Read the AI rules first
MLH says AI and LLM usage must be credited. MLH also says the project should not be a reskin of an existing AI tool. Put that at the top of the checklist before model choice, stack choice, or idea choice.
A safe AI build can explain what the model does, where human work remains, and what was reused. If your answer is just a wrapper around a hosted chatbot, expect judges to ask what you built.
Draft the submission early
Devpost recommends starting a draft submission early because it can be edited before the deadline. For online hackathons, Devpost also recommends submitting at least a week early so organizers can confirm eligibility before the deadline.
Do not wait for the product to feel finished. The draft forces useful questions: what is the project URL, where is the repo, what is the demo video, what problem does the AI solve, and which tools need credit?
Build toward the required artifact
HackAI 2025 required a Devpost explanation, a Google Form track selection, a GitHub repo, and a demo video of 2 minutes max. AI ATL 2025 required a Devpost submission, a GitHub repo created between the start and end of hacking, and a roughly 5-minute demo video.
That means the checklist has to include artifact work, not just feature work. Keep a running README, keep the repo public when allowed, and record proof before your model endpoint, GPU quota, or Wi-Fi gets flaky.
Prove the AI is functional
Google Cloud Rapid Agent Hackathon required a functional AI agent powered by Gemini and Google Cloud Agent Builder, with one partner MCP server integrated. That is a stricter bar than a text box that calls a model once.
Your demo should show the AI doing the hard part: extracting, planning, routing, ranking, checking, or controlling something. If the same product would work with a static rules engine, say that honestly or pick a sharper problem.
Make model risk visible
MLH's credit rule is the floor. For stronger AI projects, show inputs, outputs, failure cases, and guardrails. Judges need to see where the model can be trusted and where the user still has to approve.
This is why AI demos with logs, citations, confidence labels, or human approval steps tend to read cleaner. They let the judge audit the system instead of guessing whether the model hallucinated its way through the demo.
Submit before the panic window
Devpost's week-early advice is aimed at online events, but the lesson scales down to weekend events. Submit early enough that an organizer can flag eligibility issues and you can still fix the page.
For AI projects, the final QA pass should include API keys removed, credits listed, repo visible, demo video playable, and prompts or datasets named where the rules require it. Most disqualifying mistakes are boring. That is why they are checklist items.
FAQ
Do I need to disclose AI tool usage?
Yes when the rules ask for it, and it is a safe default. Source: MLH's rules guide says submissions should credit LLM and AI tool use, and it warns against reskins of existing AI tools.
What makes an AI hackathon project feel original?
The model should power a specific workflow, not replace your whole build. Source: Google Cloud Rapid Agent Hackathon required a functional Gemini and Agent Builder agent with a partner MCP server, which makes the integration and workflow part of the judged artifact.
What should be in the final AI submission check?
Check the repo, hosted URL, demo video, AI credits, model or dataset names, prompt-sensitive behavior, and any required sponsor tool. Source: HackAI 2025 and AI ATL 2025 both specify concrete submission artifacts rather than only an idea description.
How do I make model risk visible to judges?
Show the input, output, failure case, citation, log, or human approval step inside the demo. Source: MLH sets credit as the floor, while Devpost judge guidance rewards submissions that make the project easy to inspect and score.
< source-linked guide · dated when published >
browse hackathons