also includes everyone that's been forced to use Google's reCAPTCHA service to prove your humanity on some website.⁷⁶ Whether you're aware of it or not, whether you consent to it or not, you are forced to become an unpaid Google employee every you access some basic website function hidden behind a reCAPTCHA.

6. We do not advance private interests at the expense of library users, colleagues, or our employing institutions. 7. We distinguish between our personal convictions and professional duties and do not allow our personal beliefs to interfere with fair representation of the aims of our institutions or the provision of access to their information resources.⁷⁷

When considering AI, we must base our decisions and discussions of AI firmly in facts and ensure our opposition to or support of their use is inline with the rest of the Code of Ethics.

8. We strive for excellence in the profession by maintaining and enhancing our own knowledge and skills, by encouraging the professional development of co-workers, and by fostering the aspirations of potential members of the profession.⁷⁸

When it comes to enhancing our knowledge and skills, staying informed on developments in AI is certainly in line with the ALA Code of Ethics. While doing so, of course, it is important to make sure we distinguish facts from misleading marketing hype.

9. We affirm the inherent dignity and rights of every person. We work to recognize and dismantle systemic and individual biases; to confront inequity and oppression; to enhance diversity and inclusion; and to advance racial and social justice in our libraries, communities, profession, and associations through awareness, advocacy, education, collaboration, services, and allocation of resources and spaces.⁷⁹

Generative AI programs have an extensive and well-documented problem with racial and gender biases.⁸⁰ They promote harmful racial and gender stereotypes around what kind of people hold what kind of job and what race "criminals" are, and hypersexualizeswomen.⁸¹ They do so even based on the dialect a person speaks. Current techniques to reduce racial bias can worsen the problem by obscuring, but not removing the racist weighting an AI program uses.⁸² Racial bias in AI appears to get worse the more input data they have.⁸³

The bigotry built into commercial AI sometimes becomes apparent in dramatic ways, like when chatbots start spewing racist, misogynistic, and anti-Semitic language as has happened with Microsoft's Tay and xAI's Grok.⁸⁴

Other times, it makes itself apparent in subtler, but more damaging ways, like when AI is used to discriminate in hiring practices,⁸⁵ as it has on more than one occasion. Amazon scrapped an AI recruiting program that exhibited bias against female job candidates.⁸⁶ iTutorGroup settled a lawsuit in 2023 over its AI recruiting software filtering out female job candidates over the age of 55 and male candidates over the age of 60.⁸⁷

AI bias also shows up when people seek medical treatment.⁸⁸ Studies have found that AI used in clinical settings and the data used to train them both demonstrate racial bias.⁸⁹ One such study found that AI used in mental health treatment recommended inferior treatment for Black patients.⁹⁰ Another recent study found similar bias against female patients. The AI programs they studied treated male patients more consistently while they were much more likely to erroneously reduce care for female patients.⁹¹

The combined racism and misogyny embedded in generative AI programs means that BIPOC women are subject to a double dose of discrimination.

As long as racial, gender, and other biases remain in AI programs, we cannot endorse or promote their use within libraries. To do so violates point nine of the Code of Conduct and would inhibit the ability of affected members of our community to fully exercise their right to intellectual freedom within our libraries. Morally and ethically it would take us back to the days, not that long ago, when white librarians across the US proudly enforced segregation in their libraries.⁹² It would take us back to the very ethical stance that point 9 was so recently adopted to address.⁹³

There are various ethical concerns around the use of generative AI programs beyond the scope of the Code of Ethics that are worth considering. I will briefly touch on a few of them.

Some studies have shown that use of generative AI can have negative effects on learning outcomes and cognition.⁹⁴ One study comparing essay writers who used AI to ones who didn't found that the AI user group demonstrated reduced neural connectivity and activity, reduced memory performance, had a reduced sense of agency in the writing process, and wrote more homogenous essays than the groups that did not use AI.⁹⁵ A study of high school students using AI tutors found that those who used a standard GPT chatbot performed worse on tests when they couldn't use the program compared to students who had no AI assistance. Those who used a GPT chatbot specifically designed for educational use showed a statistically insignificant performance reduction compared to the control group.⁹⁶ In the medical field, several studies have found that reliance on AI programs can reduce the skill of medical professionals, including making it harder for them to correctly identify cancer.⁹⁷ As knowledge workers in institutions of learning, we have a responsibility not to promote tools that inhibit learning.

There have also been instances where use of generative AI has exacerbated or possibly caused mental health problems.⁹⁸ In the case of two teenagers who died by suicide, ChatGPT and a Character.AI chatbot have been accused of encouraged teens to kill themselves and of helping them plan their suicides.⁹⁹ A custom chatbot used by the National Eating Disorders Association gave users weight loss advice, potentially worsening their existing eating disorders.¹⁰⁰ A man in Connecticut killed himself and his mother after months of having his paranoid delusions confirmed and reinforced by ChatGPT.¹⁰¹ The danger this poses only increases as more people are turning to generative AI for therapy, a task that such programs are woefully unsuited for as they tend to reinforce delusions and stigmatize mental health problems.¹⁰²

Beyond the direct human impacts of AI we also need to consider its environmental impacts. As the effects of human-driven climate change continue to become more severe we need to be more conscious than ever about the environmental impact of the technology we use. Generative AI programs are wildly inefficient and require enormous amounts of energy to function. For answer seeking tasks, AI requires far more energy than search engines.¹⁰³ The voracious energy consumption of generative AI programs has led to a massive increases in carbon emissions.¹⁰⁴ This unfettered energy consumption requires billions of gallons of water to cool the machines powering commercial AI programs and threatens the water supply for communities that live near AI-hosting data centers.¹⁰⁵ To use commercially available AI in libraries or to promote its use is to participate in environmentally devastating business practices that worsen living conditions for people around the globe.

Using commercial AI in libraries also means doing business with companies that have long histories of unethical business practices. As previously detailed, exploitative labor practices and disregard for intellectual property rights and user privacy are endemic to the industry. These issues alone put these companies and products at odds with our professional ethics. Additionally, the industry has a documented history of making misleading and exaggerated statements about their products that the uncharitable could characterize as fraud.¹⁰⁶ Do we really want to use our institutions to launder the reputations of companies like these all for inferior products that don't work as advertised?

Given how deeply at odds with our professional ethics most existing AI programs are, we face great difficulties in using them without violating those ethics. There may be ways to use or create generative AI programs that are ethical, but utmost caution must be taken at every step of their development and evaluation to avoid repeating the ethical pitfalls of existing options. Often, it may be better to heed the words of Rea N. Simons and recognize the wisdom that sometimes inaction is the best path.¹⁰⁷ If we cannot use AI the right way, it is better we do not use it at all rather than collaborate in the harm it causes. Does this mean we'll miss out on the AI revolution? Maybe, but since generative AI has yet to produce any demonstrable benefits, we're not missing out on much.