AI projects for students: the ones that are already obsolete, and the ones we actually ask about
Most AI projects students put on a resume were good signals in 2020 and are dead ones now. I have screened these as a CTO and I co-wrote a book of them, so I can tell you exactly which projects make a hiring manager's eyes glaze and which ones make them ask a follow-up question. The list changed. Most students are building the old one.
If you are a student building an AI project to get noticed, the first thing to know is that the target moved and most advice did not move with it. The projects that were strong signals five years ago are now the ones that tell a hiring manager you followed a tutorial, because everyone followed the same one. I have screened these projects as a CTO, and I co-authored a book of AI projects, so this is not a guess about what impresses. It is what I actually did when the resume was in front of me.
The projects that are already obsolete
The MNIST digit classifier. The Titanic survival model. The IMDB sentiment classifier. The spam detector. And now, the newest member of the dead list, the GPT-wrapper chatbot that calls an API and puts a chat box in front of it. These are not bad exercises. They are fine ways to learn. They are terrible ways to stand out, because their supply is effectively infinite and a signal that everyone can produce is not a signal at all. When I see the same five projects on the hundredth resume, the projects section has stopped telling me anything except that you can complete a well-documented tutorial, which I already assumed.
What we actually ask about in 2026
The projects that earn a follow-up question are the ones that touch the problems teams are actually paying to solve right now. Something built with retrieval over your own data, where you had to deal with the retrieval being wrong. An agent that does a multi-step task, where you hit the part where step three fails and had to handle it. Anything with an evaluation harness, because knowing how to measure whether your AI output is good is the rarest and most valuable skill in the field right now. Anything that runs on-device or under a real constraint. The common thread is not that these are trendy. It is that they force you into the actual hard parts, and the old projects protect you from them.
The test underneath all of it
None of this is really about the topic. A boring project done for real beats an impressive project done from a tutorial, every time. What I am actually looking for is the same three things I look for in any project on a resume: did you ship it so someone who is not you used it, did you handle messy real data instead of a clean download, and can you explain one decision you made and what it cost you. A RAG project built by pasting a tutorial fails all three exactly as hard as a Titanic model does. The topic is not the signal. The realness is.
So if you are picking an AI project to get hired, pick a current problem and then make it real. Not because “current” impresses, but because the current problems have not been tutorialised into meaninglessness yet, and because they drag you into the parts of the work that are actually worth demonstrating. If you have no experience and are trying to break in, that project is your substitute for experience, which is a whole argument on its own that I made in how to get an ML job with no experience. Build the current one. Then be able to tell me what broke.