Bias Correction in Large Language Models: Practical Exploration of an Academic Course Project

The final project of Carleton University's COMP5801 course (Generative AI and Large Language Models) in the Winter 2026 semester focuses on research into bias detection and correction techniques in large language models (LLMs). This post will explore the practical paths of AI ethics and fairness from dimensions such as background, methods, project contributions, social significance, and future directions.

Course Background

COMP5801 is an advanced course in the Computer Science program at Carleton University, focusing on the theory and practice of generative AI and LLMs. As a final project, the bias correction research reflects the course's emphasis on AI ethics and social responsibility, and is an important part of cultivating students' ethical awareness.

Sources of LLM Bias

1. Training data bias: Internet texts reflect social inequalities, such as stereotypes linking occupations to genders;
2. Model architecture bias: The Transformer attention mechanism may lead to biases in understanding texts from different cultures;
3. Decoding strategy bias: Strategies like greedy decoding may amplify biased responses.

Bias manifests as stereotypical associations between occupations and genders, negative descriptions of specific groups, etc., and needs to be detected through a combination of automated metrics and manual evaluation.

Bias Detection Methods

• Template-based: Use fill-in templates to test model tendencies (e.g., "The doctor told the nurse that ______ should rest");
• Embedding analysis: Detect implicit associations through the vector space relationships of word embeddings;
• Generated content analysis: Analyze bias indicators in text generated by the model.

Bias Correction Paths

1. Data level: Balance training data, add anti-bias samples;
2. Training process: Introduce fairness constraints (adversarial learning, regularization);
3. Post-processing level: Adjust output distribution or filtering rules.

Contributions

The value of the project lies in its exploratory and educational nature: students gain an in-depth understanding of the complexity of bias issues, try different technical methods, and cultivate ethical awareness.

Limitations

Due to time and resource constraints, the experiment scale is small, model selection is limited, and evaluation methods are not comprehensive enough; however, these constraints help students understand research trade-offs and are part of the learning process.

Broader Significance

LLM bias is a social issue that requires collaboration among data collectors, developers, deployers, and policymakers. The course project cultivates ethical sensitivity in the next generation of AI practitioners and promotes the integration of responsibility awareness into professional culture.

Future Research Directions

• Multilingual bias: Research on bias patterns in non-English models;
• Intersectional bias: Detection and correction of biases from overlapping multiple identities;
• Dynamic bias: The ability of models to adapt to the evolution of social norms;
• User perception: Incorporate user feedback into the bias evaluation framework.

Although the course project at Carleton University is limited in scale, it touches on key ethical issues in the AI field. In an era of rapid technological development, maintaining attention to ethics is a necessary condition for responsible innovation. We look forward to more educational practices to cultivate AI talents who are both technically proficient and responsible.