For many researchers, ChatGPT has become the first place to ask a question.
Need a quick explanation of an unfamiliar method? Ask ChatGPT. Looking for a summary of a dense paper? Paste it into the chat. Trying to draft an email, rephrase an abstract, or brainstorm research directions? It does all of that in seconds, and you get an answer that sounds authoritative enough to either let you sleep or send you down a rabbit hole.
The Appeal Is Obvious
Why wouldn’t you just use ChatGPT?
The answer isn’t that ChatGPT is broken. It’s that it’s built for a different job entirely. The reason scientists reach for chat LLMs is totally legitimate — there’s friction everywhere else. Learning a new search interface takes time. Instead of opening half a dozen browser tabs, searching databases, and piecing together information manually, you get a polished answer almost instantly. For busy researchers juggling experiments, teaching, grants, and publications, that speed is hard to ignore.
But speed isn’t the same as scientific reliability.
The problem starts when we treat a summarization tool as though it were also a discovery tool, without realizing they are fundamentally different things. LLMs are made to generate convincing language. Scientific discovery depends on evidence that can be traced, challenged, reproduced, and updated. Every claim needs a source. Every conclusion exists within years — sometimes decades — of prior work.
What Scientific Discovery Actually Requires
Here’s what makes science different from, say, drafting an email or getting recipe ideas. Every claim you stake has consequences. If you cite something wrong in a paper, reviewers catch it. If you build a hypothesis on a finding that doesn’t actually exist or has been superseded, you’ve wasted months. If you make a clinical decision based on misread evidence, it affects people.
That means research work depends on four foundations that general-purpose chat models weren’t designed to provide.
Provenance Comes Before Answers
You need to know not just what the answer is, but where it comes from.
In research, when someone claims that a protein regulates inflammation or a drug improves patient outcomes, the immediate follow-up isn’t “Interesting.”
It’s “Where did that come from?”
Researchers need to inspect the original experiment, evaluate the methods, understand the sample size, examine limitations, and decide whether the conclusion still holds in light of newer evidence. Without provenance, even a correct answer becomes difficult to trust, because there’s no straightforward way to verify it.
Research Lives in Relationships, Not Individual Papers
Scientific knowledge isn’t a collection of isolated publications. Each paper exists within a much larger network of supporting studies, conflicting findings, follow-up experiments, systematic reviews, and evolving consensus. A newly published result might confirm years of previous work or completely overturn it. Understanding those relationships, and navigating connections across the literature, is often more valuable than reading any single paper.
Scientific Knowledge Never Stops Moving
Scientific literature updates constantly. Every day, new preprints appear, journals publish fresh findings, and ongoing debates shift as additional evidence accumulates. What represented the best understanding six months ago may already have been refined or challenged today. That’s why researchers continually revisit the literature rather than relying on static knowledge — and why any tool frozen at a training cutoff is working with an incomplete map.
Confidence Matters as Much as Conclusions
You need to know not just what someone found, but whether you should believe it. That means understanding the study design, the sample size, and the limitations of a study. It is what lets a researcher ask: why should I trust this? ChatGPT optimizes for none of these things. It optimizes for being fluent.
Where General Chat LLMs Run Short
None of this means ChatGPT is a bad tool. The problem is that researchers often expect it to solve a problem it wasn’t designed to solve. No amount of fluent language compensates for citation hallucination, and no amount of speed compensates for being months behind the literature.
Citation Hallucination
This is the failure everyone knows about, but the mechanism is worth understanding. When a language model generates text, it isn’t retrieving citations from a database. It’s generating tokens based on statistical patterns in its training data. When it produces “Smith et al., 2021 found X,” it isn’t looking up the paper and checking whether that is what it actually said. It’s completing a pattern, which is sometimes accurate. Sometimes it invents a paper that doesn’t exist, or attributes a finding to the wrong paper, or quotes a study backwards.
Studies on frontier models have found that when asked to cite sources in technical queries, they produce fabricated or non-existent references at a meaningful rate. Not catastrophic, but frequent enough that you can’t simply trust what you’re told. For researchers, that’s a critical distinction.
Bigger Context Windows Don’t Solve the Problem
Recent AI models advertise context windows capable of processing hundreds of thousands or even millions of tokens. On paper, this sounds transformative. In practice, scientific literature is vastly larger.
A single comprehensive review in computational biology might be 15,000 words. A single year of high-volume conference papers in machine learning is probably 50+ million tokens. A 100K-token window covers maybe a few dozen papers if you manually feed them in. And if you’re just asking a question, the model is seeing zero papers. It’s working entirely from parametric memory — the patterns it learned during training.
The comforting lie is that a big context window means “access to the literature.” It doesn’t. It means you could upload some literature if you went to the trouble. For most queries, you don’t.
No Native Relationship Model
A chat model can tell you that Paper A cites Paper B. It can even tell you the gist of the relationship. But it can’t represent the graph underneath it the way a knowledge graph does. It can’t natively query “all papers that cite X but contradict Y,” or “all genes implicated in this disease and their interaction patterns,” or “clinical trials on this compound organized by outcome.” Those queries require an actual graph structure. A language model simulates understanding them — sometimes convincingly, other times disastrously.
Training Cutoff Lag
Every chat model has a training cutoff, and the literature does not stop at it. Within months of any cutoff date, new work has already shifted what the state of the art looks like. If you’re working on something moving fast, as most science is, you’re working partly from an outdated map.
The deeper point worth noting is that chat models have no structural mechanism for knowing when they’re wrong. A language model doesn’t have a little light that goes off when it generates an inaccurate citation. A knowledge graph does. It knows what it knows, it knows what it doesn’t know, and it knows what’s in conflict.
What Structured, Graph-Grounded Infrastructure Changes
Instead of asking an AI to “remember” everything it learned during training, graph-grounded systems organize scientific knowledge into explicit, traceable relationships. Every paper, author, dataset, gene, disease, drug, or experimental finding becomes part of a connected network that can be queried, verified, and continuously updated. That changes the role AI plays in research.
We noticed this in previous articles. After Papers with Code was shut down in July 2025, what we really saw was the cost of losing an unstructured, crowd-maintained resource. Information did not just disappear — a whole layer of relational metadata vanished with it. This made the community realize how structure matters for durability.
Similarly, when we discussed PubMed Knowledge Graph 2.0, the real value wasn’t in adding more data. It was in encoding relationships explicitly: gene-to-disease, disease-to-drug, citation lineage, study methodology. You can query those relationships. A researcher looking at a disease can ask “what drugs are in clinical trials?” and get back not a list of LLM-generated guesses, but an actual graph of connections with evidence attached to each edge.
That’s the infrastructure layer that chat models don’t have and can’t simulate. And that’s where Axy is laying groundwork that matters.
What Axy does differently is ground generation in an explicit, queryable scientific knowledge graph instead of parametric recall. When you ask a question, you’re not getting prose that sounds like it came from the literature. You’re getting an answer that traces back through an actual graph of sources, with relationships intact. That means provenance is built in — you’re not playing the “hope this citation exists” game. Freshness becomes possible, because the graph updates with the literature. And audit trails become native: you can follow the chain from claim back to source in the fundamental structure of the answer.
The Right Tool for the Right Layer of the Workflow
Chat LLMs and knowledge graphs solve different problems in the research workflow.
ChatGPT is genuinely great for scaffolding. If you already know your domain, or have a rough sense of the answer you’re looking for and just want help summarizing it, a language model is fast and useful.
Structured scientific infrastructure is the backbone beneath that. It ensures the prose is grounded in verifiable fact, that relationships are explicit, and that the map doesn’t go out of date the moment a new paper lands.
The real power comes from both. Use structured tools for discovery, verification, and synthesis. Use language models for explanation, drafting, and sense-checking. Understand what each one is actually good at.
The scientists who are going to build the most reliable research workflows aren’t the ones asking “should I use ChatGPT or a knowledge graph?” They’re the ones building systems where structured infrastructure and language models have clear, separate roles. They let LLMs communicate, not discover. Discovery is grounded in a graph that clearly shows what it knows and can prove it. Because in science, the best answer isn’t the one that sounds the most convincing — it’s the one you can prove.
That right there is the architecture that actually works for real scientific discovery, and it’s finally becoming possible to build at scale.
Call to Action
Every scientific breakthrough starts with understanding what is already known. At Axy, we’re building the infrastructure that helps researchers spend less time searching and more time discovering. See how graph-grounded AI is reshaping scientific research.
