Project summary

I am seeking US$4,000 for a three-month, part-time pilot evaluating the reliability of AI-generated mathematical proofs and AI-assisted scientific criticism. The project will combine mathematical analysis, counterexamples, and reproducible checks to investigate when AI contributes useful insights and when its reasoning becomes misleading.

What are this project's goals? How will you achieve them?

I will construct 30 tasks drawn from mathematical analysis and discrete dynamical systems. These will cover assessment of valid and invalid arguments, counterexample construction, missing assumptions, and whether proposed criticism is mathematically justified. Selected public correction notes from my previous work may inform the collection after renewed verification. Every task will have a checked reference argument or counterexample and explicit assumptions.

I will evaluate at least two accessible AI models, recording prompts, model identifiers, dates, and available settings. Selected tasks will be repeated to assess consistency, including whether models revise incorrect conclusions when challenged with precise mathematical objections. Responses will be classified as correct, incomplete, incorrect, or appropriately uncertain. Numerical agreement will not be presented as proof. Modified examples cannot guarantee absence of related material from model training data.

Month 1: construct tasks, verify reference arguments, and establish evaluation criteria. Month 2: evaluate models and repeat selected tests. Month 3: complete checks, seek independent review of selected cases, and publish the materials and report. Work will begin after funding is available.

How will you know it worked?

The deliverables are 30 documented tasks with reference arguments, evaluation records and model responses where redistribution is permitted, reproducible checking scripts, and a technical report. Completion will be assessed against these outputs rather than a predetermined rate of model failure. Results will be limited to tested tasks and model versions, not presented as a general model ranking or a demonstration of broad AI safety.

The resources will be freely accessible to researchers and educators in Algeria, elsewhere in Africa, and internationally. Their value will depend on the quality, transparency, and usefulness of the mathematical evidence. Examples will focus on arguments rather than personal judgments about authors. Private correspondence will not be published without permission.

How will this funding be used?

Total requested: US$4,000.

Research time for task construction, proof checking, and analysis: $2,400.

Model API access and computing: $600.

Independent mathematical review of selected cases: $600.

Documentation, release preparation, and dissemination time: $400.

An independent reviewer has not yet been appointed. Material changes to this allocation will be explained transparently.

Who is on your team? What's your track record on similar projects?

I am Rafik Zeraoulia, an Algerian PhD researcher in Mathematical Analysis and Applications at Université Djilali Bounaama de Khemis Miliana, with extensive secondary-school mathematics teaching experience and university teaching experience in analysis. This is a solo project, with independent review of selected cases planned.

My research includes rational difference equations, mathematical analysis, discrete dynamical systems, and computer-assisted counterexamples. My publications include a coauthored proof of Conjecture 1 of Kulenović, Ladas and Overdeep and a sole-authored computer-assisted counterexample concerning the three-dimensional Lamé operator.

I have used AI assistance while critically examining more than 200 research papers and providing constructive scientific criticism. This has led to correction notes identifying errors, missing assumptions, or limitations in a number of published papers. Some researchers have responded positively to my notes and the issues raised. These responses do not establish that every proposed correction is valid.

My correction notes are listed in my ORCID record: https://orcid.org/0000-0002-5436-3320. Reviewers can consult them to assess the scope and substance of this work. My CV is available on request.

What are the most likely causes and outcomes if this project fails?

Reference arguments and evaluations can themselves contain errors. I will prioritize careful checking, seek independent review of selected cases, and document unresolved disagreements. Model access constraints or unexpected task difficulty may limit coverage; any shortfall will be reported alongside completed materials. A small sample may yield findings with limited generalizability. This project addresses specific reliability failures, not a demonstrated reduction in catastrophic risk.

How much money have you raised in the last 12 months, and from where?

I have received US$0 in grant or project funding during the last 12 months. I have never received grant or project funding.

I have applied to Renaissance Philanthropy and BlueDot Impact. Any overlapping support will be disclosed, and the same expenses will not be charged to multiple grants.

AI assistance disclosure

This application was prepared with AI assistance using my CV and project discussions. I remain responsible for the accuracy of its claims, the commitments undertaken, and verification of all mathematical outputs.

Contact: zeraoulia@univ-dbkm.dz