As models like GPT-4 got good enough to write convincingly, a new practical problem showed up: telling human writing apart from machine writing. This project is an end-to-end answer to that problem — not just a classifier, but the pipeline and deployment around it.
Why detection matters
The case for building this isn't abstract:
- Misinformation control — false information generated at scale is harder to contain if it can't be identified.
- Plagiarism prevention — academic and published work depends on knowing what was actually written by the author.
- Ethical considerations — contexts like journalism need transparency about a text's origin.
- Trust and transparency — knowing a source strengthens the credibility of the content itself.
The pipeline
The system is built as a modular production pipeline, structured in clear stages from ingestion through to evaluation — each stage independently testable rather than one monolithic script. Text is vectorized with TF-IDF and classified with a custom TensorFlow neural network trained to separate human-written from LLM-generated text.
Shipping it
A classifier that only exists in a notebook doesn't answer the original question for anyone else. This one is fully deployed on AWS EC2, with a Streamlit interface in front of it for real-time inference — type in a passage, get a read on where it likely came from.
That deployment step is deliberate: the point of this project was never just the model, but a working answer to a real and current problem.