Guidelines for Responsible AI Flashcards

Data augmentation helps balance the dataset by creating new samples from underrepresented groups. By applying transformations such as rotation, cropping, or flipping to the existing images, the researcher can reduce the bias in the dataset and improve the fairness of the model’s outputs.

• Model monitoring for accuracy drift is incorrect because monitoring for accuracy drift helps track model performance over time but does not correct bias in the training data. It identifies performance issues but doesn’t resolve them.
• Retrieval Augmented Generation (RAG) is incorrect because RAG is a method for improving generative models by retrieving relevant information from external sources, which doesn’t help balance the data or correct the underlying bias in this scenario.
• Edge detection for image processing is incorrect because edge detection is a technique used to identify boundaries in images, which is unrelated to addressing dataset bias. It focuses on feature extraction, not on balancing data or mitigating bias.

Reference:
Class Imbalance (CI)

Ensuring the model generates unbiased outputs is crucial for responsible AI deployment, especially in sensitive tasks like legal drafting. Using diverse data helps the model generalize better and reduces bias in its predictions.

• Adjust the model’s temperature to increase output variety: Adjusting the temperature controls output randomness but does not address bias or fairness.
• Limit training epochs to prevent overfitting: While useful for preventing overfitting, this doesn’t address bias or fairness concerns in model outputs.
• Apply regularization techniques to tune hyperparameters: Regularization can improve model performance but does not directly address fairness or bias.

Reference:
Transform Responsible AI from Theory into Practice

Amazon SageMaker Clarify helps identify and mitigate bias in machine learning datasets and models, ensuring fairness during both data preparation and model training.

• Implements a Retrieval Augmented Generation (RAG) pipeline is incorrect because RAG is a method used to improve generative models by retrieving relevant information, and SageMaker Clarify does not handle RAG workflows.
• Tracks the performance of machine learning models in production is incorrect because tracking model performance in production is handled by tools like SageMaker Model Monitor, not SageMaker Clarify.
• Captures key metadata about machine learning models is incorrect because documenting metadata about models is a feature of SageMaker Model Cards, not SageMaker Clarify.

Reference:
Amazon SageMaker Clarify

Including labeled examples in the prompt helps guide the LLM by demonstrating how to classify sentiments, improving the model’s accuracy in identifying whether a new passage is positive or negative.

• Include a thorough explanation of sentiment analysis techniques and how LLMs work in the prompt is incorrect because providing an explanation of the underlying techniques does not help the LLM in performing the specific task of sentiment classification. The model responds better to concrete examples rather than theoretical explanations.
• Input the new text passage without any examples or context and ask the model to classify it is incorrect because not providing examples or context decreases the model’s ability to accurately perform sentiment classification, especially if it lacks prior instructions on how to approach the task.
• Include the new text passage along with examples of other tasks, like text summarization or translation, in the prompt is incorrect because mixing different tasks like summarization or translation dilutes the model’s focus on sentiment analysis and can lead to incorrect results.

Reference:
Amazon Bedrock

Plagiarism occurs when content generated by AI is copied without proper attribution, leading to ethical and academic concerns. In this scenario, the student is using AI-generated text without crediting the source, which constitutes plagiarism.

• Misinformation is incorrect because misinformation refers to false or inaccurate information being generated, not copying content without proper credit.
• Hallucinations is incorrect because hallucinations refer to AI generating false or misleading information that wasn’t part of the original input.
• Bias is incorrect because bias refers to unfair or unbalanced outputs from AI systems, not the unauthorized copying of content.

Reference:
Transform Responsible AI from Theory into Practice

Partial dependence plots (PDPs) help stakeholders understand how specific features influence the model’s predictions. By showing the relationship between inputs and outputs, PDPs provide insights into the model’s decision-making process, enhancing transparency and explainability.

• The raw data that was used to test the model is incorrect because providing raw data doesn’t explain how the model makes predictions. Raw data alone does not enhance the interpretability of the model.
• A summary of the model’s memory usage during training is incorrect because memory usage is related to model efficiency, not its transparency or explainability.
• The dataset schema is incorrect because while the schema shows the structure of the dataset, it does not help explain how the model makes decisions.

Reference:
Model Explainability

Inference is the process where a trained model is used to analyze new data and make predictions. In this case, when the deployed object detection model analyzes a new image to identify objects, it is performing inference. The model is applying what it learned during the training phase to make predictions on unseen data.

• Training is incorrect because training refers to the process of teaching the model using a labeled dataset. The model has already been trained and is now being used to make predictions.
• Model deployment is incorrect because model deployment refers to the process of moving the trained model into a production environment. The question focuses on what happens after the model has been deployed.
• Bias correction is incorrect because bias correction involves adjusting the model to prevent biased outputs. The question does not mention any bias-related issues, and it focuses on object detection in a production environment.

Reference:
Deploy Models for Inference

Hallucination refers to the problem where a model generates content that appears factual but is not based on accurate or relevant data. In this case, the LLM is generating plausible-sounding but incorrect marketing content, which is a hallmark of hallucination in language models.

• Data leakage is incorrect because data leakage happens when the model has access to information during training that it wouldn’t have during real-world use, which leads to overly optimistic results. It does not explain why the generated content is incorrect.
• Feature selection is incorrect because feature selection refers to choosing the most relevant input variables in traditional machine learning models. It does not apply to the generation of incorrect text by LLMs.
• Gradient explosion is incorrect because gradient explosion refers to issues in the training process of deep learning models, where gradients grow uncontrollably, but it is unrelated to generating incorrect but plausible text.

Reference:
Improve LLM performance with Human and AI Feedback on Amazon SageMaker for Amazon Engineering

Standardized benchmark datasets are specifically designed to evaluate models for bias and fairness. These datasets are pre-built and provide a low-effort way to assess how the LLM behaves across different demographics or writing styles, without the need for significant manual data collection or preparation.

• Historical grading records is incorrect because using historical grading records requires significant manual work to analyze for bias, making it less efficient than using benchmark datasets.
• Grading rubrics is incorrect because rubrics provide grading criteria but do not serve as a data source for evaluating the model’s fairness or bias.
• Peer-reviewed journal articles is incorrect because journal articles can provide theoretical insights but are not practical data sources for evaluating model outputs for bias.

Reference:
Transform Responsible AI from Theory into Practice

Identify potential biases or discrepancies in the patient data: By addressing biases or discrepancies in the patient data, the organization ensures that the model doesn’t make biased treatment recommendations that favor or disadvantage certain patient groups.

Regularly assess the model’s outputs for fairness and share the results with healthcare regulators: Ongoing evaluation of the model’s outputs for fairness helps the organization remain compliant with healthcare regulations and ensures the model is used responsibly.

• Run the model continuously to improve prediction consistency is incorrect because continuously running the model does not directly address the issue of bias or fairness in its outputs.
• Use F1-score evaluations to ensure the model provides the highest accuracy in all cases is incorrect because the F1 score measures model performance, but it does not address fairness, transparency, or bias in healthcare recommendations.
• Ensure that the model processes data at a sufficient speed for real-time recommendations is incorrect because processing speed does not impact responsible AI usage or the reduction of bias in patient treatment recommendations.

Reference:
Transform Responsible AI from Theory into Practice

Adding safeguards in the prompts teaches the LLM to detect suspicious or harmful inputs and avoid giving unsafe responses or revealing sensitive information. This helps protect the chatbot from being manipulated by malicious prompts.

• Increase the number of tokens the model can process to handle more context is incorrect because increasing the token limit affects how much text the model can handle but doesn’t prevent manipulation.
• Use a model that updates frequently for better accuracy is incorrect because frequent updates improve accuracy but don’t directly prevent the chatbot from being tricked by malicious prompts.
• Lower the temperature setting to make the model’s responses shorter is incorrect because lowering the temperature makes responses more predictable, but it doesn’t protect against prompt manipulation.

Reference:
Transform Responsible AI from Theory into Practice

Sampling bias occurs when the data used to train the model is not representative of the entire population, leading to biased predictions. In this case, the model might have been trained on data that over-represents or under-represents certain ethnic groups, resulting in disproportionate flagging of individuals from that group.

• Measurement bias is incorrect because measurement bias occurs when data is inaccurately measured or recorded, not when specific groups are disproportionately flagged.
• Cognitive bias is incorrect because cognitive bias refers to the biases in human decision-making, not the biases that occur in training data for machine learning models.
• Algorithmic bias is incorrect because while algorithmic bias can happen, the issue here is rooted in the data sampling process rather than the algorithm itself.

Reference:
Transform Responsible AI from Theory into Practice

Amazon SageMaker Clarify helps in improving transparency and explainability by generating metrics, reports, and explanations about how the model makes its predictions. This is essential for meeting regulatory requirements in healthcare.

• Configure security and compliance settings using AWS Shield is incorrect because AWS Shield is focused on protecting against DDoS attacks and is unrelated to model transparency and explainability.
• Encrypt and secure the data using AWS Key Management Service (KMS) is incorrect because while securing data is important, encryption does not address transparency or explainability requirements for the model’s predictions.
• Use Amazon Transcribe to convert medical reports into text for further analysis is incorrect because Amazon Transcribe is for speech-to-text conversion, which is unrelated to ensuring model transparency or explainability.

Reference:
Amazon SageMaker Clarify

Designing and training a model from scratch gives the company full control and responsibility over the entire lifecycle, including data security, model training, and infrastructure. The company must manage all aspects of compliance and security, giving it the most ownership of the process.

• Using a pre-built AI model from a third-party platform to analyze clinical trial data is incorrect because the security and maintenance of the model are primarily managed by the third-party platform, leaving the company with minimal control.
• Developing a healthcare chatbot using an existing generative AI service hosted by a cloud provider is incorrect because the cloud provider handles much of the foundational security for the AI model, leaving the company with less responsibility.
• Fine-tuning a third-party generative AI model with patient records and research data is incorrect because, while the company is responsible for the fine-tuning process, the core model’s security and maintenance remain the responsibility of the third-party provider.

Reference:
Transform Responsible AI from Theory into Practice

Pre-built bias detection tools can automatically scan the LLM-generated content for biases, harmful language, or stereotypes, allowing the company to ensure compliance and maintain quality with minimal manual effort.

• User feedback reports is incorrect because relying on user feedback introduces delays and requires more manual review after the content has been generated and published.
• Internal moderation guidelines is incorrect because guidelines alone do not automatically review or detect bias. They require manual enforcement, which increases administrative overhead.
• Historical product reviews is incorrect because product reviews are user-generated content and do not help directly in evaluating the bias of newly generated product descriptions.

Reference:
Pre-training Data Bias