PoliticAI

Your safety thinking is correct. Answer only from trusted sources, show the source, and refuse when there is no good evidence. After you shared the code I can see there is a working app now, a chat screen connected to Gemini, so it is not only a plan and I raised the execution score for that. But the app is not the system the report describes. There is no real retrieval, no source database, and no refusal step. It is a normal Gemini chatbot with a small hardcoded list of political records, plus a lot of Google permissions (full Gmail, Drive, Sheets) that have nothing to do with the safety idea. I also see in the repository that the same project was sent to other hackathons under the name ElectAI, so it looks reused. Please build the retrieval and the refuse parts that you describe, remove the leftover cite tags in the report, and remove the permissions you do not need.

The project addresses an important challenge in political AI systems by prioritizing retrieval, source attribution, and refusal over unconstrained generation. The emphasis on evidence-grounded responses is well aligned with responsible deployment in politically sensitive contexts. To strengthen the project, I would recommend moving beyond the system design and demonstrating how the proposed safeguards perform in practice. For example, evaluating groundedness, citation accuracy, refusal behavior under insufficient evidence, and resilience against adversarial or misleading queries would provide much stronger evidence that the approach improves safety. It would also be helpful to compare the proposed pipeline against a standard RAG baseline to better quantify the benefits of the additional safety mechanisms. Overall, this is a promising direction that would benefit from stronger empirical validation.

Clear write-up, and the safety instinct is right: ground answers in sources, cite them, refuse when evidence is weak. The problem is real. But this is a proposal, not a project. It's written in the conditional from start to finish, with no system built, no data, no retrieval output, and no numbers. The evaluation section names good metrics, none of which were run, and the architecture is standard RAG with nothing new added. To lift this, build even a thin working slice and report real results on a small labeled query set, and add one concrete novel mechanism (contradiction detection, or a calibrated evidence-sufficiency threshold) instead of restating the pattern. The manuscript would also benefit from removing placeholder citation tags and presenting concrete results.