Inspiration
Taking inspiration from the many up-and-coming attempts to combat the rise in students using AI to circumvent their assignments, we attempted to use our expertise as both engineers and students to tackle the issue in a way that offered a seamless experience to teachers and professors alike. GPT-Zero, and softwares alike try to use AI as the means of fighting against it, and because of the infancy of AI, these softwares lack the certainty teachers would like when confronting AI usage. We offer a foolproof way of finding AI use, by leveraging prompt engineering to provide foolproof evidence.
What it does
What the program does is take a teacher's Word document as an input. This is ideally the assignment that they plan to distribute to their class electronically. We take their file and a prompt injection of their choice that they would like to embed as a fingerprint. They are then greeted with a visually identical Word Document that only shows its true colors if copied and pasted into ChatGPT. All the teacher has to do is to look for responses to the hidden fingerprints in student essays and they can confidently say that the student turned in an AI written essay.
In a poll of 25+ of our college peers, a majority of respondents said they do not read their prompts after copying into chatGPT.
How we built it
This project is a composition of three unique, but equally important parts. The frontend is the most publicly accessible aspect of the project, and it offers a unique and easy experience to view past files that were uploaded, experience what ChatGPT would respond with given the prompt injection, and allow them to upload a file in order to add the unique fingerprinting throughout it. We chose to use the widely popular React framework for this part, which allows us to build robust components that offer the most accessibility to all people, regardless of their tech literacy. The second part is the backend, where we leverage Google Cloud Platforms to handle our storage of these files for the long-term. The third and final part is a Python middleware, which acts as the bridge between our GCP backend and our React Frontend. Using RESTful API’s we use endpoints to send data from the frontend to Python, where the middleware acts upon the word file and embeds the prompt within the file, inaccessible to any viewers. From there, Python sends the data to GCP for storage.
Challenges we ran into
We couldn’t get FastAPI’s cors settings to produce the expected result of allowing both production and development servers access, in order to solve this problem we wrote our own cors header management function.
Working with rich text documents like word and pdf quickly proved challenging, we tested many python libraries including ironpdf, aspose, and PyPDF2, before settling on python-docx
Accomplishments that we're proud of
Throughout the hackathon, and the challenges we faced along the way, we are proud of the final product that we have delivered, and all that we learned along the way. Learning and appreciating the sheer amount of work that is required to do document editing gives a new appreciation to the tech leaders in the space, like Google or Adobe. The specialization that all of us went through as we focused on our specific tasks gave us a new understanding of how separate parts of a project can all come together into one cohesive piece.
What we learned
During the development of DetectorInspector we gained a more complete understanding of the entire process of creating the front and backend of a website. We learned how to use tailwind to quickly program CSS embedded into, google cloud based storage, firestore, to work as our database, and we learned how to buy and upload a domain name.
What's next for DetectorInjector
The future is twofold. The first is scaling up. We plan to offer the tool and service back to our high-school teachers, who are very keen on using this in their classrooms. The second is testing and future development. We want to bring this feature to more than just Word documents, but PDF and any text editing format. Using these two opportunities at our disposal, keeps the future of DetectorInjector alive and flourishing as it continues to grow into a service that helps students continue to learn without having the urge to rely on 3rd party applications as a crutch.
Built With
- cloudflair
- godaddy
- google-cloud
- python-fastapi
- react
- vercel
Log in or sign up for Devpost to join the conversation.