Stronniczość, sprawiedliwość i przejrzystość

Welcome! In this lesson, we'll tackle a critical topic: bias, fairness, and transparency in AI agents. As AI becomes more powerful, ensuring it acts fairly and predictably is essential.

AI bias occurs when an AI system produces results that are systematically prejudiced or unfair towards certain groups or individuals. This can lead to discriminatory outcomes.

Where does bias come from? Often, it's not intentional but a reflection of the data and processes used to build AI systems, especially Large Language Models (LLMs).

• Training Data: If the data used to train an LLM contains societal biases (e.g., historical stereotypes), the model will learn and perpetuate them.
• Human Labeling: Biases can be introduced during data annotation or reinforcement learning from human feedback.
• Model Design: Sometimes, the architecture or algorithms themselves can inadvertently amplify biases.

Bias manifests in many forms. Recognizing them is the first step to mitigation:

• Historical Bias: Reflects past societal prejudices present in historical data.
• Representational Bias: Under-representation or over-representation of certain groups in data.
• Stereotypical Bias: Reinforcing harmful stereotypes (e.g., gender roles in professions).
• Allocation Bias: AI systems unfairly allocate resources or opportunities (e.g., loan approvals).
• Aggregational Bias: Performance varies significantly across different subgroups.

Identifying bias requires systematic testing and evaluation. Here are some approaches:

• Probing: Presenting the LLM with specific inputs designed to reveal biased responses (e.g., asking about different genders in leadership roles).
• Fairness Metrics: Using statistical measures to compare model performance across different demographic groups.
• Human Evaluation: Having diverse groups of people review outputs for fairness and appropriateness.

Addressing bias often starts at the data level:

• Data Augmentation: Creating synthetic data or modifying existing data to balance representation.
• Data Filtering: Removing or down-weighting biased examples.
• Diverse Data Collection: Actively seeking out and including data from underrepresented groups.
• Re-balancing: Adjusting the proportion of different groups in the training set.

For LangChain agents, prompt engineering is a powerful tool to guide LLMs towards fairer outputs:

• Explicit Instructions: Directly instruct the LLM to be neutral, fair, and avoid stereotypes.
• Contextual Nuance: Provide context that disambiguates potentially biased terms.
• Role-Playing: Ask the LLM to adopt a persona that is inherently fair or unbiased.
• Reframing: Rephrase questions to avoid triggering known biases.

Beyond data and prompts, techniques can be applied directly to the model:

• Fine-tuning with Debiased Data: Training pre-trained LLMs on carefully curated, bias-reduced datasets.
• Adversarial Debiasing: Using adversarial networks to train models to be less sensitive to sensitive attributes.
• In-Context Learning: Providing examples of fair responses within the prompt to guide the model.

Fairness isn't a single concept; it has many definitions. Choosing the right one depends on the application:

• Equal Accuracy: Ensuring the model performs equally well for all groups.
• Equal Opportunity: Ensuring that false negative rates (e.g., wrongfully denying a loan) are similar across groups.
• Demographic Parity: Ensuring the positive outcome rate (e.g., loan approval rate) is similar across groups.

These definitions can sometimes be in conflict, requiring careful trade-offs.

Transparency, often linked with Explainable AI (XAI), means understanding how an AI agent arrived at a particular decision or output. This is crucial for building trust and accountability.

• Debugging: Helps developers understand and fix issues.
• Auditing: Allows regulators and users to verify fairness and compliance.
• Trust: Users are more likely to trust a system they can understand.

While LLMs are often 'black boxes,' LangChain's architecture helps in achieving transparency by breaking down complex tasks into observable steps. Tools like LangSmith (covered in a previous lesson) visualize these chains.

Even without complex tools, we can design agents to output their 'thought process'. Try this conceptual example:

An AI agent is trained on historical job application data where male applicants historically received more interview offers for engineering roles, even with similar qualifications.

Which type of bias is most likely to be perpetuated by this agent if not mitigated?

We've explored the critical concepts of bias, fairness, and transparency in AI agents. You learned about the sources and types of bias, strategies for detection and mitigation (data, prompting, model-level), and the importance of XAI.

Ensuring ethical AI is an ongoing process that requires vigilance in data curation, model development, and continuous monitoring. As you build agents, always consider the potential for bias and strive for fair and transparent outcomes.