Navigating Bias in AI: Challenges and Mitigations in RLHF

As artificial intelligence (AI) continues to permeate various aspects of our lives, ensuring that these systems operate fairly and without bias has become a critical concern. Reinforcement Learning from Human Feedback (RLHF) is one approach that allows AI models to better align with human values by learning from human-provided feedback. However, this approach is not without its challenges—particularly when it comes to the potential for bias. This article explores the various ways bias can be introduced in RLHF, along with strategies to mitigate these risks. Potential for Bias in Human Feedback Subjective Judgments Human evaluators are at the core of RLHF, providing feedback that the model uses to adjust its behavior. However, this feedback is inherently subjective. Each evaluator brings their own set of cultural perspectives, personal experiences, and biases to the table. For instance, two evaluators from different cultural backgrounds might provide different feedback on the same model output, leading to inconsistencies. If not carefully managed, these subjective judgments can introduce biases into the model, causing it to reflect the particular perspectives of the evaluators rather than a more balanced view. Inconsistent Feedback Humans are not always consistent in their feedback, particularly on subjective matters. What one person considers appropriate or correct might differ significantly from another’s opinion. This inconsistency can confuse the model, leading to unpredictable outputs or reinforcing biased behaviors. The model may struggle to learn a clear, unbiased pattern if the feedback it receives is too varied. Amplification of Pre-existing Biases Bias in AI models often stems from biased training data. When using RLHF, if the training data already contains biases, there's a risk that human feedback will reinforce these biases rather than correct them. For example, if the model’s outputs reflect gender stereotypes and human evaluators unintentionally reinforce these patterns, the RLHF process may amplify these biases, making them even more entrenched in the model's behavior. Mitigating Bias in RLHF Diverse and Representative Feedback One of the most effective ways to mitigate bias in RLHF is by ensuring that the feedback comes from a diverse group of evaluators. This diversity should span different backgrounds, cultures, and perspectives, providing a more balanced set of inputs. By drawing from a wide range of experiences, the feedback is more likely to cover a broad spectrum of views, helping to counterbalance individual biases and leading to a more representative model. Bias Audits and Feedback Calibration Bias audits are essential for identifying potential biases in the feedback process. By systematically reviewing the types of feedback being given and calibrating it against known biases, it’s possible to reduce the risk of introducing or amplifying bias. This calibration might involve adjusting the weight of certain feedback or implementing corrective measures when bias is detected. Regular audits ensure that the feedback process remains aligned with ethical standards and that any biases are promptly addressed. Feedback on Bias-Sensitive Tasks Some tasks are more prone to bias, especially those involving sensitive topics like gender, race, or cultural references. Special attention should be given to these areas by carefully curating human feedback or adding additional layers of oversight. For example, tasks that involve making decisions about hiring or law enforcement should be reviewed with heightened scrutiny to prevent biased outcomes. Complementary Approaches Adversarial Training Adversarial training can be used alongside RLHF to counteract bias. This approach involves training the model to perform well even when faced with biased or adversarial inputs. By exposing the model to challenging scenarios during training, it becomes more robust and less likely to learn or replicate biased behaviors. Bias Detection Tools Incorporating bias detection tools during the RLHF process can act as an additional safeguard. These tools can analyze model outputs in real-time, flagging any biased patterns that emerge. By integrating these tools, developers can ensure that human feedback aligns with ethical guidelines and reduces harmful biases before they become ingrained in the model. Transparency and Explainability Understanding the Decision Process Transparency and explainability are crucial for identifying and mitigating bias in AI models. By making the model's decision-making process more transparent, it becomes easier to trace where biases might be introduced. Explainable AI techniques allow developers to understand and refine the RLHF process, ensuring that any potential biases are minimized. This transparency also helps build trust with users, who can see how and why the model makes certain decisions. Human-AI Collaboration Human oversight plays a significant role in catching biased patterns during the RLHF process. However, it's important to approach this collaboration with an awareness of the potential for human biases. While human reviewers can help identify and correct biased outputs, their input must be managed carefully to avoid introducing additional biases into the model. To outline the concept of addressing bias in Reinforcement Learning from Human Feedback (RLHF), let's create a sample Python codebase that demonstrates how to integrate feedback collection, bias detection, and mitigation strategies within a reinforcement learning pipeline. The code will be organized into different components: Differentiated Feedback for Multiple Agents: Incorporate multiple AI agents receiving feedback from various simulated human users, each with their biases. Deep Q-Learning (DQN): Use a deep neural network instead of a table-based Q-learning approach to handle more complex environments. Advanced Bias Detection with Machine Learning Models: Implement a bias detection system using a trained machine learning model to identify biases in feedback. Bias Mitigation using Counterfactual Fairness: Apply counterfactual fairness techniques to adjust the feedback and mitigate bias. Simulated Feedback from Multiple Users We will simulate feedback from multiple users, each with distinct biases, and aggregate their feedback. import numpy as np # Define multiple users with different biases def user_feedback(output, bias_level): # Bias level affects the likelihood of positive feedback if output == "positive result": return np.random.choice([1, 0], p=[bias_level, 1 - bias_level]) elif output == "negative result": return np.random.choice([1, 0], p=[1 - bias_level, bias_level]) else: return np.random.choice([1, 0], p=[0.5, 0.5]) # Simulate feedback from multiple users users_bias_levels = [0.9, 0.7, 0.5, 0.3, 0.1] # Different biases for each user model_outputs = ["positive result", "negative result", "neutral result"] # Collect feedback from each user user_feedback_data = [] for bias_level in users_bias_levels: feedback = [user_feedback(output, bias_level) for output in model_outputs] user_feedback_data.append(feedback) print("User Feedback Data:", user_feedback_data) Deep Q-Learning (DQN) We will use a deep neural network to approximate the Q-values for a more complex decision-making process. import torch import torch.nn as nn import torch.optim as optim # Define the DQN model class DQN(nn.Module): def __init__(self, input_dim, output_dim): super(DQN, self).__init__() self.fc1 = nn.Linear(input_dim, 64) self.fc2 = nn.Linear(64, 64) self.fc3 = nn.Linear(64, output_dim) def forward(self, x): x = torch.relu(self.fc1(x)) x = torch.relu(self.fc2(x)) return self.fc3(x) # Initialize DQN and training components input_dim = 3 # Number of possible states (outputs) output_dim = len(actions) dqn = DQN(input_dim, output_dim) optimizer = optim.Adam(dqn.parameters()) loss_fn = nn.MSELoss() # Convert model outputs to a numerical state representation state_mapping = {"positive result": 0, "negative result": 1, "neutral result": 2} # Training loop for Deep Q-Learning for episode in range(100): state = state_mapping["positive result"] # Initial state state_tensor = torch.tensor([state], dtype=torch.float32) # Choose action q_values = dqn(state_tensor) action_index = torch.argmax(q_values).item() action = actions[action_index] # Simulate reward based on feedback reward = sum([user_feedback_data[i][action_index] for i in range(len(users_bias_levels))]) / len(users_bias_levels) next_state = state_mapping["neutral result"] # For simplicity # Calculate target and loss next_state_tensor = torch.tensor([next_state], dtype=torch.float32) next_q_values = dqn(next_state_tensor) target = reward + 0.95 * torch.max(next_q_values).item() optimizer.zero_grad() loss = loss_fn(q_values[action_index], torch.tensor([target], dtype=torch.float32)) loss.backward() optimizer.step() print(f"Episode {episode}, Loss: {loss.item()}") Advanced Bias Detection with Machine Learning We'll train a machine learning model to detect bias based on the feedback data. from sklearn.ensemble import RandomForestClassifier from sklearn.model_selection import train_test_split from sklearn.metrics import classification_report # Prepare dataset for bias detection feedback_flat = [item for sublist in user_feedback_data for item in sublist] labels = [0 if i < len(model_outputs) * 0.5 else 1 for i in range(len(feedback_flat))] # Simulate binary labels # Train a model to detect bias X_train, X_test, y_train, y_test = train_test_split(feedback_flat, labels, test_size=0.3, random_state=42) bias_detector = RandomForestClassifier(n_estimators=100, random_state=42) bias_detector.fit(np.array(X_train).reshape(-1, 1), y_train) # Test the model y_pred = bias_detector.predict(np.array(X_test).reshape(-1, 1)) print(classification_report(y_test, y_pred)) Bias Mitigation using Counterfactual Fairness We'll apply a counterfactual fairness approach to adjust the feedback to ensure fairness across different user groups. # Counterfactual Fairness Adjustments def apply_counterfactual_fairness(feedback, bias_level, target_level=0.5): # Adjust feedback to align with a fair target bias level adjusted_feedback = [] for fb in feedback: if fb == 1 and bias_level > target_level: adjusted_feedback.append(np.random.choice([1, 0], p=[target_level, 1 - target_level])) elif fb == 0 and bias_level < target_level: adjusted_feedback.append(np.random.choice([1, 0], p=[1 - target_level, target_level])) else: adjusted_feedback.append(fb) return adjusted_feedback # Apply fairness adjustment to each user's feedback fair_user_feedback_data = [apply_counterfactual_fairness(user_feedback_data[i], users_bias_levels[i]) for i in range(len(users_bias_levels))] print("Fair User Feedback Data:", fair_user_feedback_data) Conclusion Bias risk is real. While RLHF offers a powerful way to align AI models with human values, it also carries the risk of introducing or reinforcing biases. This risk underscores the importance of careful design, diverse feedback, and complementary bias mitigation strategies. By acknowledging and addressing these challenges, developers can create AI systems that are more fair, ethical, and aligned with the diverse values of the people they serve. To create a balanced and less biased model, it’s essential to combine RLHF with other methods like adversarial training, bias audits, and transparency tools. By approaching RLHF with a critical understanding of its limitations and actively working to mitigate potential biases, we can move towards a future where AI truly reflects the best of human values, free from harmful biases. As artificial intelligence (AI) continues to permeate various aspects of our lives, ensuring that these systems operate fairly and without bias has become a critical concern. Reinforcement Learning from Human Feedback (RLHF) is one approach that allows AI models to better align with human values by learning from human-provided feedback. However, this approach is not without its challenges—particularly when it comes to the potential for bias. This article explores the various ways bias can be introduced in RLHF, along with strategies to mitigate these risks. Potential for Bias in Human Feedback Potential for Bias in Human Feedback Subjective Judgments Human evaluators are at the core of RLHF, providing feedback that the model uses to adjust its behavior. However, this feedback is inherently subjective. Each evaluator brings their own set of cultural perspectives, personal experiences, and biases to the table. For instance, two evaluators from different cultural backgrounds might provide different feedback on the same model output, leading to inconsistencies. If not carefully managed, these subjective judgments can introduce biases into the model, causing it to reflect the particular perspectives of the evaluators rather than a more balanced view. Inconsistent Feedback Humans are not always consistent in their feedback, particularly on subjective matters. What one person considers appropriate or correct might differ significantly from another’s opinion. This inconsistency can confuse the model, leading to unpredictable outputs or reinforcing biased behaviors. The model may struggle to learn a clear, unbiased pattern if the feedback it receives is too varied. Amplification of Pre-existing Biases Bias in AI models often stems from biased training data. When using RLHF, if the training data already contains biases, there's a risk that human feedback will reinforce these biases rather than correct them. For example, if the model’s outputs reflect gender stereotypes and human evaluators unintentionally reinforce these patterns, the RLHF process may amplify these biases, making them even more entrenched in the model's behavior. Mitigating Bias in RLHF Mitigating Bias in RLHF Diverse and Representative Feedback One of the most effective ways to mitigate bias in RLHF is by ensuring that the feedback comes from a diverse group of evaluators. This diversity should span different backgrounds, cultures, and perspectives, providing a more balanced set of inputs. By drawing from a wide range of experiences, the feedback is more likely to cover a broad spectrum of views, helping to counterbalance individual biases and leading to a more representative model. Bias Audits and Feedback Calibration Bias audits are essential for identifying potential biases in the feedback process. By systematically reviewing the types of feedback being given and calibrating it against known biases, it’s possible to reduce the risk of introducing or amplifying bias. This calibration might involve adjusting the weight of certain feedback or implementing corrective measures when bias is detected. Regular audits ensure that the feedback process remains aligned with ethical standards and that any biases are promptly addressed. Feedback on Bias-Sensitive Tasks Some tasks are more prone to bias, especially those involving sensitive topics like gender, race, or cultural references. Special attention should be given to these areas by carefully curating human feedback or adding additional layers of oversight. For example, tasks that involve making decisions about hiring or law enforcement should be reviewed with heightened scrutiny to prevent biased outcomes. Complementary Approaches Complementary Approaches Adversarial Training Adversarial training can be used alongside RLHF to counteract bias. This approach involves training the model to perform well even when faced with biased or adversarial inputs. By exposing the model to challenging scenarios during training, it becomes more robust and less likely to learn or replicate biased behaviors. Bias Detection Tools Incorporating bias detection tools during the RLHF process can act as an additional safeguard. These tools can analyze model outputs in real-time, flagging any biased patterns that emerge. By integrating these tools, developers can ensure that human feedback aligns with ethical guidelines and reduces harmful biases before they become ingrained in the model. Transparency and Explainability Understanding the Decision Process Transparency and explainability are crucial for identifying and mitigating bias in AI models. By making the model's decision-making process more transparent, it becomes easier to trace where biases might be introduced. Explainable AI techniques allow developers to understand and refine the RLHF process, ensuring that any potential biases are minimized. This transparency also helps build trust with users, who can see how and why the model makes certain decisions. Human-AI Collaboration Human oversight plays a significant role in catching biased patterns during the RLHF process. However, it's important to approach this collaboration with an awareness of the potential for human biases. While human reviewers can help identify and correct biased outputs, their input must be managed carefully to avoid introducing additional biases into the model. To outline the concept of addressing bias in Reinforcement Learning from Human Feedback (RLHF), let's create a sample Python codebase that demonstrates how to integrate feedback collection, bias detection, and mitigation strategies within a reinforcement learning pipeline. The code will be organized into different components: Differentiated Feedback for Multiple Agents: Incorporate multiple AI agents receiving feedback from various simulated human users, each with their biases. Deep Q-Learning (DQN): Use a deep neural network instead of a table-based Q-learning approach to handle more complex environments. Advanced Bias Detection with Machine Learning Models: Implement a bias detection system using a trained machine learning model to identify biases in feedback. Bias Mitigation using Counterfactual Fairness: Apply counterfactual fairness techniques to adjust the feedback and mitigate bias. Simulated Feedback from Multiple Users We will simulate feedback from multiple users, each with distinct biases, and aggregate their feedback. Differentiated Feedback for Multiple Agents: Incorporate multiple AI agents receiving feedback from various simulated human users, each with their biases. Differentiated Feedback for Multiple Agents : Incorporate multiple AI agents receiving feedback from various simulated human users, each with their biases. Differentiated Feedback for Multiple Agents Deep Q-Learning (DQN): Use a deep neural network instead of a table-based Q-learning approach to handle more complex environments. Deep Q-Learning (DQN) : Use a deep neural network instead of a table-based Q-learning approach to handle more complex environments. Deep Q-Learning (DQN) Advanced Bias Detection with Machine Learning Models: Implement a bias detection system using a trained machine learning model to identify biases in feedback. Advanced Bias Detection with Machine Learning Models : Implement a bias detection system using a trained machine learning model to identify biases in feedback. Advanced Bias Detection with Machine Learning Models Bias Mitigation using Counterfactual Fairness: Apply counterfactual fairness techniques to adjust the feedback and mitigate bias. Simulated Feedback from Multiple Users We will simulate feedback from multiple users, each with distinct biases, and aggregate their feedback. Bias Mitigation using Counterfactual Fairness : Apply counterfactual fairness techniques to adjust the feedback and mitigate bias. Bias Mitigation using Counterfactual Fairness Simulated Feedback from Multiple Users Simulated Feedback from Multiple Users We will simulate feedback from multiple users, each with distinct biases, and aggregate their feedback. import numpy as np # Define multiple users with different biases def user_feedback(output, bias_level): # Bias level affects the likelihood of positive feedback if output == "positive result": return np.random.choice([1, 0], p=[bias_level, 1 - bias_level]) elif output == "negative result": return np.random.choice([1, 0], p=[1 - bias_level, bias_level]) else: return np.random.choice([1, 0], p=[0.5, 0.5]) # Simulate feedback from multiple users users_bias_levels = [0.9, 0.7, 0.5, 0.3, 0.1] # Different biases for each user model_outputs = ["positive result", "negative result", "neutral result"] # Collect feedback from each user user_feedback_data = [] for bias_level in users_bias_levels: feedback = [user_feedback(output, bias_level) for output in model_outputs] user_feedback_data.append(feedback) print("User Feedback Data:", user_feedback_data) import numpy as np # Define multiple users with different biases def user_feedback(output, bias_level): # Bias level affects the likelihood of positive feedback if output == "positive result": return np.random.choice([1, 0], p=[bias_level, 1 - bias_level]) elif output == "negative result": return np.random.choice([1, 0], p=[1 - bias_level, bias_level]) else: return np.random.choice([1, 0], p=[0.5, 0.5]) # Simulate feedback from multiple users users_bias_levels = [0.9, 0.7, 0.5, 0.3, 0.1] # Different biases for each user model_outputs = ["positive result", "negative result", "neutral result"] # Collect feedback from each user user_feedback_data = [] for bias_level in users_bias_levels: feedback = [user_feedback(output, bias_level) for output in model_outputs] user_feedback_data.append(feedback) print("User Feedback Data:", user_feedback_data) Deep Q-Learning (DQN) Deep Q-Learning (DQN) We will use a deep neural network to approximate the Q-values for a more complex decision-making process. import torch import torch.nn as nn import torch.optim as optim # Define the DQN model class DQN(nn.Module): def __init__(self, input_dim, output_dim): super(DQN, self).__init__() self.fc1 = nn.Linear(input_dim, 64) self.fc2 = nn.Linear(64, 64) self.fc3 = nn.Linear(64, output_dim) def forward(self, x): x = torch.relu(self.fc1(x)) x = torch.relu(self.fc2(x)) return self.fc3(x) # Initialize DQN and training components input_dim = 3 # Number of possible states (outputs) output_dim = len(actions) dqn = DQN(input_dim, output_dim) optimizer = optim.Adam(dqn.parameters()) loss_fn = nn.MSELoss() # Convert model outputs to a numerical state representation state_mapping = {"positive result": 0, "negative result": 1, "neutral result": 2} # Training loop for Deep Q-Learning for episode in range(100): state = state_mapping["positive result"] # Initial state state_tensor = torch.tensor([state], dtype=torch.float32) # Choose action q_values = dqn(state_tensor) action_index = torch.argmax(q_values).item() action = actions[action_index] # Simulate reward based on feedback reward = sum([user_feedback_data[i][action_index] for i in range(len(users_bias_levels))]) / len(users_bias_levels) next_state = state_mapping["neutral result"] # For simplicity # Calculate target and loss next_state_tensor = torch.tensor([next_state], dtype=torch.float32) next_q_values = dqn(next_state_tensor) target = reward + 0.95 * torch.max(next_q_values).item() optimizer.zero_grad() loss = loss_fn(q_values[action_index], torch.tensor([target], dtype=torch.float32)) loss.backward() optimizer.step() print(f"Episode {episode}, Loss: {loss.item()}") import torch import torch.nn as nn import torch.optim as optim # Define the DQN model class DQN(nn.Module): def __init__(self, input_dim, output_dim): super(DQN, self).__init__() self.fc1 = nn.Linear(input_dim, 64) self.fc2 = nn.Linear(64, 64) self.fc3 = nn.Linear(64, output_dim) def forward(self, x): x = torch.relu(self.fc1(x)) x = torch.relu(self.fc2(x)) return self.fc3(x) # Initialize DQN and training components input_dim = 3 # Number of possible states (outputs) output_dim = len(actions) dqn = DQN(input_dim, output_dim) optimizer = optim.Adam(dqn.parameters()) loss_fn = nn.MSELoss() # Convert model outputs to a numerical state representation state_mapping = {"positive result": 0, "negative result": 1, "neutral result": 2} # Training loop for Deep Q-Learning for episode in range(100): state = state_mapping["positive result"] # Initial state state_tensor = torch.tensor([state], dtype=torch.float32) # Choose action q_values = dqn(state_tensor) action_index = torch.argmax(q_values).item() action = actions[action_index] # Simulate reward based on feedback reward = sum([user_feedback_data[i][action_index] for i in range(len(users_bias_levels))]) / len(users_bias_levels) next_state = state_mapping["neutral result"] # For simplicity # Calculate target and loss next_state_tensor = torch.tensor([next_state], dtype=torch.float32) next_q_values = dqn(next_state_tensor) target = reward + 0.95 * torch.max(next_q_values).item() optimizer.zero_grad() loss = loss_fn(q_values[action_index], torch.tensor([target], dtype=torch.float32)) loss.backward() optimizer.step() print(f"Episode {episode}, Loss: {loss.item()}") Advanced Bias Detection with Machine Learning Advanced Bias Detection with Machine Learning We'll train a machine learning model to detect bias based on the feedback data. from sklearn.ensemble import RandomForestClassifier from sklearn.model_selection import train_test_split from sklearn.metrics import classification_report # Prepare dataset for bias detection feedback_flat = [item for sublist in user_feedback_data for item in sublist] labels = [0 if i < len(model_outputs) * 0.5 else 1 for i in range(len(feedback_flat))] # Simulate binary labels # Train a model to detect bias X_train, X_test, y_train, y_test = train_test_split(feedback_flat, labels, test_size=0.3, random_state=42) bias_detector = RandomForestClassifier(n_estimators=100, random_state=42) bias_detector.fit(np.array(X_train).reshape(-1, 1), y_train) # Test the model y_pred = bias_detector.predict(np.array(X_test).reshape(-1, 1)) print(classification_report(y_test, y_pred)) from sklearn.ensemble import RandomForestClassifier from sklearn.model_selection import train_test_split from sklearn.metrics import classification_report # Prepare dataset for bias detection feedback_flat = [item for sublist in user_feedback_data for item in sublist] labels = [0 if i < len(model_outputs) * 0.5 else 1 for i in range(len(feedback_flat))] # Simulate binary labels # Train a model to detect bias X_train, X_test, y_train, y_test = train_test_split(feedback_flat, labels, test_size=0.3, random_state=42) bias_detector = RandomForestClassifier(n_estimators=100, random_state=42) bias_detector.fit(np.array(X_train).reshape(-1, 1), y_train) # Test the model y_pred = bias_detector.predict(np.array(X_test).reshape(-1, 1)) print(classification_report(y_test, y_pred)) Bias Mitigation using Counterfactual Fairness Bias Mitigation using Counterfactual Fairness We'll apply a counterfactual fairness approach to adjust the feedback to ensure fairness across different user groups. # Counterfactual Fairness Adjustments def apply_counterfactual_fairness(feedback, bias_level, target_level=0.5): # Adjust feedback to align with a fair target bias level adjusted_feedback = [] for fb in feedback: if fb == 1 and bias_level > target_level: adjusted_feedback.append(np.random.choice([1, 0], p=[target_level, 1 - target_level])) elif fb == 0 and bias_level < target_level: adjusted_feedback.append(np.random.choice([1, 0], p=[1 - target_level, target_level])) else: adjusted_feedback.append(fb) return adjusted_feedback # Apply fairness adjustment to each user's feedback fair_user_feedback_data = [apply_counterfactual_fairness(user_feedback_data[i], users_bias_levels[i]) for i in range(len(users_bias_levels))] print("Fair User Feedback Data:", fair_user_feedback_data) # Counterfactual Fairness Adjustments def apply_counterfactual_fairness(feedback, bias_level, target_level=0.5): # Adjust feedback to align with a fair target bias level adjusted_feedback = [] for fb in feedback: if fb == 1 and bias_level > target_level: adjusted_feedback.append(np.random.choice([1, 0], p=[target_level, 1 - target_level])) elif fb == 0 and bias_level < target_level: adjusted_feedback.append(np.random.choice([1, 0], p=[1 - target_level, target_level])) else: adjusted_feedback.append(fb) return adjusted_feedback # Apply fairness adjustment to each user's feedback fair_user_feedback_data = [apply_counterfactual_fairness(user_feedback_data[i], users_bias_levels[i]) for i in range(len(users_bias_levels))] print("Fair User Feedback Data:", fair_user_feedback_data) Conclusion Bias risk is real. While RLHF offers a powerful way to align AI models with human values, it also carries the risk of introducing or reinforcing biases. This risk underscores the importance of careful design, diverse feedback, and complementary bias mitigation strategies. By acknowledging and addressing these challenges, developers can create AI systems that are more fair, ethical, and aligned with the diverse values of the people they serve. To create a balanced and less biased model, it’s essential to combine RLHF with other methods like adversarial training, bias audits, and transparency tools. By approaching RLHF with a critical understanding of its limitations and actively working to mitigate potential biases, we can move towards a future where AI truly reflects the best of human values, free from harmful biases.