When Knowledge Loses Its Footnotes

Citations Built Trust. AI Is Testing It.

AI can produce writing that looks credible without showing its sources. Have you used AI to help you with your writing, and then tried to trace where its claims came from? If you’re clever with your prompting, you ask for citations to be included. It will often oblige. But on closer inspection, you realize some of those citations were misleading, irrelevant, or entirely fabricated. What you’re left with is writing that sounds strong, looks well supported, but is actually built on a house of cards. 

This isn’t just a quirk of the tool. Studies are already showing that hallucinated citations are making their way into published research, where they can be reused and repeated, polluting the knowledge ecosystem over time. A domino effect of junk citations, if you will. 

What Citations Actually Do for Us

Let’s back up a second and look at the importance of citations. Citations are so embedded in academic and professional writing that we rarely stop to ask why they exist. But they do more work than we usually give them credit for.

1. Making Thinking Visible

Citations are often treated as a requirement, something added at the end of writing to meet a standard. But they do something much more important. They make the thinking behind a piece visible.

A citation signals where an idea came from, how it was shaped, and what it’s in conversation with. At their best, they reveal the process. They show what was selected, what was read, and how ideas were brought into conversation with one another. A citation is a small marker of a much larger act of judgment.

They give the reader a way to trace the thinking, to see not just what is being said, but how it was built: from selecting sources, evaluating them, and deciding how they fit into an argument. A path can be followed and, if needed, questioned or challenged. These steps are largely invisible in AI-generated writing. 

2. Creating Structure and Consistency

Citations don’t just support individual claims. They create shared structure. Citation styles (APA, MLA, Chicago, etc.) standardize how knowledge is presented, making it easier to read, compare, and connect across sources.

This structure allows information to move between systems—across articles, databases, and disciplines—without losing its shape. It’s structured trust, which matters when working with structured data.  

3. Providing Pathways

Citations give readers a way to move beyond a single piece of writing. They point outward, allowing readers to follow threads, explore sources, and situate ideas within a broader network of knowledge.

A well-cited work isn’t a dead end. It’s a knot along a longer string. 

4. Supporting Accountability and Ethical Practice

Citations distribute responsibility. They make it possible to trace claims back to their sources, to challenge them, and to correct them over time. They also serve the ethical function of giving credit to other writers.

Citations make knowledge not just shareable, but inspectable. Citations are infrastructure for trust, not just attribution.

Citation Errors Aren’t New

Of course, citation errors aren’t new in the publishing world. Humans make errors frequently in mistranscribing a DOI or citing the wrong source. 

Additionally, citations can be selective, inconsistent, and sometimes performative (how many of us can admit to using a citation that we didn’t fully read?) But they still provide something essential: a visible path back to source material.

AI is changing not only the speed, but the structure

AI isn’t just accelerating writing, it’s restructuring how knowledge is produced. In a traditional writing or publishing model writers read/consume/interpret a body of knowledge, synthesize it into something new, and produce a new work, often a piece of writing with citations. 

Now, AI models are taking a prompt, aggregating data, predicting an outcome, and generating a new work. AI does not retrieve and cite sources in the way a human writer does. It generates text based on patterns, including citations that look plausible but may not correspond to real or relevant sources. The result is writing that sounds grounded, without showing its ground. It’s math that doesn’t show its work. 

What Gets Lost

When we strip away citations, we lose more than just our own ability to synthesize ideas and create something new. We lose history of thought, we lose provenance (more on that later). Sources become obscured, and the lineage of thought is no longer traceable in a meaningful way. 

And when inaccurate or fabricated citations are introduced, they don’t just fail to support a claim—they risk being repeated and embedded, gradually distorting the knowledge they were meant to support. 

At the same time, the distinction between knowing something and sounding like we know it begins to blur. AI can produce coherent, confident writing, but based on what? Accountability weakens. When claims are harder to trace, they are harder to challenge. The iterative process of critique and correction begins to break down.

This isn’t just a problem of misinformation. It’s a problem of untraceable information. And that has real implications. It becomes harder to evaluate credibility, harder to build on existing knowledge.

The AI Upside

I don’t want to sound totally alarmist. We know that they are upsides to AI writing and content creation. It can lower barriers to access to knowledge. AI can help us quickly draw connections across large bodies of data. It can help speed up a brainstorming, data collection, or early-stage thinking. 

Some are even using AI as a tool to detect these false or fabricated citations. An analysis by Nature used an AI tool (Veracity by Grounded AI) to check citations in articles submitted to them and flag “invalid” or “irrelevant” references. They estimate “more than 110,000 of the 7 million or so scholarly publications from 2025 contain invalid references.”

The irony of using AI to detect AI isn’t lost on me. 

What’s Next? From Citations to Provenance

If citations helped us trust knowledge in a slower, human system, what replaces them in a fast, generative one? The next challenge isn’t producing more information. It’s making that information traceable again.

Citations were designed for human-produced knowledge. They are static, selective, and often appended (like footnotes or a references list). AI requires something different. Provenance–the history, origins, and context of our information–needs to be embedded to help our content regain trust. 

Citations (naturally)

Karcher, Sebastian, and Philipp Zumstein. Citation Styles: History, Practice, and Future. Authorea, 2018, https://www.authorea.com/users/102264/articles/124920-citation-styles-history-practice-and-future 

Mikanovich, Troy. “AI Writing and Attribution: AI Cannot Cite Anything.” Graduate Writing Coach, University of Southern California, 21 Nov. 2023, https://sites.usc.edu/graduate-writing-coach/ai-writing-and-attribution-ai-cannot-cite-anything/ 

Naddaf, Miryam, and Elizabeth Quill. “Hallucinated Citations Are Polluting the Scientific Literature. What Can Be Done?” Nature, vol. 652, 2026, pp. 26–29. https://doi.org/10.1038/d41586-026-00969-z