The Epistemic Stakes of Large Language Models

Large language models (LLMs) like ChatGPT are astonishing tools. They can draft essays, generate code, and carry on conversations that feel remarkably fluent. But their rise also raises urgent questions: what sort of knowledge, if any, do they provide? And what kind of responsibility do we take on when we use them?

Most public debates about AI fixate on ethics, safety, and jobs. These are vital concerns. But before we even get to those, there’s a deeper, more basic issue: epistemology, the philosophical study of knowledge and of what makes believing reasonable. In other words, we need to ask not just what AI does, but what it means to know something when it comes from AI.

Why Epistemology Matters

When you hear a fact from a teacher, read a book, or look through a microscope, you’re doing something epistemic: acquiring information that you treat as knowledge. Each of those practices has rules and responsibilities attached. A teacher should be competent and honest. A scientific tool should be reliable and transparent. And you, as the user, are expected to exercise judgment.

LLMs don’t fit neatly into any of these categories. They’re not people offering testimony, even though they sound like they are. And they’re not quite tools like calculators or thermometers, because they don’t just process information—you can’t point them at the world and get a measurement. Instead, they generate language by predicting the most statistically likely next word, based on patterns in their training data. That can be useful. But it means their output isn’t grounded in truth-seeking in the way we usually expect of epistemic sources.

The Mistake: Treating LLMs as Testimony

One of the easiest traps to fall into is treating LLMs as if they were reliable speakers. They are not. ¹

When ChatGPT says “Paris is the capital of France,” it is not “telling” you a fact in the way a teacher or a newspaper might. Instead, it’s stringing together words that tend to appear together in its training data. Most of the time, this gives you something true—but not because the system has access to the truth. The danger is that users, lulled by fluency and confidence, treat the system as if it were offering informed testimony, when in fact it is producing plausible text without regard to whether it is true.

That distinction matters. Human sources—teachers, books, newspapers—can certainly be mistaken, but their errors usually stem from lapses in memory, judgment, or accuracy within an underlying orientation toward truth. An LLM, by contrast, is not oriented toward truth at all: it generates text by following statistical patterns, indifferent to whether the result is correct. If you take such output at face value, you risk mistaking coherence for reliability, and the confident “hallucinations” it produces expose the misplaced trust of treating it as if it were testimony.

The Model: LLMs as Tools

A more accurate way to think about LLMs is to treat them as epistemic tools, like microscopes or calculators. Tools don’t speak; they generate outputs that you interpret. A microscope gives you an image, but you still need training to know what you’re looking at. A calculator gives you a number, but you need to understand the math to use it well.

Seen this way, LLMs are linguistic instruments: they give us sentences that reflect patterns in human discourse. That can help us brainstorm, rephrase, translate, or surface possibilities we hadn’t thought of. But the burden of interpretation rests with us. We can’t just accept the outputs as truths. We have to ask what role they play in reasoning and whether we’re using them responsibly.

The Stakes: Responsible Use

This reframing carries real stakes. If we treat LLMs as testifiers and hand them too much authority, we risk eroding our standards for knowledge. We lower the bar for what counts as evidence, expertise, or reasoning. Over time, that could corrode trust in public discourse itself.

But if we treat them as tools, we can fold them into responsible knowledge practices. That means cultivating epistemic virtues for using LLMs in particular:

• Interpretive literacy — understanding what kind of outputs these systems provide and what they don’t.
• Operational competence — knowing when and how to use them effectively (and when not to).
• Epistemic humility — resisting the temptation to over trust or outsource judgment.

Responsible use also means recognizing limitations. LLMs are trained on vast but partial data sets. They often reflect biases, amplify stereotypes, and lack transparency. Users need to ask critical questions about what information is missing, who controls the training data, and what purposes the system is serving.

The Upshot

LLMs aren’t useless. They’re extraordinary tools with real potential to expand human creativity and efficiency. But their epistemic role is unlike anything we’ve had before. They’re not trustworthy speakers, and they’re not neutral mirrors of reality. They’re pattern machines that generate plausible text.

Our challenge is to integrate them responsibly into our epistemic practices—without letting them hollow out our standards of knowledge. That requires distinct habits of judgment, literacy, and humility.

The epistemic stakes of LLMs are high because they cut to the foundation of how we know what we know. If we get thiswrong, the cost won’t just be bad AI systems. It will be a culture that forgets how to tell the difference between seeming right and being right.

‍

See Embedded Ethics in Action

Discover how Verizon worked with Northeastern’s Responsible AI team to integrate ethics directly into product development, policy, and culture—making ethics a business advantage, not a bottleneck.

Read the Case Study

‍