Fine-Tuning Fairness: Building AI That Works for Everyone

Artificial intelligence is rapidly shaping how we work, learn, communicate, and access opportunity. But, AI models don't magically develop their understanding of the world; they learn from us. And if they're trained on biased data, they can reinforce and amplify pre-existing societal stereotypes and inequities.

During a recent LGBT Tech webinar session, "Fine-Tuning Fairness: Curating Datasets to Eliminate Gender Bias in Technical LLMs," AI engineer and Fundlish founder Zaria Hallager (She/They) spoke with LGBT Tech's Director of Programs, Kristen Kelley (She/They), to explore where bias enters large language models (LLMs), how engineers can reduce it, and why fairness must be treated as a core measure of AI quality.

AI Learns More Than Facts

One of the central themes of the discussion was that AI models don't just absorb information; they also learn patterns. Hallager shared research showing that coding assistants such as GitHub Copilot historically default to male pronouns for engineering roles, while other models more frequently associate women with nursing, caregiving, and administrative work.

"This isn't one product's bug," Hallager explained. "It's the same systematic fingerprint everywhere a model trains on uncorrected human text."

Because LLMs are trained on enormous collections of internet content, they absorb society's historical biases alongside factual information. If engineers don't intentionally address bias during development, we risk AI systems presenting stereotypes as objective truths.

Fairness Starts Before Training

A common misconception is that bias can simply be fixed after a model is finished; however, Hallager argued the opposite, emphasizing that engineers can and must design for fairness on day one.

That begins with understanding the data itself. Before training a model, developers should audit their datasets, measuring representation and identifying imbalances. One example of an audit could be conducting pronoun ratio checks throughout a training corpus to determine whether different genders are represented equitably.

For Hallager, building fair AI requires treating fairness like any other engineering priority by:

• Defining fairness criteria before development begins.

• Instrumenting development pipelines to identify bias throughout the training process.

• Assigning ownership and accountability so bias mitigation becomes an ongoing engineering responsibility rather than a one-time cleanup effort.

Quality Over Quantity

The discussion also explored practical techniques developers can use to reduce bias without sacrificing performance. Rather than simply collecting more information, Hallager stressed the importance of higher-quality data. "More data is not better data," they explained.

Hallager outlined a couple of approaches to ensure higher-quality data is being used:

• Data distillation, where larger models help identify the highest-quality, lowest-bias examples for training smaller models. 

• Counterfactual data augmentation (CDA), which creates balanced examples by swapping demographic characteristics while keeping every other part of the scenario identical. 

These methods help teach models that job performance, expertise, or competence should not be associated with gender or other demographic traits.

Fairness Doesn't Require Starting From Scratch

Most people assume that retraining an entire AI model to correct bias means starting over, which is often prohibitively expensive. Instead, Hallager discussed Parameter-Efficient Fine-Tuning (PEFT) and Low-Rank Adaptation (LoRA), techniques that modify only a small portion of a model's parameters. These methods make fairness improvements far more accessible for organizations that may not have the resources to rebuild large models from scratch.

However, Hallager cautioned that no amount of fine-tuning can compensate for poor-quality data. "Data quality isn't optional here," she explained. "It's the whole mechanism."

Measuring What Matters

Hallager urged simultaneous benchmarks rather than trying to find a single “perfect” metric. They offered three categories every AI team should evaluate simultaneously:

• Compatibility, which ensures the model still performs its intended tasks accurately.

• Fairness, which measures whether demographic groups receive equitable treatment using benchmarks such as WinoBias, pronoun default rates, and counterfactual testing.

• Robustness, which evaluates how models respond to adversarial prompts and whether human reviewers consistently agree with model behavior.

Importantly, Hallager emphasized that success requires passing all three categories: "A model that clears one gate and fails the second gate but passes the third does not ship." She noted that her team's results demonstrated that reducing gender bias did not require sacrificing model capability.

Building Accountability Into AI

Beyond technical implementation, Hallager encouraged organizations to build fairness into their entire AI development lifecycle, and to have these evaluations as ongoing quality checks rather than one-time exercises.

Her recommended six-step process includes:

• Auditing datasets before training begins.

• Running fairness benchmarks such as standard and explainable attention (SEAT) and WinoBias as baseline measurements.

• Generating balanced synthetic examples where representation gaps exist.

• Fine-tuning models using bias-corrected datasets.

• Re-running fairness evaluations after training.

• Integrating continuous testing into deployment pipelines to prevent future regressions.

Why Inclusive AI Matters

The conversation reinforced that AI systems increasingly influence hiring, education, healthcare, software development, customer service, and countless other aspects of daily life. If these technologies inherit society's biases unchecked, they risk reinforcing existing inequities at scale and disproportionately affecting marginalized communities, like the LGBTQ+ community.

For LGBT Tech, ensuring AI is inclusive is about more than improving algorithms—it's about expanding opportunity and ensuring emerging technologies serve everyone fairly.

As Hallager concluded: "Build the measurement in and you build the fairness in. Fairness isn't a constraint on your model's capability, but a dimension of its quality.”