You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
My paper, The Generative No and Sourced Refusal in Human-AI Co-Creation, was accepted at NeurIPS 2026's main conference in Sydney (Creative AI Track). I wrote it alone as an MIT undergraduate, independent of any lab and with no funding, and I'm raising $5,000 so I can go present it.
The question behind it is simple. When an AI system refuses you, whose rule is it enforcing? A platform policy, a model's training, an interface designer and your own earlier instruction can all produce the same "I can't help with that," and each one gets challenged somewhere different. I ran seven boundary prompts through four open-weight model families. Thirteen of the 28 responses refused or redirected. Every one named something it was protecting, and none said who imposed the constraint, what it rested on or how to contest it.
I also built The Generative No, a small web probe where each refusal names its source and basis and logs what the user does next. The paper is public at https://openreview.net/forum?id=ssPJkk8w7E.
The immediate goal is to present the paper in Sydney. I'm the only author. The presentation happens only if I'm there. While I'm there I want the idea attacked by people who work on evaluation and oversight. Within 60 days I'll publish a short technical note recording what I heard and laying out the next experiment. That study would show people missing, correct and deliberately misleading source labels and measure whether they can tell who is responsible and pick the right way to push back.
The bigger bet is that as AI agents take on real work, every block, rewrite or redirect should leave a trail back to the rule that caused it. A reviewer could then spot a bad intervention, or a system misdescribing its own decision. My paper doesn't prove that yet. It shows the trail is missing today and offers one way to build it.
All of it goes to the trip. My working budget is $2,000 for economy airfare from Boston, $1,400 for seven hotel nights, $700 registration, $420 meals, $150 local transport, $150 Australian entry authorization, $100 travel insurance, $30 poster printing and $50 transfer fees. Those are estimates and I'll swap in real quotes. I'm also seeking NeurIPS financial assistance and other travel grants. If any of them covers part of the trip, I'll say so here and lower the goal. If NeurIPS covers registration and hotel, what's left is about $2,900.
Just me. I'm an MIT undergraduate in Mathematics with Computer Science. This is my first ML paper, and before it I worked on two papers in mathematics. I took this project from a blank page in March 2026 to a NeurIPS acceptance on my own, including the audit design, the analysis and the probe.
The most likely way this falls short is that the idea doesn't survive contact with experts. The audit is small, one output per model and prompt, coded by one person. The probe's labels are claims about where a refusal came from, not verified traces from inside a real system. People in Sydney may convince me the labels don't help users, or that they can't be made trustworthy in deployed systems. If that happens, the technical note will say so, and that's still a useful result for anyone thinking about building this.
The other risk is plain logistics. If the trip doesn't come together, I'll tell donors and ask before using the money any other way.
Nothing. The project has had $0 in funding and $0 in direct spending since I started it in March 2026. There's no lab or institution behind it.
There are no bids on this project.