How to Address Generative AI Bias

Ethan Altmann
April 10, 2024
Discover with anecdotes how to address generative AI bias

How to Address Generative AI Bias 

Imagine a Generative AI echoing society's whispered biases; this isn't just hypothetical, it's today's reality. Generative AI can be influenced by biased information from public sources. This is because the output of GenAI Large language models (LLMs) is shaped by their training data, which can sometimes originate from sources that aren't trustworthy or validated. As a result, the output might reflect the same bias in machine learning or prejudices in the original data. Such a generative AI bias can encompass issues like racial or gender discrimination and other sensitive attributes. This not only risks legal consequences for the organizations involved but also potential harm to their reputation.  

To ensure the reliability of GenAI outputs, it's essential when building an LLM to train it with unbiased data. A robust strategy mitigating bias in AI includes understanding the risks when it comes to these data sets, implementing controls to reduce this risk, and having governance in place to guide the organization. 

The urgency to find solutions has never been greater, especially as the commercial AI market surges forward. The repercussions are already visible. For instance, in August, the agency responsible for overseeing anti-discrimination laws marked its first settlement related to AI-biased discrimination in the workplace. The Equal Employment Opportunity Commission (EEOC) disclosed that tutoring firm iTutor agreed to a $365,000 settlement after allegations that its AI-driven hiring tool systematically turned down female candidates over 55 and male candidates over 60. As AI integration in recruitment and hiring escalates — with current estimates suggesting 79% to 85% of employers are using some AI form — this percentage is only set to grow. This rapid adoption will inevitably prompt employers to seek guidance on how to best comply with legal standards and prevent AI bias in hiring systems.

In the following post, we will explore generative AI bias examples and how to reduce bias in AI.

Top 5 Types of Dataset-Related Bias in AI 

For organizations venturing into the frontier of Generative AI, the safe and unbiased integration of this technology is not just a benefit, but a necessity, particularly in areas where AI Compliance regulations have been introduced (such as in the EU - see EU AI Act). Compliance with this regulation is heavily reliant on recognizing and addressing potential biases, particularly those concealed within LLM generative AI. To navigate this complex landscape, it's crucial to be informed and proactive.

Here are the top 5 types of AI biases that may arise during the training and implementation of generative AI solutions.

1. Historical Bias

Historical biases arise from historical data that may be tainted with societal prejudices from the past. Since GenAI models are trained on vast amounts of data from the internet, they can pick up and even amplify the biases present in that data. For example, if a GenAI system is trained on books and literature primarily from the 18th and 19th centuries, then it might pick up and reinforce antiquated views on gender roles, racial hierarchies, or colonial perspectives, not reflecting the more progressive and inclusive views of modern society.

2. Measurement Bias

Measurement bias occurs when the data collection process is flawed, leading to systematic errors. In medical research, if an AI system is trained primarily on data collected from patients using a specific type of medical equipment or diagnostic test, it might not provide accurate predictions or diagnoses for data gathered from alternative equipment or tests, thus unintentionally favoring one measurement method over another.

3. Confirmation Bias

GenAI can inadvertently encourage confirmation bias by providing outputs that users expect to see based on their pre-existing beliefs. If not addressed, this can lead to a reinforcing loop of existing biases. For example, if users often search for articles confirming that "coffee is good for health" and the AI primarily retrieves such articles over time, it might start to overly emphasize that viewpoint, even if significant research shows potential downsides.

4. Group Attribution Bias

This form of bias occurs when the AI ascribes specific characteristics or behaviors to an entire group based on a limited sample. This can reinforce harmful generalizations. For example, if an AI tool analyzes sales performance data and notices that sales representatives from urban areas tend to close deals more quickly than those from rural areas, it might start recommending higher bonuses or promotions primarily to representatives from urban areas, assuming their inherent superiority in sales skills.

5. Availability Bias

This data bias in AI  is rooted in the models' exposure to copious publicly available data, leading them to favor prevalent content and potentially overlook less common perspectives or data. For example, consider a company using an AI tool to forecast market trends based on online articles and news reports. If there's been a recent surge in articles about the boom in the electric vehicle (EV) market, AI might overemphasize the potential of the EV sector and underrepresent other growing sectors. As a result, a business might mistakenly allocate a disproportionate amount of resources and investment into the EV market.

Mitigating Bias in AI

Awareness of these potential biases is the first step toward ensuring more equitable AI systems. Only with understanding can we implement measures to prevent, detect, and correct them. Potential solutions for how to prevent bias in AI include:

Diverse Training Data: 

Ensuring that training data is comprehensive and diverse can mitigate many of the generative AI biases above. For example, an e-commerce company using AI to recommend products to users could face bias if the training data is predominantly based on purchases from urban demographics. By integrating diverse purchasing data from both urban and rural areas, AI can make more inclusive product recommendations.

Transparent Algorithms: 

Open-sourcing algorithms and making them transparent can help identify and rectify inherent AI algorithm biases. For example, a recruitment agency using AI to filter potential candidates for businesses might decide to share its algorithm's decision-making process. This transparency allows companies to understand the criteria and potentially identify any inadvertent AI biases in hiring systems, ensuring a more diverse candidate shortlist.

Regular Audits: 

Continual auditing of AI models can help identify biases and recalibrate the models accordingly. For example, a fintech company using AI for credit scoring can periodically review its model. Regular audit procedures can be done to validate the output from AI is as expected compared to non-AI procedures, for instance, the AI model is unjustly assigning lower scores to startups from specific industries or regions. Validating that the KPI’s for the model are aligned to the expected output can be refined to ensure fairness.

Ethical Guidelines: 

Setting strict ethical guidelines for AI development and deployment can go a long way in ensuring fairness. For example, a business analytics firm using AI to provide insights to clients can establish strict ethical guidelines. These might include fair data sourcing practices, transparent data processing methods, and open dialogue about AI's predictive limitations, ensuring clients receive unbiased and reliable business intelligence.

Working Together to Eradicate Generative AI Bias

Efforts to combat bias in generative AI models are gaining momentum as awareness grows about the ethical implications of biased outputs. Hugging Face, a pioneering AI community and library, is at the forefront of this movement, actively pushing for transparency, inclusivity, and fairness in AI training processes. Similarly, Snowflake has been developing tools and methodologies to ensure that AI models are designed with neutrality and fairness from the outset. Additionally, the GenAI Toolkit by anecdotes – pioneers in Compliance automation – offers a structured approach to identifying, managing, and rectifying biases in AI systems under Bias in Output Data Risk ANEC-AI-8 that is addressed by input validation (7.1) and output validation (7.2). This framework emphasizes continual evaluation and recalibration based on feedback and real-world results.  Collectively, these initiatives signify the industry's concerted push to create AI models that respect and understand human diversity, ensuring that AI-generated content is as unbiased as possible. 

Ethan Altmann
Compliance Product Owner - Chief framework cross-referencer, Control-understander, Evidence-mapper