AI Governance: Key Terms

AI Governance Key Terms

• Accountability: Ensures AI systems operate ethically, fairly, transparently, and compliantly, tracing actions back to the responsible entity; includes obligations and responsibilities of AI developers and deployers to ensure ethical operation, fairness, transparency, and compliance with regulations.
• Accuracy: Measures how correctly an AI system performs its task; critical for AI model reliability, especially in high-precision applications like medical diagnoses; measures system performance and effectiveness in producing correct outputs from input data.
• Active Learning: A machine learning subfield where algorithms selectively learn from data by requesting specific data points to improve learning.
• Adaptive Learning: A method that customizes educational content to meet individual student needs and learning styles; provides personalized and optimized learning catering to diverse learning styles.
• Adversarial Attack: Exploits AI model vulnerabilities via malicious input data, causing malfunctions and unsafe outputs; manipulating inputs can cause malfunctions, like fooling a self-driving car into perceiving a stop light as green, impacting safety.
• AI Assurance: Frameworks and controls ensuring safe, reliable, and trustworthy AI; includes conformity, impact, and risk assessments, AI audits, certifications, testing, evaluation, and compliance w/ standards.
• AI Audit: Review of AI systems ensuring intended operation and legal compliance, identifying risks, and offering mitigation; review and assessment of an AI system ensures it operates as intended, complies with laws, regulations, standards, helps identify and map risks and offer mitigation strategies
• AI Governance: Laws, policies, and processes that manage AI tech development ethically and responsibly at international, national, and organizational levels; helps stakeholders implement, manage, oversee, regulate AI technology, manages associated risks to ensure alignment with objectives, ethical development, legal and regulatory compliance
• Algorithm: A procedure or set of rules designed to perform a specific task or solve a computer problem.
• Artificial General Intelligence (AGI): AI with human-level intelligence and broad generalization capability; contrasted with "narrow" AI
• Artificial Intelligence (AI): Engineered systems that use computation to perform tasks, simulate intelligence, and automate; a field of computer science that simulates intelligent behavior, may include automated decision making.
• Automated Decision-Making: Making choices via AI without human intervention; a process of making decisions by technological means either wholly or partly.
• Bias: Systematic errors in AI predictions due to model assumptions or flawed data; computational or machine bias is a systematic error from a model's assumptions or the data itself, can impact outcomes and risk individual rights and liberties; cognitive bias refers to inaccurate judgement or distorted thinking, societal bias leads to systemic prejudice, favoritism, discrimination (selection bias ie biases in selecting data for model training)
• Bootstrap Aggregating: A machine learning method improving model stability and accuracy by averaging multiple model versions trained on random data subsets; also known as bagging
• Chatbot: AI simulating human conversations through natural language processing and deep learning; designed to simulate human-like conversations and interactions, uses natural language processing, deep learning to understand and respond to text or speech.
• Classification Model: Categorizes input data in machine learning; sometimes referred to as classifiers
• Clustering: An unsupervised machine learning to group similar data points; sometimes referred to as clustering algorithms
• Compute: Processing resources in a computer system, essential for data handling, applications, and visual rendering; includes hardware components such as the central processing unit or graphics processing unit that is essential for memory, storage, processing data, running applications, rendering graphics etc.
• Computer Vision: AI field for processing images/videos, enabling applications like facial/object recognition; uses computers to process and analyze images, videos, and other visual inputs
• Conformity Assessment: Analysis to determine if an AI system meets requirements for risk management, data governance, and cybersecurity; often performed by an entity independent of a model developer in order to establish a risk management system, data governance, record-keeping, transparency, and cybersecurity practices
• Contestability: Ensuring AI systems' decisions are questioned or challenged, promoting accountability; helps promote accountability in AI governance, outputs and actions depend on transparency, ability to contest or challenge the outcomes (principle of ensuring AI systems, their decision-making processes can be questioned/challenged by humans) may be called redress
• Corpus: A data collection used by computers to find patterns/make predictions; may include structured or unstructured data and cover a specific topic or a variety of topics
• Data Leak: Unintentional exposure of sensitive data due to security lapses or human error; can be from poor security defenses, human error, storage misconfigurations or a lack of robust policies around internal and external data sharing practices; a data breach is intentional and done in bad faith
• Data Poisoning: Adversarial attacks that inject false data to corrupt the training process; involves malicious users injecting false data into a model, leading to compromised performance, unintentional, misleading, or harmful outputs
• Data Provenance: Tracking data history to ensure integrity, transparency, and governance; tracks and logs the history and origin of records in a dataset, encompassing the entire life cycle from its creation/collection to its transformation to its current state, essential for data transparency and governance and promote a better understanding for the data.
• Data Quality: How well a dataset meets requirements for its intended use and is accurate, complete, valid, consistent and timely; high quality is accurate, complete, valid, consistent, timely, fit for purpose, directly impacts the quality of AI outputs and the system performance.
• Decision Tree: A machine learning model showing branching decisions and their potential outcomes; supervised learning that represents decisions and their potential consequences in a branching structure
• Deep Learning: AI subfield using neural networks, excels in processing raw data like images/speech; a subfield of AI and machine learning that uses artificial neural networks
• Deepfakes: AI-altered audio or video for misinformation.
• Diffusion Model: Generates images by refining a noise signal.
• Discriminative Model: Directly maps input features to class labels in machine learning; used in machine learning that directly maps input features to class labels and analyzes for patterns, is often used for text classification tasks, like identifying the language of a piece of text or detecting spam; examples are traditional neural networks and random forests
• Disinformation: Manipulated content intended for harm; can be spread through deepfakes who have malicious intentions
• Entropy: Measures unpredictability in machine learning datasets; a higher entropy signifies greater uncertainty in predicting outcomes
• Expert System: Emulates human decision-making in fields like medical diagnoses using a rule-based AI; a form of rules based AI that draws inferences from a knowledge base provided by human experts to replicate decision-making abilities within a specific field such as medical diagnoses
• Explainability: Providing sufficient information about the how an AI system works, to maintain transparency and trust; important in maintaining transparency and trust in AI
• Exploratory Data Analysis: Techniques to gain insights into a dataset, uncover relationships, and identify anomalies before training a machine learning model.
• Fairness: An AI quality that prioritizes equal treatment in decisions, avoiding adverse impacts on sensitive attributes like race or gender; system decisions should not adversely impact, whether directly, disparately on sensitive attributes
• Federated Learning: Training models on decentralized data while preserving privacy; allows models to be trained on the local data of multiple edge devices - only the updates of the local model are sent to a central location to improve the global model
• Fine-tuning: Specialized training of pre-trained learning models to make them perform a specialized task; a process taking a pretrained deep learning model and training it further for a specialized task through supervised learning, involves taking a foundation model that has already learned general patterns to train it for a specific task using a much smaller and labelled dataset
• Foundation Model: Large-scale model trained on diverse datasets to be a base for specific applications; a large-scale model that has been trained on extensive and diverse datasets to enable broad capabilities, can function as the base for use-specific applications, also called general purpose AI model and frontier AI.
• Generalization: The ability of a model to apply learned patterns to new, unseen data.
• Generative AI: AI for creating content from large datasets, responding to prompts; uses deep learning trained on large datasets to create content such as data, images, code, music, simulations, videos in response to user prompts, unlike discriminative models it makes predictions existing data vs new data
• Greedy Algorithms: Algorithms that make immediate, optimal choices without considering long-term solutions.
• Ground Truth: True state against which AI is evaluated; serves as the real-world reference against which the outputs are measured for accuracy and reliability.
• Hallucinations: Factually incorrect information created by AI; also called confabulations
• Human-Centric AI: Prioritizes human well-being and autonomy; an approach to AI design, development, deployment and use in order to amplify and augment human abilities vs undermining them
• Human-in-the-Loop (HITL): AI design including human oversight; a design paradigm that incorporates human oversight, intervention, interaction or control over the operation and decision-making processes of an AI system
• Impact Assessment: Process to identify the ethical, legal and societal impact of AI; an evaluation process designed to document and mitigate the potential implications of an AI system for a specific use case.
• Inference: Model-based predictions from input data; a type of machine learning process used to make predictions or decisions based on input data.
• Input Data: Data used by a learning algorithm to produce an output; the basis for machine learning models to learn, make predictions and carry out tasks.
• Interpretability: Ability to explain a model's reasoning in human terms; involves designing models that inherently facilitate understanding through their structure, emphasizes designing models that facilitate understanding through their structure, features or algorithms, requires significant domain expertise
• Large Language Model (LLM): Utilizes deep learning to analyze patterns in text-based tasks; a form of AI that utilizes deep learning algorithms to create models (see also machine learning model, foundation model and fine-tuning) pretrained on massive text datasets, performs text-based tasks; generally, two types: generative and discriminative models
• Machine Learning (ML): Algorithms that learn and improve from data to make decisions; implements various algorithms that learn and improve by experience, includes data collection and preparation, feature engineering, training, testing and validation
• Machine Learning Model: Representation of data patterns used for future predictions; a learned representation of underlying patterns and relationships in data, created via an AI algorithm to a training dataset, can be used to make predictions or perform tasks on new, unseen data
• Misinformation: False content spread without harmful intent; can be spread unintentionally through deepfakes by those who lack harm
• Model Card: A document with information about the purpose and performance of an AI model; discloses information about an AI model like explanations about intended use, performance metrics benchmarked in various conditions
• Multimodal Models: Processing multiple data types at once; can process more than one type of output, has various uses like image captioning and speech recognition
• Natural Language Processing (NLP): Helps computers understand human language; helps computers to understand, interpret, and apply human language by transforming into content and enabling machines to translate language, read text or spoken language etc
• Neural Networks: Models mimicking brain structure for complex tasks; the layered approach enables machine-based neural networks to model complex nonlinear relationships and patterns within data
• Open-Source Software: Promotes transparency and collaboration; a decentralized development model that provides free and open access to source code, promotes shared collaboration and learning
• Overfitting: Model that is too specific to training data and fails on new data; unable to generalize to unseen data
• Oversight: Supervising AI for regulatory compliance and risk management; important for effective AI governance
• Parameters: Internal variables adjusted during training; the model adjusts to, during the training process so it can make predictions on new data.
• Post-Processing: Adjusting model outputs.
• Preprocessing: Preparing data for machine learning; can include cleaning, handling missing values, normalizing, extracting features, etc
• Prompt: An input or instruction provided to an Al model or system to generate an output.
• Prompt Engineering: Structuring prompts to optimize desired outputs; a deliberate process of structuring prompts or series of prompts to influence model behavior to generate more desirable outputs.
• Random Forest: A supervised machine learning algorithm using multiple decision trees to increase accuracy and stability in prediction; helpful to use with datasets that are missing values or are very complex
• Red Teaming: Testing AI safety and security through simulated adversarial attacks; reveals security risks, model flaws, biases etc and the results of such testing are passed onto the model developers for evaluation and remediation.
• Reinforcement Learning: Training a model to optimize actions via reward.
• Reinforcement Learning with Human Feedback (RLHF): Learning through human preferences; human feedback is provided on the model's output to help align the AI's behavior.
• Reliability: Consistent and accurate AI performance; ensures it behaves as expected and performs its intended function even with new data
• Robotics: Designing robots that interact with the physical world
• Robustness: AI's resilience and functionality under attack; ensures system's functionality, performance and accuracy in variety of environments/circumstances even when faced with changed inputs or security attacks
• Safety: Minimizing AI harms and risks; also encompasses the prevention of existential or unexpected risks
• Semi-Supervised Learning: Combines supervised and unsupervised learning; generative AI commonly relies on semi-supervised learning
• Small Language Models: LLMs optimized for efficiency; optimize them for efficiency and better suiting them for deployment in environments that need faster training
• Supervised Learning: Trained on labeled data; can be useful for classification or regression
• Synthetic Data: Artificially created data resembling real data; often used for testing or training machine learning models, particularly in cases with limited, unavailable, sensitive data.
• System Card: Documents on how AI models collaborate; discloses information about how various AI models work together within a network of AI systems, promotes explainability of the overall system
• Testing Data: Assessing model performance.
• Training Data: Datasets to train a model.
• Transfer Learning Model: Using knowledge learned from one task to perform another.
• Transformer Model: Neural network focusing on sequence data relationships; by attending to the surrounding words a word can be comprehended
• Transparency: Comprehensibility and openness in AI; implies openness, comprehensibility and accountability in the way AI algorithms function and make decisions, specific meaning can vary
• Trustworthy AI: Ethical, secure, and accountable AI; interchangeable with terms responsible AI and ethical AI, refer to principle-based AI development and AI governance,
• Turing Test: Measuring AI's ability to exhibit human behavior; originally thought of the test to be an AI's ability to converse through a written text, such that a human reader would not be able to tell a computer-generated response from that of a human
• Underfitting: Model failing to capture training data complexity.
• Unsupervised Learning: Discovering patterns in unlabeled data; with minimal human supervision, the provided AI is provided with preexisting unlabeled datasets and then analyzes those for patterns
• Validation Data: Assessing model performance during training; and preventing overfitting
• Variables: Measurable attribute with different values; can be numerical/quantitative or categorical/qualitative, also called features.
• Variance: How data spreads from the average.
• Watermarking: Embedding patterns in AI-generated content for transparency.