Most people who use AI tools every day have hit these problems without knowing the name for them. The chatbot that forgets what you pasted. The model that scores brilliantly in testing and collapses in production. The confident answer with an invented citation.
These 120 questions target that gap between experience and vocabulary. They are not definition drills — each one describes a situation and asks what went wrong, which is a harder and more useful thing to be able to do. The security questions in particular reward knowing why a control exists rather than what it is called.
Every question below shows all four options, the keyed answer, and the one-line explanation. If you want to turn these into video, the OmniFrame Quiz Studio is free and this pack is built in.
Question 1
Two AI tools answer a legal question. One invents cases; the other requests sources. Which is safer?
Why: A confident answer is not proof. Source-grounded uncertainty is safer than fluent hallucination.
Question 2
A chatbot remembers your first message, then suddenly forgets the document you pasted earlier. Nothing crashed. What invisible limit did you hit?
Why: The model can only keep a limited amount of text in active context.
Question 3
A strong password is reused everywhere. One site is breached, then the cloud account is stolen. What failed?
Why: A strong password still fails if the same one is reused on a breached site.
Question 4
A team runs the same AI prompt twice. One answer is careful, the next is wildly creative. Which setting probably changed?
Why: Temperature controls randomness and variation in model output.
Question 5
Local AI is private but slower; cloud AI is faster but sends data away. What is the tradeoff?
Why: Local compute keeps data under control; cloud compute usually trades privacy for speed.
Question 6
A developer shares a screenshot containing a secret string. Strangers use the paid account. What leaked?
Why: API keys act like software passwords and should never appear in public screenshots.
Question 7
An AI summarizes 200 pages but misses a warning on page 187. What went wrong?
Why: Long documents can exceed what the model actually considers unless the workflow chunks and checks every section.
Question 8
A support email links to a login domain with one changed letter. What attack is this?
Why: The attacker uses a convincing message and lookalike domain to steal credentials.
Question 9
A customer message tells an AI to ignore its rules and reveal them. What attack is this?
Why: Prompt injection hides malicious instructions inside user-controlled text.
Question 10
An AI trained on customer chats later repeats a private phone number. What should have been removed?
Why: Training or retrieval data should be cleaned so private information is not exposed.
Question 11
An AI answers confidently but invents the source. What is the failure called?
Why: The model generated plausible text that was not grounded in facts.
Question 12
What is the hidden unit most chatbots read instead of whole words?
Why: Models split text into token chunks before processing.
Question 13
Why can a GPU train AI faster than a CPU?
Why: GPUs run many similar calculations at the same time.
Question 14
What is a prompt injection trying to change?
Why: It attempts to override or manipulate the system behavior.
Question 15
A disease affects 1 in 1000. A model predicts nobody has it and scores 99.9% accurate. Why is it useless?
Why: With rare positives, always guessing 'no' scores well while catching none of the real cases.
Question 16
What makes a website load faster by saving repeated files?
Why: Cached files avoid downloading the same assets again.
Question 17
What is data called after it is labeled for training?
Why: Labels turn raw examples into usable model training material.
Question 18
What tech word means a fake test environment?
Why: A sandbox isolates risky actions from the real system.
Question 19
What does an API key act like?
Why: It authorizes code to use a service.
Question 20
What is the danger of copying secret keys into a public repo?
Why: Anyone can use exposed keys if they find them.
Question 21
A search engine finds similar ideas even when words differ. What representation makes that possible?
Why: Embeddings place semantically related content near each other in a numeric vector space.
Question 22
A trained model answers a new prompt without updating its weights. What phase is this?
Why: Inference uses learned parameters to generate or predict outputs for new inputs.
Question 23
Which transformer mechanism weighs relevant words against each other while processing a sequence?
Why: Attention assigns different weights to tokens based on their relevance to the current representation.
Question 24
Your model's accuracy falls month after month, but nobody changed the code or the model. What is happening?
Why: The world moved: live data no longer resembles the data the model was trained on.
Question 25
A base model is trained further on curated customer-service examples. What process is this?
Why: Fine-tuning adapts a pretrained model using a smaller task-specific dataset.
Question 26
A chatbot searches company documents before answering. Which architecture is it using?
Why: RAG retrieves relevant external information and supplies it to the model before generation.
Question 27
Before a language model reads text, what process splits it into smaller units?
Why: Tokenization converts text into token IDs the model can process.
Question 28
You fine-tune a model on a new task and it suddenly fails at a task it used to do well. What happened?
Why: Training hard on new data can overwrite the weights that encoded the earlier skill.
Question 29
Which sampling control restricts choices to a cumulative probability mass?
Why: Top-p sampling draws from the smallest token set whose cumulative probability reaches the chosen threshold.
Question 30
Which instruction layer defines an assistant's highest-level role and behavior?
Why: A system prompt sets high-priority behavior before ordinary user messages.
Question 31
You show a model three labeled examples before asking a similar task. What technique is this?
Why: Few-shot prompting demonstrates the desired pattern with a small number of examples.
Question 32
You ask a model to classify text without giving any examples. What technique is this?
Why: Zero-shot prompting asks for a task using instructions but no demonstrations.
Question 33
An answer is checked against an approved document instead of model memory. What improves?
Why: Grounding ties a model response to supplied evidence rather than unsupported recall.
Question 34
One model accepts text, images, and audio. What kind of model is it?
Why: A multimodal model can process or generate more than one type of media.
Question 35
Users wait five seconds after clicking Generate. Which product metric is hurting?
Why: Latency measures the delay between a request and its response.
Question 36
A model becomes smaller by storing weights with fewer bits. What technique was used?
Why: Quantization reduces numerical precision to lower memory and compute requirements.
Question 37
A small model learns to imitate a larger model's outputs. What is this called?
Why: Distillation trains a smaller student model using signals from a larger teacher model.
Question 38
What are the learned numerical values inside a neural network commonly called?
Why: Parameters are adjusted during training and determine how the model transforms inputs.
Question 39
Which optimization method repeatedly adjusts weights to reduce model error?
Why: Gradient descent updates parameters in a direction that lowers the loss function.
Question 40
One complete pass through a training dataset is called what?
Why: An epoch is one full traversal of the available training examples.
Question 41
A model scores 99% in testing and fails immediately in production. A column secretly contained the answer. What went wrong?
Why: A feature carried information unavailable at prediction time, so the test score measured nothing real.
Question 42
Which dataset should remain separate until final model evaluation?
Why: A held-out test set estimates performance on data not used for training or tuning.
Question 43
A model scores highly because test questions appeared in its training data. What happened?
Why: Contamination makes evaluation results misleading because the model may have seen the answers.
Question 44
A model memorizes training examples but performs poorly on new data. What is this?
Why: Overfitting occurs when a model learns training details that do not generalize.
Question 45
An AI hiring tool disadvantages one group because historical data reflects past inequality. What risk appeared?
Why: Models can reproduce unfair patterns present in data, labels, or deployment decisions.
Question 46
A medical AI flags a serious case, but a clinician makes the final decision. What safeguard is this?
Why: Human review keeps accountable experts involved in consequential decisions.
Question 47
A model's 90% confidence predictions are correct only half the time. What is poor?
Why: Calibration measures whether stated confidence levels match observed accuracy.
Question 48
You need repeatable outputs for a strict workflow. Which setting should usually be lowered?
Why: Lower temperature generally reduces sampling randomness and improves repeatability.
Question 49
An app must receive valid JSON from a model every time. What should constrain the output?
Why: Schema-constrained generation defines required fields, types, and structure.
Question 50
A research answer includes links that do not support its claims. What verification step failed?
Why: Citations must be opened and matched to the specific claims they supposedly support.
Question 51
A 500-page manual overwhelms retrieval. What preprocessing step divides it into searchable sections?
Why: Chunking splits long documents into smaller passages for indexing and retrieval.
Question 52
Search retrieves twenty passages, then a second model orders the best five. What is that step?
Why: A reranker scores retrieved candidates again to improve final relevance.
Question 53
A search finds 'automobile' when the query says 'car.' What kind of search is this?
Why: Semantic search uses meaning-based representations rather than exact keyword matches alone.
Question 54
Which measure commonly compares the direction of two embedding vectors?
Why: Cosine similarity measures the angle between vectors and is widely used for embedding comparison.
Question 55
A chatbot refuses to reveal private account data even when asked creatively. What control helped?
Why: Guardrails constrain behavior around safety, privacy, or business rules.
Question 56
A team deliberately attacks its own AI before launch. What practice is this?
Why: Red teaming probes a system for abuse paths, failures, and unsafe behavior.
Question 57
A prompt update quietly breaks ten working use cases. What test should catch this?
Why: Regression tests rerun known cases to detect behavior that changed unexpectedly.
Question 58
Half of users see one AI summary design and half see another. What experiment is this?
Why: An A/B test compares measured outcomes between two controlled variants.
Question 59
A cancer screener must avoid missing positive cases. Which metric deserves special attention?
Why: Recall measures how many actual positive cases the system successfully identifies.
Question 60
A spam filter blocks a legitimate invoice. What kind of error is that?
Why: A false positive incorrectly labels a negative case as positive.
Question 61
A fraud detector flags 100 accounts and only 3 are truly fraudulent. Which metric is poor?
Why: Precision asks how many flagged cases were correct; here 97 of 100 alerts were false alarms.
Question 62
A hidden instruction inside a webpage manipulates the AI reading it. What attack is this?
Why: Indirect injection embeds malicious instructions in external content an AI later processes.
Question 63
A user repeatedly rephrases requests to bypass an AI's safety policy. What are they attempting?
Why: A jailbreak is an attempt to circumvent a model's intended safety restrictions.
Question 64
Attackers insert malicious examples into training data to alter future behavior. What attack is this?
Why: Poisoning corrupts training or feedback data to influence a model.
Question 65
Someone queries a model to reconstruct details from its training records. What privacy attack is this?
Why: Model inversion attempts to infer sensitive attributes or representative inputs from model behavior.
Question 66
An attacker tests whether one person's record was used for training. What attack is this?
Why: Membership inference estimates whether a specific example belonged to a model's training data.
Question 67
Before sending support logs to an AI, names and phone numbers are removed. What step is this?
Why: Redaction removes or masks personally identifiable information before processing.
Question 68
An AI assistant can read invoices but cannot delete them. Which security principle is this?
Why: Least privilege grants only the permissions required for the task.
Question 69
Untrusted AI-generated code runs inside an isolated environment. What protects the host?
Why: A sandbox limits the code's access to files, processes, and networks.
Question 70
A repository blocks commits containing API keys. What automated control is running?
Why: Secret scanning detects credential patterns before or after they enter source control.
Question 71
A realistic video makes a public figure appear to say words they never spoke. What is it?
Why: A deepfake uses synthetic media techniques to imitate a person's appearance or voice.
Question 72
Which metadata can record where digital media came from and how it changed?
Why: Provenance records can document a media asset's origin and editing history.
Question 73
An urgent email uses AI-written language and a fake login page. What attack remains the core threat?
Why: AI can improve the message, but the attack still tricks users into revealing credentials.
Question 74
A caller perfectly imitates your manager and demands money. What should you verify through another channel?
Why: A separate trusted channel helps detect voice-cloning and impersonation scams.
Question 75
A stolen password alone cannot open an account because a second proof is required. What protected it?
Why: MFA requires an additional factor beyond the password.
Question 76
Data is encrypted while moving between an app and its AI provider. What protection is this?
Why: Transport encryption protects data as it travels across networks.
Question 77
Stored AI conversation logs are unreadable without a key. What protection is this?
Why: Encryption at rest protects stored files and databases from unauthorized reading.
Question 78
A company needs to know who asked an AI to access a customer record. What should it keep?
Why: Audit logs record actions, identities, and timestamps for later review.
Question 79
An AI vendor keeps prompts forever without a business need. Which policy should limit this?
Why: A retention policy defines how long data is stored and when it is deleted.
Question 80
A model runs entirely on a laptop without sending prompts away. What privacy advantage does this provide?
Why: Local processing can keep sensitive inputs on the user's device.
Question 81
An AI agent checks a calendar by calling an external service. What capability is it using?
Why: Tool use lets a model invoke software functions or external systems.
Question 82
An agent divides a travel request into flights, hotels, and scheduling steps. What capability is this?
Why: Planning decomposes a goal into ordered subtasks and decisions.
Question 83
An assistant remembers a user's preferred writing style next week. What feature stores that preference?
Why: Persistent memory carries selected information across separate sessions.
Question 84
An AI drafts a bank transfer, but a person must approve it before sending. What control is this?
Why: Approval gates prevent high-impact actions from executing without accountable review.
Question 85
A payment tool receives the same AI request twice but charges only once. What property prevented duplication?
Why: An idempotent operation can be repeated without creating additional effects.
Question 86
An API fails temporarily, so the agent waits longer after each retry. What strategy is this?
Why: Backoff spaces retries progressively to reduce overload and allow recovery.
Question 87
An AI service rejects requests after too many arrive per minute. What control triggered?
Why: Rate limiting caps request frequency to protect capacity and prevent abuse.
Question 88
A tool call hangs forever unless the agent stops waiting after thirty seconds. What limit is needed?
Why: A timeout ends work that exceeds an allowed duration.
Question 89
An agent edits the wrong record and restores the previous state. What recovery action occurred?
Why: A rollback reverses a change by returning to a known earlier state.
Question 90
A regulated workflow must reconstruct every AI action and tool result. What record is essential?
Why: An audit trail preserves the sequence of decisions, inputs, actions, and outcomes.
Question 91
A team tracks which prompt produced each release result. What practice supports this?
Why: Prompt versioning records changes so outputs can be reproduced and compared.
Question 92
Several AI steps run in a controlled order with data passed between them. What manages this?
Why: Orchestration coordinates steps, dependencies, state, and failures across a workflow.
Question 93
A model selects a typed function and supplies its arguments. What feature is this?
Why: Function calling lets a model request a defined operation using structured arguments.
Question 94
Before an agent sends an email, software checks the address and required fields. What step is this?
Why: Validation checks that generated data satisfies rules before an action runs.
Question 95
The primary AI provider fails, and the app switches to another model. What resilience pattern is this?
Why: A fallback provides an alternate path when the preferred model is unavailable.
Question 96
An agent stops before a tool call would exceed its spending limit. What control protected the budget?
Why: A cost cap prevents an automated workflow from spending beyond an approved threshold.
Question 97
An agent repeats the same two actions without progress. What safeguard should stop it?
Why: Loop detection identifies repeated states or actions and terminates stalled workflows.
Question 98
A research agent may search the web but cannot send emails. Which principle limits its tools?
Why: Restricting tools and permissions reduces the damage an agent can cause.
Question 99
An agent answers using the actual database result instead of guessing. What is the response grounded in?
Why: Using returned tool data grounds the answer in observed system state.
Question 100
A workflow moves through Draft, Review, Approved, and Sent. What structure controls these transitions?
Why: A state machine defines allowed states and transitions between them.
Question 101
An image begins as random noise and becomes a scene through repeated denoising. What model process is this?
Why: Diffusion image generation iteratively removes predicted noise to form an image.
Question 102
Two neural networks compete, with one generating images and one judging them. What architecture is this?
Why: A generative adversarial network trains a generator against a discriminator.
Question 103
An image prompt says what must not appear in the result. What is that instruction called?
Why: A negative prompt describes elements the generation process should avoid.
Question 104
AI increases a small image's resolution while reconstructing detail. What process is this?
Why: AI upscaling predicts higher-resolution pixels from a lower-resolution source.
Question 105
A meeting recording becomes searchable text. Which AI task performed the conversion?
Why: Automatic speech recognition converts spoken audio into text.
Question 106
Written narration becomes a natural-sounding voice. Which AI task created it?
Why: Text-to-speech synthesizes spoken audio from written text.
Question 107
Artificial examples are generated to supplement scarce real training records. What are these examples called?
Why: Synthetic data is artificially created to resemble useful properties of real data.
Question 108
A shopping app predicts products a user may like. What type of system is this?
Why: Recommender systems rank items using user behavior, item features, or both.
Question 109
An AI labels each email as urgent or normal. What machine-learning task is this?
Why: Classification assigns an input to one of a defined set of categories.
Question 110
An AI groups customers by behavior without predefined labels. What task is this?
Why: Clustering discovers groups based on similarity without requiring known class labels.
Question 111
A payment suddenly differs from a customer's normal pattern. What AI task can flag it?
Why: Anomaly detection identifies observations that deviate strongly from expected patterns.
Question 112
A model predicts next month's demand from historical sales. What task is this?
Why: Forecasting estimates future values from past patterns and relevant variables.
Question 113
An AI detector labels human writing as generated. Why should its verdict not be treated as proof?
Why: AI-text detectors can produce false positives and should not be the sole evidence.
Question 114
Sales rose after an AI tool launched. What must be tested before claiming the AI caused it?
Why: Timing alone shows correlation; experiments or stronger evidence are needed to establish causation.
Question 115
A manager accepts an AI recommendation without checking obvious errors. What bias is this?
Why: Automation bias is overreliance on automated output despite reasons to question it.
Question 116
A chatbot says 'I feel worried.' Why is treating that sentence as proof of emotion risky?
Why: Human-like language can cause people to attribute feelings or intentions that have not been established.
Question 117
Training a larger model uses more computation and electricity. What operational concern grows?
Why: More training compute generally requires more electrical energy and supporting infrastructure.
Question 118
A phone analyzes a photo without sending it to the cloud. Where is the AI running?
Why: Edge AI performs inference on or near the device producing the data.
Question 119
A large model will not fit on a graphics card. Which hardware resource is exhausted?
Why: Model weights, activations, and caches must fit within available GPU memory.
Question 120
A model answers with today's private inventory even though training ended last year. What likely supplied the update?
Why: External retrieval or tool access can provide current data not stored in the model's training weights.
These questions are free to use — in a classroom, a pub quiz, a video, or anywhere else. No attribution required. If you think one is ambiguous or an explanation is wrong, tell us and we will fix the page.
For how questions like these are constructed, see what makes a quiz question hard.