Behind the Curtain: AI & You

# Behind the Curtain: AI & You Hello! Whether you have or haven't used one of the many different artificial intelligence interfaces available to us today, it's important to understand how these programs retrieve and provide information to you — and why you should always double-check the answers AI provides. AI makes mistakes, whether by hallucinating (or making up) information, or simply providing an answer to a different question altogether. However, there's more to checking an AI's answer than just finding out if it is right or wrong. Sometimes questions are open-ended, or maybe you were just trying to generate an image to share on social media. Unfortunately, even these seemingly general requests can produce answers that accidentally — or purposefully — contain bias you may not even be aware of. Let's break down what is happening behind the wizard's curtain that is an AI prompt. --- [[Pull back the curtain →->The Big Picture]]

# The big picture When you type a question into a search engine or an AI chatbot, it feels like magic. You ask, it answers. Simple. But behind that answer is a long chain of human decisions — and every single link in that chain is a place where bias can creep in. <div class="callout">Bias doesn't mean someone sat down and decided to be unfair. It often means that a system quietly reflects the assumptions, habits, and blind spots of the people who built it — without anyone noticing.</div> There are five main places this happens. We'll walk through each one. --- [[Let's start at the beginning →->Stage 1: The Data]]

<div class="stage-label">Stage 1 of 5 · Where the data comes from</div> # Not everyone is equally represented online Search engines and AI tools learn from content that already exists on the internet. But the internet is not an equal place. It overrepresents people with reliable internet access, English speakers, and those from wealthier countries. Meanwhile, many communities, languages, and ways of knowing are barely present at all — not because their knowledge is less valuable, but because the circumstances were never in place for it to be recorded and published online. When an AI learns from this uneven pool of information, it learns an uneven picture of the world. <div class="warning">Think of it this way: if you only surveyed people in one neighborhood to write a guide for the whole city, you'd miss a lot. Internet data has the same problem — just at a much bigger scale.</div> --- [[Tell me more about this →->Stage 1: Example]] [[Move on to Stage 2 →->Stage 2: The Training]]

<div class="stage-label">Stage 1 · Going deeper</div> # A real-world example Researchers studying AI language tools have found that they perform significantly better in English than in other languages — even languages spoken by hundreds of millions of people. Why? Because the training data was mostly in English. The AI didn't choose that. It simply learned from what it was given. This means that someone searching for medical information in Swahili, or asking a question in Cherokee, may get a much less accurate or complete answer than someone asking the exact same question in English. It also means that historical knowledge passed down through oral traditions, community archives, or non-digitized collections may be invisible to the AI entirely — not because it doesn't exist, but because the system was never designed to look for it. <div class="callout">Libraries are specifically built to preserve and provide access to materials that don't show up in a basic internet search. That's part of what makes them different.</div> --- [[Back to Stage 1 →->Stage 1: The Data]] [[Move on to Stage 2 →->Stage 2: The Training]]

<div class="stage-label">Stage 2 of 5 · Who teaches the system</div> # Humans train AI — and bring their own assumptions Before an AI can answer questions, people have to teach it what a "good" answer looks like. This process is called training, and it involves thousands of human judgment calls. Trainers might be asked: Is this search result relevant? Is this image appropriate? Is this sentence offensive? Each answer shapes how the AI behaves. When the group of people doing that training is not very diverse — in terms of background, culture, language, or life experience — those judgments can reflect a narrow set of views, even when everyone involved is trying to do a good job. <div class="warning">Think of it this way: it's like asking a small group of people to write the dictionary. Their word choices would tell you a lot about who they are — and whose words didn't make the cut.</div> --- [[Tell me more about this →->Stage 2: Example]] [[Move on to Stage 3 →->Stage 3: The Pattern]]

<div class="stage-label">Stage 2 · Going deeper</div> # A real-world example Studies have found that AI tools used in hiring sometimes ranked resumes from men higher than identical resumes from women — not because anyone programmed them to do that, but because they were trained on historical hiring data from companies where men had been hired more often. The AI learned a pattern: "people who got hired looked like this." Then it repeated that pattern, confidently, at scale. The humans who built the system didn't intend to discriminate. But the data they trained it on reflected decades of discrimination that had already happened — and the AI had no way to know that was a problem. <div class="callout">This is one of the reasons it matters to ask: who built this tool, and what were they trying to optimize for? The answer isn't always obvious.</div> --- [[Back to Stage 2 →->Stage 2: The Training]] [[Move on to Stage 3 →->Stage 3: The Pattern]]

<div class="stage-label">Stage 3 of 5 · How the system learns</div> # Patterns get repeated — and strengthened The AI looks for patterns in everything it was taught and repeats them. The problem is that it cannot tell the difference between a pattern that reflects the world as it really is and one that reflects how things have been unfairly skewed. Both get learned the same way. And here's the part that surprises most people: biased patterns don't just get preserved — they often get amplified. A small skew in the training data can become a strong, confident-sounding answer in the AI's output. <div class="warning">Think of it this way: if most historical photos of scientists show men, an AI trained on those photos may start to associate "scientist" with men — not because that's true, but because that's what it saw most often. Then it passes that assumption along to you, wrapped up in a confident-sounding answer.</div> --- [[Tell me more about this →->Stage 3: Example]] [[Move on to Stage 4 →->Stage 4: The Ranking]]

<div class="stage-label">Stage 3 · Going deeper</div> # A real-world example When researchers asked AI image generators to create pictures of "a CEO," "a nurse," or "a criminal," the results reflected strong stereotypes about gender and race — even though no one explicitly told the AI to do that. The AI had simply learned from millions of images on the internet, where those stereotypes already existed. It found the pattern and reproduced it — amplified, polished, and presented without any caveat. This matters beyond images. AI writing tools can produce text that assumes certain defaults — who a "typical" person is, what a "normal" family looks like, whose history is worth mentioning. These assumptions often go unnoticed because they match what we're already used to seeing. <div class="callout">The question to ask yourself: whose "normal" is this answer based on? And is that the same as my normal?</div> --- [[Back to Stage 3 →->Stage 3: The Pattern]] [[Move on to Stage 4 →->Stage 4: The Ranking]]

<div class="stage-label">Stage 4 of 5 · What gets pushed to the top</div> # Popular doesn't mean accurate Even after an AI or search engine has decided what results are relevant, it still has to decide what order to show them in. And that ranking is shaped by more than just quality. Results rise to the top because of how many times they've been clicked, whether a company has paid to promote them, how long they keep people on the page, and how well the page has been optimized to game the algorithm. A result at the top of the list may be there because it's well-funded or attention-grabbing — not because it's the most trustworthy or complete. <div class="warning">Think of it this way: a loud voice in a crowded room isn't necessarily saying the most important thing. It's just easier to hear.</div> --- [[Tell me more about this →->Stage 4: Example]] [[Move on to Stage 5 →->Stage 5: The Loop]]

<div class="stage-label">Stage 4 · Going deeper</div> # A real-world example Search engine optimization — often called SEO — is an entire industry built around making content rank higher in search results, regardless of whether it's better than what's ranked below it. A large company with a marketing budget can publish a mediocre article about a health topic and have it appear at the top of search results, while a carefully researched study from a small university press appears on page four — where most people never look. This same dynamic applies to AI tools that draw on web sources. They are often pulling from the loudest, most frequently cited, most commercially promoted content — not necessarily the most accurate. <div class="callout">Library databases work differently. The sources in them have been evaluated and selected by information professionals — not ranked by ad revenue or click counts.</div> --- [[Back to Stage 4 →->Stage 4: The Ranking]] [[Move on to Stage 5 →->Stage 5: The Loop]]

<div class="stage-label">Stage 5 of 5 · What the system learns about you</div> # Your search history shapes what you see next Many platforms track what you click on, how long you read it, and what you search for — and they use that information to personalize your future results. Over time, you see more of what you've already engaged with, and less of what might challenge or expand your thinking. This can happen so gradually you don't notice it happening. Two people searching the exact same words can get meaningfully different results — each quietly tailored to confirm what they already believe. <div class="warning">Think of it this way: it's like a bookstore that only ever recommends books you'd already like. You'd never stumble across something surprising — or something that changed your mind.</div> --- [[Tell me more about this →->Stage 5: Example]] [[I've seen all five — now what? →->So What Do I Do]]

<div class="stage-label">Stage 5 · Going deeper</div> # A real-world example Researchers studying social media and search platforms have documented what's sometimes called a "filter bubble" — the way personalization algorithms gradually narrow the range of information a person encounters. Someone who clicks on a few articles skeptical of a particular medical treatment may find that their future searches on the topic return more and more skeptical sources — even if the scientific consensus says something different. The algorithm isn't trying to mislead them. It's just giving them more of what they seemed to want. This is one of the reasons that the same question can lead different people to very different conclusions — and why it can sometimes feel like people are living in entirely different information realities. <div class="callout">One of the best ways to break out of a filter bubble is to use a source that isn't personalized to you at all — like a library database, a reference librarian, or a vetted print resource.</div> --- [[Back to Stage 5 →->Stage 5: The Loop]] [[I've seen all five — now what? →->So What Do I Do]]

# So what do I do with all of this? Knowing about bias in AI and search results doesn't mean you can never use them — it just means you use them more thoughtfully. Here are five questions worth asking any time you get an answer from a search engine or AI tool: - **Who wrote this?** Is there a named author, and what are their credentials or interests? - **Who is this written for?** Is this aimed at you, or at a different audience entirely? - **What might be missing?** Whose perspective, language, or experience isn't represented here? - **Why is this at the top?** Is it ranked highly because it's trustworthy, or because it's popular? - **Where else can I look?** Have you checked more than one source? You don't have to answer all of these every time. But making a habit of asking even one of them will make you a sharper, more confident researcher. --- [[How can my library help? →->Your Library]] [[Test what I've learned →->Quiz 1]]

# Your library can help Libraries are specifically designed to get around many of the problems we've talked about. Here's how we can help you find more complete, trustworthy information. **Use library databases.** Tools like JSTOR, ProQuest, and subject-specific databases are curated by information professionals — not ranked by clicks or ad revenue. The sources in them have been evaluated before they were included. **Ask a librarian.** We're trained to find sources that don't show up on the first page of a Google search — including local history, government data, and specialized archives. Reference help is free and there's no such thing as a question too small. **Try more than one source.** No single database or search engine covers everything. Comparing a few different sources — especially ones from different regions or perspectives — gives you a much fuller picture. **Look for peer-reviewed or vetted sources.** These have gone through a review process before publication, which adds a layer of accountability that most web content doesn't have. A librarian can help you tell the difference. <div class="callout">You don't have to navigate this alone. That's what we're here for.</div> --- [[Test what I've learned →->Quiz 1]] [[Start over →->Start]] [[View works cited →->Works Cited]]

# Quick check: let's see what stuck **Question 1 of 3** An AI image generator asked to show "a scientist" keeps producing images of men in lab coats. This is most likely because: --- [[The AI was programmed to prefer men →->Quiz 1: Wrong A]] [[The training data showed scientists as mostly men →->Quiz 1: Right]] [[Scientists are mostly men →->Quiz 1: Wrong B]]

**Not quite.** It's rarely the case that someone sat down and deliberately programmed a bias like this. More often, the bias is unintentional — baked into the training data without anyone realizing it. The AI learned from images that already existed on the internet, and those images reflected existing stereotypes. It didn't choose to do that. It just learned the pattern it was given. --- [[Try again →->Quiz 1]]

**Not quite — and this is worth pausing on.** While demographics in certain scientific fields are unequal, the bigger issue here is about what the AI *learned* to picture, not what the workforce actually looks like. Even in fields with more gender balance, AI tools trained on historical data can still produce skewed images — because history has been unequal, and the AI can't tell the difference between "how things were" and "how things are." --- [[Try again →->Quiz 1]]

**That's right.** The AI learned from images that already existed online, and those images reflected existing stereotypes. It didn't choose to do that — it just found a pattern and repeated it. This is why who creates the training data matters so much. The AI inherits the assumptions already present in whatever it learned from. --- [[On to question 2 →->Quiz 2]]

# Quick check: let's see what stuck **Question 2 of 3** A search result appears at the very top of the page. This most likely means: --- [[It has been reviewed by experts →->Quiz 2: Wrong A]] [[It is the most accurate result available →->Quiz 2: Wrong B]] [[It has been clicked often, promoted, or optimized for search →->Quiz 2: Right]]

**Not quite.** Expert review is not part of how most search engines decide what to rank highly. Peer review, editorial fact-checking, and professional curation happen at the *source* level — but search algorithms don't know or care whether a source has been through that process. High rankings are mostly about popularity, optimization, and in some cases, paid promotion. --- [[Try again →->Quiz 2]]

**Not quite.** This is one of the most common assumptions about search results — and one of the most important to unlearn. Ranking high means a result was popular, well-optimized, or paid to appear there. It says nothing about accuracy or completeness. Some of the most trustworthy sources end up buried on page three because they don't have the resources to compete with larger, better-funded websites. --- [[Try again →->Quiz 2]]

**Exactly right.** Search rankings are shaped by click counts, advertising, and search engine optimization — not by a panel of experts evaluating trustworthiness. High placement is a sign of visibility, not accuracy. This is one of the key reasons library databases are valuable: their rankings aren't driven by commercial interests. --- [[On to question 3 →->Quiz 3]]

# Quick check: let's see what stuck **Question 3 of 3** You and a friend search the same question and get different results. The most likely reason is: --- [[One of you made a typo →->Quiz 3: Wrong]] [[The search engine is personalizing results based on your histories →->Quiz 3: Right]] [[One of you has a faster internet connection →->Quiz 3: Wrong]]

**Not quite — though those things can matter a little.** The bigger factor here is personalization. Search platforms track your history and tailor your results to match your past behavior. Two people asking the same question can get meaningfully different answers based on what the algorithm has learned about each of them. --- [[Try again →->Quiz 3]]

**That's right.** Personalization is one of the most invisible forms of algorithmic influence. The results feel neutral — but they've already been shaped by everything you've clicked on before. This is one reason why it's worth occasionally stepping outside your usual sources and looking for information somewhere that doesn't know your history — like a library database or a reference librarian. --- [[See how I did →->Quiz End]]

# Well done for working through all of this. The goal here isn't to make you distrust every search result you ever see. It's to help you ask better questions — and to remind you that behind every AI answer and every search ranking, there are human decisions that shaped what you received. **The five things to remember:** 1. The data AI learns from is uneven — not everyone is represented equally. 2. The people who train AI bring their own assumptions, even when they try not to. 3. Patterns in data get learned and amplified, not just copied. 4. Popular results are not the same as accurate results. 5. Personalization quietly narrows what you see over time. <div class="callout">Your library is here to help you find information that search engines and AI tools might miss. Ask us anything — that's genuinely what we're for.</div> --- [[Start over from the beginning →->Start]] [[View works cited →->Works Cited]] [[Learn how your library can help →->Your Library]]

# Works cited The following sources informed the content of this presentation. All three are peer-reviewed and freely available online. --- **Ferrara, E. (2024).** Fairness and bias in artificial intelligence: A brief survey of sources, impacts, and mitigation strategies. *Sci, 6*(1), Article 3. https://doi.org/10.3390/sci6010003 This open-access survey provides a comprehensive overview of where bias originates in AI systems — including data collection, algorithm design, and human decision-making. It covers real-world examples in hiring, healthcare, and generative AI image tools, and examines how bias gets amplified rather than simply preserved. --- **Ilori, O. V., & Dube, L. (2024).** Navigating algorithm bias in AI: Ensuring fairness and trust in Africa. *Frontiers in Research Metrics and Analytics, 9*, Article 1486600. https://doi.org/10.3389/frma.2024.1486600 Written from a library and information science perspective, this article examines how algorithmic bias affects access to fair information — particularly in communities underrepresented in training data. It addresses the relationship between trust and AI, and the real-world consequences of biased systems in hiring, finance, and healthcare. --- **Bashir, M. A., Osei-Bonsu, A., Salami, T. O., & Amoah-Mensah, A. (2025).** Artificial intelligence algorithm bias in information retrieval systems and its implication for library and information science professionals: A scoping review. *Journal of Education for Library and Information Science*. https://doi.org/10.1080/07317131.2025.2512282 This scoping review draws on 76 peer-reviewed studies from 2019–2024 and is specifically addressed to library and information professionals. It examines how AI bias affects the credibility of information retrieval systems and what LIS practitioners can do about it. --- [[Go back to the beginning →->Start]] [[How your library can help →->Your Library]]