Detecting Ghostwritten College Essays

ISEF Category: Behavioral and Social Sciences

Subcategory: Other  ·  Difficulty: Advanced  ·  Setup: University Lab  ·  Time: Full Year

The Hook

A few paragraphs can hide a lot more than a story. Writing leaves patterns, and those patterns can act like a fingerprint. That means a computer may spot when an essay does not sound like the student who submitted it. Your project asks how far those signals can go before the guess becomes shaky.

What Is It?

Forensic linguistics studies how language can reveal who wrote a text. Stylometry is the part that measures writing style, such as sentence length, word choice, punctuation, and how often a person repeats certain patterns. In this project, you compare essays that are truly original with essays that are meant to look ghostwritten, then ask whether a machine can tell them apart.

Think of writing style like a voiceprint. Two people can tell the same story, but one may use shorter sentences, more transitions, or a different mix of common words. A small transformer is a machine learning model that reads text and learns patterns from examples. You would pair those learned patterns with simple stylometric features, then test whether both methods separate authentic essays from ghostwritten ones.

Because real college essays are private, many students build a synthetic-but-realistic dataset with mentor guidance. That means the texts are created to match real essay themes and skill levels without using private admissions files. The research question becomes not only whether the model works, but also which writing signals stay visible when the text is edited, polished, or imitated.

Why This Is a Good Topic

This is a strong science fair topic because you can test a clear yes-or-no claim with measurable text features. It connects to a real problem in admissions integrity, authorship detection, and digital forensics. You can learn data cleaning, feature engineering, classification, and error analysis without needing a wet lab. The project also has a strong ethics angle, since you can study detection limits without exposing anyone's private essays.

Research Questions

• How does the accuracy of a stylometry model change when it classifies authentic and ghostwritten college essays?
• What is the effect of adding a small transformer to stylometric features on classification accuracy?
• Does essay length change the model's ability to detect ghostwriting?
• To what extent do punctuation and sentence-length features separate authentic essays from ghostwritten ones?
• Which writing features matter most when the model predicts authorship style?
• How does performance change when the ghostwritten essay is edited to sound more personal?
• What is the effect of training on one essay prompt and testing on a different prompt?

Basic Materials

• Computer with Python support and enough memory for text modeling.
• Spreadsheet or CSV file of essay texts and labels.
• Text editor for reviewing samples and annotations.
• Notepad or lab notebook for recording feature choices and model settings.
• Mentor-created synthetic essay dataset with authentic and ghostwritten labels.
• Headphones or quiet workspace for careful reading and annotation.

Advanced Materials

• University or lab workstation with a dedicated GPU.
• Python environment with PyTorch or TensorFlow.
• Hugging Face Transformers library for a small pretrained model.
• Scikit-learn for baselines and evaluation.
• Jupyter Notebook for model comparison and plots.
• Version-controlled dataset splits and experiment logs.
• Annotation tool such as Doccano for labeling style patterns.

Software & Tools

• Python: Runs text cleaning, feature extraction, and model training.
• Jupyter Notebook: Helps you compare baselines, metrics, and error cases in one place.
• Scikit-learn: Builds simple classifiers and evaluation reports for stylometry baselines.
• Hugging Face Transformers: Fine-tunes a small transformer on essay text.
• pandas: Organizes essay metadata, labels, and derived style features.

Experiment Steps

1. Define the exact authorship question, including what counts as authentic, ghostwritten, and edited text.
2. Choose a dataset design that balances prompts, lengths, and style levels so the model does not learn shortcuts.
3. Pick a baseline set of stylometric features before adding any transformer model.
4. Plan train, validation, and test splits that keep the same writer or prompt from leaking across sets.
5. Decide which evaluation metrics will matter most, such as accuracy, F1, and confusion patterns.
6. Prepare an error-analysis plan so you can inspect the texts the model gets wrong and explain why.

Common Pitfalls

• Letting prompt topic differences act like an easy shortcut, which makes the model look smarter than it is.
• Mixing the same writer across train and test sets, which leaks authorship clues and inflates accuracy.
• Using only polished synthetic samples, which makes the detector fail on messy real student writing.
• Ignoring essay length, which can cause the model to confuse brevity with ghostwriting.
• Treating a high accuracy score as proof of authorship, even when the model breaks on edited or paraphrased essays.

What Makes This Competitive

A competitive version goes beyond a simple classifier. You would compare stylometric features against a transformer, then test both under harder conditions like prompt shifts, light editing, and shorter essays. Strong entries also explain failure cases clearly, because a detector that flags style drift but not topic drift says more than a raw accuracy score. If you add careful error analysis and fair train-test splits, your project starts to look like real forensic NLP rather than a class demo.

Project Variations

• Use personal statement prompts instead of college essays to see whether the model still separates authorship styles.
• Compare hand-crafted stylometric features against a transformer on short scholarship answers, which are even harder to classify.
• Test whether the detector still works after essays are paraphrased or lightly edited by another writer.

Learn More

• NIH PubMed: Search for review articles on authorship attribution, stylometry, and forensic linguistics.
• MIT OpenCourseWare: Look for free NLP, machine learning, and text analysis course materials.
• Scikit-learn documentation: Read the classification and metrics guides for baseline text models.
• Hugging Face course and docs: Find free lessons on transformers and text classification.
• ACL Anthology: Search peer-reviewed papers on stylometry, authorship attribution, and student writing analysis.