AI Disinformation Detector.
An AI-powered system exploring how machine learning can detect misleading information across digital environments.
Current status
RESEARCH / PROTOTYPE
Selected project / 01
AI / MACHINE LEARNING
01 / Overview
A final-year software engineering project exploring how artificial intelligence and machine learning can be applied to detecting potentially misleading information in digital environments.
Baseline test accuracy
~36%
Traditional ML experiments
Transformer validation accuracy
33.2%
DistilBERT experiment
Primary focus
NLP
Text classification and evaluation
02 / Behind the work
The problem
The challenge is not simply to label a statement as true or false. Language can be ambiguous, context matters and datasets can contain imbalanced or inconsistent labels. A useful system needs to be evaluated critically rather than judged by one headline accuracy figure.
The approach
The pipeline processes datasets into a consistent format before preparing the text for modelling. Traditional machine-learning approaches provide a point of comparison for transformer-based experiments, making it possible to investigate whether added model complexity translates into better performance.
Engineering decisions
Keeping preprocessing, model training and evaluation as distinct stages makes the experiments easier to repeat and compare. This also creates a foundation for further work on class imbalance, explainability and robustness without confusing new findings with earlier results.
03 / Engineering
Implementation & technical decisions
The implementation combines Python, Pandas and NumPy for data handling, NLTK for natural-language preprocessing, Scikit-learn for traditional machine-learning baselines and Hugging Face Transformers for transformer-based modelling. Logistic Regression and SVM baselines were evaluated before experimenting with DistilBERT.
Technology stack
04 / Outcomes
The traditional baseline models achieved approximately 36% test accuracy, while the DistilBERT experiment achieved approximately 33.2% validation accuracy. These results exposed the difficulty of the classification task, particularly with imbalanced classes and nuanced language. They are treated as findings to investigate, not evidence of a finished detection solution.
05 / Lessons
A more sophisticated model does not automatically create a better system. Data quality, class distribution, preprocessing, evaluation strategy and understanding model behaviour can matter as much as the model architecture itself.
Currently viewing: AI Disinformation Detector