AI Literacy

Q.1 What is Artificial Intelligence (AI)?

AI refers to computer systems designed to perform tasks that typically require human intelligence — such as understanding language, recognizing images, making decisions, or solving problems — by learning patterns from data or following programmed rules.

Report This Question

Q.2 What is the difference between AI, Machine Learning, and Deep Learning?

AI is the broad field of making machines act intelligently. Machine Learning (ML) is a subset of AI where systems learn patterns from data rather than being explicitly programmed. Deep Learning is a subset of ML that uses layered neural networks to learn complex patterns, especially effective for tasks like image and speech recognition.

Report This Question

Q.3 What is a chatbot, and how is it related to AI?

A chatbot is a software application that simulates conversation with users, often via text or voice. Modern chatbots are powered by AI — particularly language models — that let them understand user input and generate relevant, human-like responses.

Report This Question

Q.4 What is training data, and why is it important for AI systems?

Training data is the set of examples an AI model learns from to identify patterns and make predictions. It's important because the quality, quantity, and diversity of training data directly affects how accurate, fair, and reliable the resulting AI system is.

Report This Question

Q.5 What does it mean when we say an AI model is "biased"?

It means the model produces systematically unfair or skewed outputs — often because the data it was trained on reflected existing imbalances or prejudices — leading to results that disadvantage certain groups or situations.

Report This Question

Q.6 What is a Large Language Model (LLM)?

An LLM is a type of AI model trained on massive amounts of text data to understand and generate human language. It can answer questions, write text, summarize information, and hold conversations by predicting likely word sequences based on patterns it learned during training.

Report This Question

Q.7 What is the difference between a rule-based system and a machine learning system?

A rule-based system follows explicit, hand-coded instructions ("if X, then Y") written by humans. A machine learning system instead learns patterns and rules automatically from data, allowing it to handle situations not explicitly programmed by a developer.

Report This Question

Q.8 What is meant by "hallucination" in the context of AI, particularly with LLMs?

Hallucination refers to an AI model generating information that sounds plausible and confident but is factually incorrect or fabricated — a known limitation of language models that can occur because they predict likely-sounding text rather than verified facts.

Report This Question

Q.9 What is the difference between narrow AI and general AI?

Narrow AI (or "weak AI") is designed to perform a specific task well (e.g., facial recognition, spam filtering) and can't generalize beyond it. General AI (or "strong AI") refers to a hypothetical system with human-like, broad intelligence capable of performing any intellectual task — which does not currently exist.

Report This Question

Q.10 Why is it important to fact-check information generated by an AI system?

Because AI systems, especially language models, can produce inaccurate, outdated, or fabricated information confidently and without indicating uncertainty, so relying on AI output without verification can lead to spreading misinformation or making poor decisions.

Report This Question

Q.11 What is supervised learning, and how does it differ from unsupervised learning?

Supervised learning trains a model on labeled data (input-output pairs), so it learns to map inputs to known correct outputs (e.g., classifying emails as spam/not spam). Unsupervised learning works with unlabeled data, where the model identifies patterns or groupings on its own (e.g., clustering customers by purchasing behavior) without predefined correct answers.

Report This Question

Q.12 What is a neural network, explained conceptually?

A neural network is a computing structure loosely inspired by the brain, made up of layers of interconnected "neurons" (nodes) that process input data by applying weighted connections and mathematical functions, progressively transforming raw input into a useful output (e.g., a classification or prediction).

Report This Question

Q.13 What is the difference between a model and an algorithm?

An algorithm is the set of rules or mathematical procedures used to process data and learn (e.g., gradient descent, decision trees). A model is the resulting artifact created after an algorithm has been trained on data — it's the learned representation that can then be used to make predictions on new inputs.

Report This Question

Q.14 What is "prompt engineering," and why does it matter for using LLMs effectively?

Prompt engineering is the practice of carefully crafting inputs (prompts) to guide an AI language model toward producing more accurate, relevant, or useful outputs. It matters because the way a question or instruction is phrased significantly affects the quality and relevance of the model's response.

Report This Question

Q.15 What is the difference between generative AI and predictive AI?

Predictive AI analyzes existing data to forecast an outcome or classify something (e.g., predicting loan default risk). Generative AI creates new content — text, images, audio, code — that didn't exist before, based on patterns learned from training data (e.g., ChatGPT generating an essay, or DALL-E generating an image).

Report This Question

Q.16 What is overfitting in machine learning?

Overfitting occurs when a model learns the training data too closely — including its noise and quirks — resulting in excellent performance on training data but poor performance on new, unseen data because it failed to generalize.

Report This Question

Q.17 What are "parameters" in the context of an AI model, and why do people talk about models having "billions of parameters"?

Parameters are the internal, adjustable values a model learns during training that determine how it processes input to produce output (e.g., weights in a neural network). Models with more parameters generally have greater capacity to learn complex patterns, which is why parameter count is often (though not perfectly) used as a rough indicator of a model's scale and potential capability.

Report This Question

Q.18 What is the difference between AI "understanding" language and simply predicting the next word?

Technically, most LLMs work by predicting the statistically likely next word/token based on patterns learned from vast text data — they don't "understand" in the human sense of conscious comprehension. However, this prediction process, at scale, can produce outputs that appear to reflect understanding, which is a nuanced and actively debated topic in AI research.

Report This Question

Q.19 What is a "model's context window," and why does it matter?

The context window is the amount of text (measured in tokens) an AI model can consider at once when generating a response — including the conversation history and any provided documents. It matters because information outside this window is effectively "forgotten" by the model during that interaction.

Report This Question

Q.20 What is the difference between AI ethics and AI safety?

AI ethics focuses on ensuring AI is developed and used in line with moral principles — fairness, accountability, transparency, avoiding harm to individuals/society. AI safety focuses more specifically on ensuring AI systems behave reliably and as intended, avoiding unintended or catastrophic failures, especially as systems become more capable and autonomous.

Report This Question

Q.21 What does "explainability" mean for an AI system, and why does it matter in high-stakes decisions?

Explainability refers to the ability to understand and articulate why an AI system produced a particular output or decision. It matters in high-stakes contexts (e.g., loan denials, medical diagnoses) because affected individuals and regulators often need to understand the reasoning to ensure fairness, trust, and accountability.

Report This Question

Q.22 What is the difference between an API and a user-facing AI application?

An API (Application Programming Interface) is a way for developers to programmatically access an AI model's capabilities within their own software, without a visual interface. A user-facing application (like a chatbot website or app) provides a direct interface for end users to interact with the AI, typically built on top of an API.

Report This Question

Q.23 What is "fine-tuning" a model, and how does it differ from training from scratch?

Fine-tuning takes an already-trained (pre-trained) model and further trains it on a smaller, specific dataset to adapt it to a particular task or domain. Training from scratch means building a model's knowledge entirely from an initial, untrained state, which requires vastly more data, time, and computational resources.

Report This Question

Q.24 What is "multimodal AI"?

Multimodal AI refers to systems that can process and/or generate multiple types of data — such as text, images, audio, and video — together, rather than being limited to just one format (e.g., a model that can look at an image and answer questions about it in text).

Report This Question

Q.25 What are some common limitations of current AI systems that users should be aware of?

Common limitations include: hallucinating incorrect information confidently, reflecting biases present in training data, lacking true reasoning/common sense in some situations, having a fixed knowledge cutoff (not knowing recent events unless connected to live tools), and being sensitive to how questions are phrased.

Report This Question

Q.26 What is the difference between a "closed-source" and "open-source" AI model?

A closed-source model's underlying code and/or weights are proprietary and not publicly released — accessed typically via an API or paid product. An open-source (or open-weight) model has its weights and/or code made publicly available, allowing anyone to inspect, modify, or run it themselves.

Report This Question

Q.27 What does "AI bias mitigation" involve at a high level?

It involves identifying potential sources of unfairness (e.g., unrepresentative training data, proxy variables correlated with protected characteristics), applying techniques to reduce bias (e.g., re-balancing data, adjusting model outputs, fairness-aware algorithms), and continuously testing/monitoring the model's outputs across different demographic groups.

Report This Question

Q.28 What is the difference between "AI-assisted" and "AI-automated" decision-making?

AI-assisted decision-making means the AI provides recommendations or analysis, but a human makes the final decision. AI-automated decision-making means the AI system makes the decision directly, often without human review, which typically requires higher scrutiny and stronger safeguards given the reduced human oversight.

Report This Question

Q.29 Why do organizations need an AI usage policy for employees?

An AI usage policy sets clear guidelines on which AI tools are approved for use, what data can/cannot be input into them (to avoid leaking sensitive/confidential information), expectations around verifying AI-generated content, and accountability for decisions made with AI assistance — reducing legal, security, and reputational risks.

Report This Question

Q.30 What is "data privacy" risk when using AI tools, and give an example.

Data privacy risk arises when personal or confidential information is input into an AI tool and could be stored, used for further model training, or exposed to unauthorized parties. For example, pasting confidential client data into a public AI chatbot could result in that data being retained or inadvertently surfaced elsewhere.

Report This Question

Q.31 Explain the concept of "emergent behavior" in large AI models and why it's significant.

Emergent behavior refers to capabilities that appear in large models but were not explicitly present in smaller versions of the same architecture — such as complex reasoning or few-shot learning suddenly improving past a certain scale threshold. It's significant because it makes model capabilities harder to predict in advance, complicating both development planning and risk assessment.

Report This Question

Q.32 What is the difference between "correlation" and "causation," and why is this distinction critical when interpreting AI/ML model outputs?

Correlation means two variables move together statistically; causation means one variable directly influences the other. It's critical for AI because models often identify correlations in data without understanding true causal relationships, which can lead to spurious or misleading conclusions if decisions are made assuming causation where only correlation exists.

Report This Question

Q.33 What are "adversarial examples" in machine learning, and why do they matter for AI safety/security?

Adversarial examples are inputs deliberately crafted (often with subtle, human-imperceptible modifications) to cause an AI model to make an incorrect prediction or classification. They matter because they reveal that models can be fragile or exploitable in ways that don't align with human perception, posing security risks in applications like facial recognition, autonomous vehicles, or content moderation.

Report This Question

Q.34 What is the concept of "alignment" in AI, and why is it considered a significant challenge?

Alignment refers to ensuring an AI system's goals, behaviors, and outputs match human intentions and values, especially as systems become more capable and autonomous. It's challenging because it's difficult to fully specify complex human values and preferences in a way a model can reliably learn and generalize, and because a model optimizing for a proxy objective can produce unintended or harmful behavior even while technically achieving its stated goal.

Report This Question

Q.35 Explain the difference between "explainable AI" (XAI) and "interpretable-by-design" models, with a trade-off discussion.

Interpretable-by-design models (e.g., decision trees, linear regression) have transparent internal logic that humans can directly follow. XAI techniques (e.g., SHAP, LIME) instead provide after-the-fact approximate explanations for complex "black-box" models (like deep neural networks) that aren't inherently interpretable. The trade-off: interpretable models are often less accurate on complex tasks, while black-box models with XAI achieve higher performance but their explanations are approximations, not exact descriptions of the model's true internal reasoning.

Report This Question

Q.36 What is "reinforcement learning from human feedback" (RLHF), and how has it shaped modern LLM behavior?

RLHF is a training technique where human evaluators rank or rate model outputs, and this feedback is used to further train the model (often via a reward model) to produce outputs better aligned with human preferences — such as being more helpful, honest, and harmless. It's a major reason modern LLM-based assistants behave conversationally and follow instructions well, compared to earlier raw language models that simply predicted text continuations.

Report This Question

Q.37 What is the "black box problem" in AI, and what practical challenges does it create for regulated industries?

The black box problem refers to the difficulty of understanding exactly how a complex model (e.g., deep neural network) arrives at a specific output, since its internal decision process isn't directly human-interpretable. In regulated industries (finance, healthcare, insurance), this creates challenges around meeting legal explainability requirements, auditing for bias/compliance, and building stakeholder trust, sometimes forcing organizations to choose more interpretable but less accurate models for high-stakes decisions.

Report This Question

Q.38 Discuss the concept of "AI scaling laws" and their implications for the future development of AI capabilities.

Scaling laws describe empirically observed relationships showing that model performance tends to improve predictably as model size, training data, and compute increase, following certain mathematical patterns. Implications include: it has driven the industry trend toward ever-larger models, though it raises questions about diminishing returns, the sustainability of compute/energy costs, and whether scaling alone will continue yielding proportional capability gains or whether new architectural breakthroughs will be needed.

Report This Question

Q.39 What is "catastrophic forgetting" in neural networks, and why is it relevant to continual learning?

Catastrophic forgetting occurs when a neural network, after being trained on new data or tasks, loses previously learned knowledge/performance on earlier tasks — because the new training updates overwrite the weights that encoded the earlier learning. It's relevant to continual learning because building AI systems that can keep learning new things over time without losing prior capabilities remains a significant unsolved challenge.

Report This Question

Q.40 Explain the difference between "narrow" fine-tuning and "instruction tuning" of a base language model.

Narrow fine-tuning adapts a pre-trained model to perform well on a specific, often narrow task or domain (e.g., legal document classification). Instruction tuning trains a base model on a broad, diverse set of instruction-response examples to make it generally better at following varied user instructions across many tasks — it's a key step in turning a raw base language model into a general-purpose assistant.

Report This Question

Q.41 What is "model interpretability at scale," and why is it harder for very large models than smaller ones?

It refers to the ability to understand the internal reasoning of a model as it grows in size and complexity. It's harder for large models because they contain billions of parameters with highly distributed, non-linear representations, making it computationally and conceptually difficult to trace how specific inputs map to specific outputs — an active area of research known as "mechanistic interpretability."

Report This Question

Q.42 What is "dataset shift" (or distribution shift), and how does it affect deployed AI models over time?

Dataset shift occurs when the statistical properties of real-world data encountered by a deployed model diverge from the data it was originally trained on (e.g., changing customer behavior, new slang, evolving market conditions). This can cause the model's performance to silently degrade over time if not monitored and periodically retrained or recalibrated.

Report This Question

Q.43 Discuss the ethical and practical challenges of using AI systems for high-stakes decisions like hiring, lending, or criminal justice.

Challenges include: risk of encoding and amplifying historical biases present in training data, lack of transparency/explainability undermining due process and appeal rights, difficulty establishing accountability when decisions are contested, potential for disparate impact on protected groups even without explicit discriminatory intent, and the risk of over-reliance on AI recommendations without adequate human judgment (automation bias).

Report This Question

Q.44 What is the difference between "weak supervision" and "self-supervised learning," and why has self-supervised learning become foundational to modern LLMs?

Weak supervision uses noisy, indirect, or heuristic-based labels (rather than fully accurate human-labeled data) to train models more cheaply at scale. Self-supervised learning generates its own training signal from unlabeled data itself (e.g., predicting a masked word in a sentence), requiring no manual labeling at all. Self-supervised learning became foundational to LLMs because it allows training on massive amounts of raw internet text without the impractical cost of human-labeling it all.

Report This Question

Q.45 What is "tokenization," and why do quirks in tokenization (e.g., difficulty with counting letters) explain certain odd LLM behaviors?

Tokenization is the process of breaking text into smaller units (tokens) — which may be whole words, subwords, or characters — that a language model processes numerically. Certain odd behaviors (e.g., an LLM struggling to count the letters in a word) happen because the model doesn't see individual characters directly; it sees tokens, which may represent multiple characters or word fragments, making character-level tasks inherently harder for it to perform accurately.

Report This Question

Q.46 Discuss the concept of "AI arms race" dynamics between organizations, and what risks this competitive pressure can introduce.

Competitive pressure to be first-to-market or to out-perform rivals can incentivize organizations to cut corners on safety testing, rush deployment of insufficiently evaluated systems, or under-invest in governance/oversight in favor of speed — raising the risk of releasing systems with unaddressed harms, security vulnerabilities, or unintended societal consequences.

Report This Question

Q.47 What is "the alignment tax," and how might it influence organizational decisions about deploying AI safety measures?

The alignment tax refers to the potential cost (in performance, capability, speed, or resources) an organization incurs by implementing safety/alignment measures on an AI system (e.g., stricter content filters reducing helpfulness). This can create organizational tension where business incentives to maximize capability/performance compete with the resources allocated to safety — often requiring leadership commitment to accept some trade-offs for long-term trust and risk mitigation.

Report This Question

Q.48 What is "few-shot learning," and how does it differ from traditional supervised learning approaches?

Few-shot learning refers to a model's ability to learn or adapt to a new task from only a handful of examples provided within the prompt (in-context), without updating the model's underlying weights. Traditional supervised learning instead requires substantial labeled datasets and an explicit retraining/fine-tuning process to adapt to a new task.

Report This Question

Q.49 Explain the concept of "AI democratization" and discuss both its benefits and potential risks.

AI democratization refers to making AI tools and capabilities broadly accessible (via low-cost APIs, open-source models, no-code platforms) rather than confined to well-resourced organizations. Benefits include broader innovation, reduced barriers to entry, and wider societal benefit. Risks include increased potential for misuse by malicious actors, a wider surface for security/privacy harms, and challenges in ensuring consistent safety standards across a much larger and more diverse set of deployers.

Report This Question

Q.50 What is the difference between "capability" and "alignment" as two distinct axes of AI progress, and why is it important to consider them separately?

Capability refers to what an AI system is able to do (its raw performance, reasoning ability, task range). Alignment refers to whether the system reliably does what humans actually want it to do, safely and as intended. It's important to consider them separately because increasing capability without corresponding improvements in alignment can widen the gap between what a system can do and what it should do, potentially increasing risk even as the system becomes more impressive or useful.

Report This Question