I recently got into a debate with a friend who swears is “Against AI Slop”, but uses AI for 90% of his researching stuff. When i point it out, i got a 20 minute rant about how AI helps him learn and makes him learn faster, but i see massive issues with this. For once, using AI even as a starting point is a gamble, if not a straight damage to the learning process as a whole, mainly because sloppified information spit out by an LLM is often simplified, imprecise and oversimplified. For example, he told me to be interested in Arch Linux, but said the Arch Wiki was too “difficult to understand and took knowledge for granted” and told me that AI helped him understanding it. However i think that the wiki being verbose and designed to not be understandable by everybody is a filter to make sure that you don’t go and use Arch if you can’t understand the wiki, not for gatekeeping, but because a sloppified and short version of processes explained in the wiki will result in you knowing WHAT to do, but often not HOW and WHY. Also, it keeps you locked to AI since if you get used to the simplification (and often misinformation) given by the slop generator, it’ll be hard to understand the real sources, and it’ll be hard to learn the jargon and will lead to an incomplete result. I could go on for days and feel free to debate me, but I am tired of pretending that AI is a great learning tool. It can be, but in its current status it’s just not, and when it can be, IMO the other cons that come with AI usage outweigh the possible value of the info.
Relying purely on the knowledge from an LLMs dataset is never a good idea. Frontier models at least append their data by popular demand. Google used to do this back in the day. If a search used to show no results they would note that and start searching their indexed sites for the unresolved search. Maybe a week or so later you could try again and go Google would finally come up with something. Same thing happens with LMMs now. If a question or prompt is popular enough they scrub their indexed sites with more focus and refine the dataset to give more accurate responses. But more accurate is still not 100% reliable. Open weight models usually have Datasets up until 2020, since filtering out AI slop from current day data is a science in itself and smaller labs and institutes don’t have the rescources for that.
But where LLMs shine is in the fact that they can understand natural language and use tools. For example if I where to research a topic I’d go through hours of web searches, clicking all the links, reading all the abstracts etc. Why not just have an LLM with a websearch access do some agentic work. You give it relevant search terms, and it will go through several pages of the Google results, create summaries and provide the sources directly.
Add some MCP servers to that so LLMs can directly interact and navigate websites and it will massively speed up research. Instead of reading through several pages of a paper to figure out it is only vaguely related to your topic turns from 10-20 minutes to 1-2 minutes.
For example Qwen Modells can easily handle 300 page PDF documents without hallucinations.
And once you’ve got enough material, then you can go back on the sources and properly read through it yourself.
LLMs can very much be helpful for research, if you use them properly for what they where designed to do, namely data analysis, instead of some know-it-all Guru.