I85PREP
⌘ K
GenAIBIZ study materials
GenAIBIZ study resources

GenAIBIZ practice questions

One question at a time. Build your GenAIBIZ understanding.

First 50 of 190 questions. Each one lists the choices, the correct answer, and an explanation. The checker at the top of this page saves progress in this browser when it is available.

Continue in the practice tool above to load more questions, 50 at a time.

Question 1AI foundations, data and deployed-system lifecycle

Generative AI (GenAI) primarily differs from traditional predictive AI in that it:

  1. Is faster to train
  2. Creates new content (text, images, audio, code) rather than only predicting labels or values Correct answer.
  3. Requires less data
  4. Only runs on CPUs

Explanation

Traditional AI predicts (classify, score, forecast). Generative AI produces novel artifacts: text, images, audio, video, code, and structured data. Both are valuable; GenAI excels at creation tasks like drafting, summarization, ideation, and personalization.

Question 2AI foundations, data and deployed-system lifecycle

A foundation model is best described as:

  1. A small, task-specific model trained from scratch for one customer
  2. A large model pre-trained on broad data that can be adapted to many downstream tasks Correct answer.
  3. A rule-based expert system
  4. A specialized image classifier

Explanation

Foundation models (Stanford 2021 term) are large, generalist models pre-trained on broad data. LLMs like GPT-4, Claude, and Gemini are foundation models. They are adapted to tasks via prompting, fine-tuning, or RAG.

Question 3AI foundations, data and deployed-system lifecycle

Multimodal models can:

  1. Only process text
  2. Accept and/or produce multiple modalities such as text, images, audio, and video Correct answer.
  3. Only run in the cloud
  4. Only translate languages

Explanation

Multimodal models (GPT-4o, Claude 3.5/4, Gemini, Llama 4) can read images, listen to audio, watch video, and respond in multiple modalities. This unlocks use cases like visual QA, document understanding, and voice agents.

Question 4AI foundations, data and deployed-system lifecycle

An autonomous AI agent typically:

  1. Only answers single questions and stops
  2. Uses an LLM as a controller that plans, calls tools, and executes multi-step actions toward a goal Correct answer.
  3. Cannot use external APIs
  4. Replaces all human oversight

Explanation

The keyed option describes a common LLM-powered agent pattern in this generative-AI context. Not every AI agent uses an LLM, and planning or tool access does not establish authority, correctness or safety. Constrain actual permissions and verify outcomes rather than equate autonomous behavior with unrestricted reliable action.

Question 5AI foundations, data and deployed-system lifecycle

Retrieval-Augmented Generation (RAG) primarily reduces hallucinations by:

  1. Increasing model temperature
  2. Grounding the model on retrieved relevant documents and asking it to cite or stay within them Correct answer.
  3. Removing all training data
  4. Disabling the model

Explanation

Retrieval can supply relevant source material that helps grounding, but retrieved content can be stale, unauthorized, irrelevant or malicious, and synthesis can still be unsupported. Verify source access, evidence and evaluation. A citation-shaped answer is not proof that the cited material supports the claim.

Question 6AI foundations, data and deployed-system lifecycle

Fine-tuning an LLM is most useful when:

  1. You only need the model to answer one question
  2. You have proprietary data, distinct style, or specialized tasks where prompt engineering and RAG hit limits Correct answer.
  3. You want lower compute costs without changing behavior
  4. You want to delete the model

Explanation

Fine-tuning adapts a base model to your data and style. Use when prompts/RAG are insufficient, you need consistent format/voice, or you must deploy small specialized models. Costs more upfront and requires curated training data and evaluation.

Question 7AI foundations, data and deployed-system lifecycle

When GenAI may not be the right tool, you should:

  1. Force it to be used regardless
  2. Choose simpler analytics, traditional ML, or human workflows when they better fit accuracy, cost, or risk constraints Correct answer.
  3. Always default to GenAI
  4. Hide the failure

Explanation

Not every problem needs GenAI. For numeric forecasting, classical ML/statistics may be cheaper and more accurate. For deterministic logic, code is best. Match the tool to the problem; do not over-rotate to GenAI for everything.

Question 8AI foundations, data and deployed-system lifecycle

A token in LLM context is best described as:

  1. A single character
  2. A subword unit produced by the tokenizer; pricing and context limits are token-based Correct answer.
  3. A unit of memory
  4. A type of cryptocurrency

Explanation

A token is a unit defined by the actual tokenizer and model interface; it can represent a subword, character or other supported unit rather than always one word or one universal fragment. Limits and metering follow actual service rules. No model request or paid generation occurs in this practice.

Question 9AI foundations, data and deployed-system lifecycle

An LLM's context window refers to:

  1. The size of the user's monitor
  2. The maximum number of tokens (input + output) the model can process in a single request Correct answer.
  3. The number of users connected
  4. The model's training set size

Explanation

A context window constrains the supported amount of model context under actual input/output and other accounting rules. Tokens are not characters or a guarantee of correctly using every supplied detail. Model-specific limits, modality handling and output reservation must be checked rather than inferred from a generic number.

Question 10AI foundations, data and deployed-system lifecycle

Embedding-based search in RAG works by:

  1. Comparing exact keyword overlap only
  2. Encoding queries and documents as vectors and retrieving by cosine similarity / nearest neighbors Correct answer.
  3. Random sampling
  4. Running a relational join

Explanation

Embedding models map text to dense vectors capturing semantic meaning. Vector databases (Pinecone, Weaviate, Chroma, pgvector, Milvus) index embeddings for fast nearest-neighbor search. Hybrid (vector + BM25) usually outperforms either alone.

Question 11AI threat boundaries, agents and technical safeguards

An LLM hallucination is:

  1. A bug in the model code
  2. A confidently stated but factually incorrect or fabricated output Correct answer.
  3. A type of model compression
  4. An adversarial attack technique

Explanation

Hallucinations are plausible-sounding but false outputs. Mitigations include grounding via retrieval (RAG), citing sources, lower temperature, prompt constraints, human review, and using domain-tuned models. Critical in regulated and high-stakes domains.

Question 12AI threat boundaries, agents and technical safeguards

Bias in GenAI outputs originates primarily from:

  1. Hardware
  2. Training data, alignment choices, and prompt phrasing Correct answer.
  3. User clicks
  4. Browser settings

Explanation

Models learn statistical patterns from data - biased data yields biased outputs. Alignment (RLHF) and post-training shape behavior further. Mitigations include diverse data, debiasing techniques, evaluation across demographic slices, and human review.

Question 13AI threat boundaries, agents and technical safeguards

Prompt injection is an attack where:

  1. An attacker steals the model weights
  2. An attacker embeds malicious instructions into content the model will read, hijacking its behavior Correct answer.
  3. Hardware is overheated
  4. Latency is increased

Explanation

Prompt injection attempts to redirect an LLM application through malicious instructions in input or external content. It is not an automatic grant of higher authority. Actual consequences depend on application behavior, tool permissions, data access and enforced boundaries; no attack or tool call is performed here.

Question 14AI threat boundaries, agents and technical safeguards

Copyright concerns with GenAI typically arise around:

  1. Whether the model is open source
  2. Training data sourcing, output similarity to copyrighted works, and ownership of generated content Correct answer.
  3. GPU vendors only
  4. DNS configuration

Explanation

Active legal areas: training data scraping (NYT v OpenAI, Getty v Stability), substantial similarity in outputs, and IP ownership of AI-generated work (US Copyright Office holds purely AI-generated work uncopyrightable; meaningful human authorship matters).

Question 15AI threat boundaries, agents and technical safeguards

Deepfakes raise concerns primarily because they enable:

  1. Slower internet
  2. Realistic synthetic audio/video impersonation that can deceive viewers and enable fraud or disinformation Correct answer.
  3. Better SEO
  4. Higher CTR

Explanation

Deepfakes can impersonate executives for wire fraud, fabricate political statements, and produce non-consensual imagery. Mitigations include detection tools, content authenticity standards (C2PA), watermarking, and policy/legal responses.

Question 16AI threat boundaries, agents and technical safeguards

An employee using an unapproved consumer GenAI tool with company data is engaging in:

  1. Best practice
  2. Shadow IT, posing security, IP, and compliance risks Correct answer.
  3. An audited workflow
  4. Authorized use

Explanation

Shadow AI risks data leakage, IP exposure, and compliance violations. Mitigations: clear acceptable use policies, sanctioned enterprise tools, training, and DLP/CASB monitoring. Make compliant tools easier than non-compliant ones.

Question 17AI threat boundaries, agents and technical safeguards

A frequent misuse of LLMs that businesses should avoid is:

  1. Asking for help drafting an email
  2. Treating LLM outputs as authoritative for high-stakes legal, medical, or financial decisions without expert review Correct answer.
  3. Summarizing meeting notes
  4. Generating brainstorming lists

Explanation

LLMs are powerful drafting and ideation tools but are not licensed advisors. High-stakes decisions require qualified human experts. Document this distinction in policies and product UIs to set user expectations.

Question 18Generative AI business use cases, prompting and responsible adoption

Which is NOT a typical business use case for GenAI?

  1. Drafting marketing copy and personalized emails
  2. Generating literal physical products without manufacturing Correct answer.
  3. Summarizing long documents
  4. Helping developers write and review code

Explanation

GenAI creates digital artifacts (text, images, audio, code, designs). It cannot manufacture physical products, though it can speed up design, engineering documentation, and supply-chain decisions that lead to manufacturing.

Question 19Generative AI business use cases, prompting and responsible adoption

Customer-service chatbots powered by GenAI typically improve:

  1. Only marketing budgets
  2. First-response time, deflection of routine queries, and 24/7 availability - while complex cases escalate to humans Correct answer.
  3. Only fax volume
  4. Only legal billable hours

Explanation

Faster initial responses, routine-query handling and extended availability are potential chatbot benefits, not guaranteed outcomes for every deployment. Measure end-to-end quality, escalations, rework, cost and user impact under representative conditions. A quick answer that is wrong or blocks an appropriate human escalation can worsen the service.

Question 20Generative AI business use cases, prompting and responsible adoption

GitHub Copilot, Cursor, and similar coding assistants primarily provide:

  1. Only spell-check
  2. Inline code suggestions, generation, refactoring, and AI-assisted code review Correct answer.
  3. Build pipelines only
  4. Hardware for engineers

Explanation

AI coding assistants accelerate developers by suggesting code, completing functions, generating tests, and explaining errors. Reported productivity gains range 10-55%. Quality, security, and IP review remain essential.

Question 21Generative AI business use cases, prompting and responsible adoption

Marketing teams often use GenAI to:

  1. Print physical brochures by themselves
  2. Generate ad copy variants, personalized emails, SEO content, and creative images at scale Correct answer.
  3. Replace the CEO
  4. Manage company finances

Explanation

Marketing GenAI use cases: ad copy variants for A/B testing, personalized email and landing pages, SEO/blog drafting, social posts, image and video creative, audience research summarization. Brand voice and review workflows remain critical.

Question 22Generative AI business use cases, prompting and responsible adoption

Sales teams commonly apply GenAI for:

  1. Manufacturing parts
  2. Personalized outreach, lead enrichment, meeting summaries, proposal drafting, and call coaching Correct answer.
  3. Auditing financial statements
  4. Building data centers

Explanation

Sales GenAI: research and account briefs, personalized cold outreach, proposal/RFP drafting, meeting summaries with action items (Gong, Fireflies), call coaching, and forecast assistance. Always review for accuracy before sending.

Question 23Generative AI business use cases, prompting and responsible adoption

HR functions enhanced by GenAI include:

  1. Resume screening, interview question drafting, onboarding content, and L&D personalization Correct answer.
  2. Replacing labor laws
  3. Providing legal advice without counsel
  4. Issuing visas

Explanation

HR GenAI use cases require strong fairness controls because hiring is high-risk under the EU AI Act and US laws (NYC AEDT, EEOC guidance). Audit for bias, document processes, and keep humans in final decisions.

Question 24Generative AI business use cases, prompting and responsible adoption

Finance teams often use GenAI for:

  1. Manufacturing money
  2. Variance analysis, financial report summarization, anomaly detection, and contract review Correct answer.
  3. Setting interest rates as a substitute for central banks
  4. Replacing all auditors immediately

Explanation

Finance GenAI: summarize 10-Ks and earnings calls, draft variance commentary, scan contracts for risks, support FP&A drafting. Strict review and SOX-aware controls are mandatory; AI must not autonomously approve transactions.

Question 25Generative AI business use cases, prompting and responsible adoption

Legal teams typically use GenAI for:

  1. Issuing court rulings
  2. Contract review, clause comparison, due diligence summarization, and legal research drafts Correct answer.
  3. Practicing law without humans
  4. Voting in elections

Explanation

LegalTech GenAI: contract review (Spellbook, Harvey, Lexis+ AI), clause libraries, due diligence summaries, document review, and research drafts. Hallucinations have caused sanctions (Mata v Avianca) - attorneys must verify citations.

Question 26Generative AI business use cases, prompting and responsible adoption

Operations and supply chain GenAI examples include:

  1. Demand forecasting summaries, supplier risk monitoring, inventory optimization narratives, and SOP drafting Correct answer.
  2. Building airplanes by themselves
  3. Replacing trucks
  4. Managing warehouse temperature unilaterally

Explanation

Operations GenAI: synthesizing demand signals, summarizing supplier risk and disruption news, generating SOPs, drafting maintenance plans, and powering operator copilots. Combined with traditional analytics for KPIs.

Question 27Generative AI business use cases, prompting and responsible adoption

Personalized customer experiences with GenAI typically need:

  1. No data at all
  2. First-party customer data, proper consent, and segmentation aligned with privacy regulations Correct answer.
  3. Only public web scraping
  4. Anonymous random text

Explanation

Personalization needs reliable first-party data with consent and clear use limits. Track preferences, history, and intent; respect opt-outs. Avoid creepy surveillance and ensure messaging aligns with brand and policy.

Question 28Generative AI business use cases, prompting and responsible adoption

A document summarization use case typically benefits from:

  1. Removing the source documents
  2. Specifying summary length, audience, and key sections; verifying critical facts against source Correct answer.
  3. Random temperature
  4. Removing the title

Explanation

Effective summarization prompts include length (e.g., 5 bullets, 200 words), audience (executive, technical), and target sections. Pair with citation-required prompting and post-hoc fact checks for high-stakes content.

Question 29Generative AI business use cases, prompting and responsible adoption

Translating product content with GenAI is typically:

  1. Forbidden by all laws
  2. Effective for many language pairs but should include native-speaker review for tone, idiom, and regulatory wording Correct answer.
  3. Never useful
  4. Only allowed in one country

Explanation

Modern LLMs translate well, especially for well-resourced language pairs. Use translation memory, glossaries, and native review for regulated copy (legal, medical) and culturally sensitive marketing.

Question 30Generative AI business use cases, prompting and responsible adoption

A knowledge base assistant for employees grounded in internal docs requires:

  1. Public web access only
  2. A retrieval index of internal documents, access controls preserving permissions, citations, and evaluation Correct answer.
  3. No access controls
  4. Only voice input

Explanation

Enterprise RAG inherits source permissions (confidentiality, role-based access). Use vector + keyword retrieval, citations, freshness checks, evaluation on representative questions, and feedback loops for relevance tuning.

Question 31Generative AI business use cases, prompting and responsible adoption

A meeting summary tool's accuracy can be improved by:

  1. Disabling transcription
  2. High-quality transcription, speaker diarization, providing context (agenda, attendees), and structured output schemas Correct answer.
  3. Random output
  4. Skipping the audio

Explanation

Good meeting summaries depend on accurate ASR, speaker labels, agenda context, and structured outputs (decisions, action items with owners, due dates). Tools include Otter, Fireflies, Microsoft Teams Recap, Google Meet Take Notes for Me, and Zoom AI Companion.

Question 32Generative AI business use cases, prompting and responsible adoption

Predictive maintenance combined with GenAI typically uses:

  1. Only random guesses
  2. Sensor analytics + ML for failure prediction, with GenAI generating natural-language work orders and operator copilots Correct answer.
  3. Manual paperwork only
  4. Voice commands only

Explanation

Industrial AI combines time-series ML for anomaly detection with GenAI for narrative work orders, technician copilots, and SOP generation. Tightly integrate with CMMS/EAM systems for action and audit.

Question 33Generative AI business use cases, prompting and responsible adoption

Healthcare clinical documentation tools ("AI scribes") typically:

  1. Replace doctors entirely
  2. Listen to encounters and draft visit notes for clinician review and approval, reducing documentation burden Correct answer.
  3. Are illegal everywhere
  4. Diagnose without input

Explanation

The described assistant can draft documentation for qualified review under actual authorized clinical and privacy arrangements. Reduced burden is a potential measured benefit, not a guarantee, and a draft must not replace clinician judgment or required verification. This business example handles no patient data and makes no clinical decision.

Question 34Generative AI business use cases, prompting and responsible adoption

Generative AI for product design typically:

  1. Replaces all designers immediately
  2. Accelerates ideation, variant generation, and asset production while designers curate, refine, and own the brand Correct answer.
  3. Cannot produce images
  4. Only works in 3D

Explanation

GenAI augments designers with rapid ideation (Figma AI, Adobe Firefly), variant generation, and asset production. Human designers handle critical thinking, brand consistency, and final polish. Speed up the front of the funnel; preserve craft at the end.

Question 35Generative AI business use cases, prompting and responsible adoption

Effective prompt engineering for business users typically includes:

  1. Vague instructions and no context
  2. Clear instructions, role/persona, examples, output format, and tone guidance Correct answer.
  3. Only one-word prompts
  4. Refusing to provide context

Explanation

Strong prompts specify task, role (act as a tax accountant), few-shot examples, desired output format (JSON, markdown, table), tone (formal, casual), constraints, and audience. Iterate to refine - prompt engineering is empirical.

Question 36Generative AI business use cases, prompting and responsible adoption

Few-shot prompting means:

  1. Providing zero examples
  2. Providing a small number of input/output examples in the prompt to guide the model's output style and format Correct answer.
  3. Training the model from scratch
  4. Restricting model output to a single word

Explanation

Few-shot prompting includes 1-5 examples in the prompt. Zero-shot uses no examples. Few-shot raises consistency for niche tasks without fine-tuning. Modern instruction-tuned models often do well zero-shot; few-shot still helps for structured outputs.

Question 37Generative AI business use cases, prompting and responsible adoption

Chain-of-thought prompting encourages a model to:

  1. Output a single word
  2. Show its reasoning step by step before giving a final answer, often improving complex reasoning Correct answer.
  3. Skip computation
  4. Answer only in numbers

Explanation

The named prompting technique requests intermediate explanatory steps in an answer. Visible steps can still be incorrect, incomplete or unfaithful to how a system produced the result, and benefit depends on model and task. Validate the final result and relevant evidence rather than treat a step-by-step narrative as proof of sound reasoning.

Question 38Generative AI business use cases, prompting and responsible adoption

When you need a structured response for downstream automation, you should:

  1. Ask for free-form prose
  2. Specify the output format (JSON schema, markdown table) and use structured-output features when available Correct answer.
  3. Use voice input only
  4. Disable temperature

Explanation

Specifying schemas (JSON schema, function/tool calling, structured outputs) ensures parseable responses for automation. Major providers offer guaranteed schema compliance modes (e.g., OpenAI Structured Outputs, Anthropic tool use).

Question 39Generative AI business use cases, prompting and responsible adoption

A persona or role in a prompt ("Act as a senior tax accountant…") typically:

  1. Has no measurable effect
  2. Sets vocabulary, tone, and assumed expertise, often improving relevance for specialized tasks Correct answer.
  3. Forces the model to refuse the task
  4. Always reduces accuracy

Explanation

A role prompt can guide vocabulary, tone and task framing. It does not supply genuine professional qualifications, verified facts, legal authority or permission to make consequential decisions. In a specialized use case, validate critical content against suitable evidence and retain the appropriate qualified accountable reviewer.

Question 40Generative AI business use cases, prompting and responsible adoption

Lower temperature settings in an LLM tend to produce:

  1. More creative, varied outputs
  2. More deterministic, consistent outputs Correct answer.
  3. Faster training
  4. Bigger models

Explanation

Lower temperature can reduce sampling variability under a model’s supported implementation, but it does not guarantee deterministic or truthful results across runs, platforms or model changes. The actual decoding and serving settings matter. Evaluate output quality and business consequences rather than treat one parameter as a reliability guarantee.

Question 41Generative AI business use cases, prompting and responsible adoption

When asking an LLM to write an email, providing the recipient's role, the desired tone, and the specific call to action will typically:

  1. Have no impact
  2. Improve relevance and reduce edit time Correct answer.
  3. Cause hallucinations
  4. Triple cost

Explanation

Specifying audience, tone, length, structure, and CTA dramatically improves first-draft quality. Combine with a brand voice example (few-shot) and constraints. Edit-to-publish ratios are a useful internal metric.

Question 42AI risk, accountability and responsible use

Responsible AI principles commonly include:

  1. Only speed
  2. Transparency, accountability, fairness, privacy, security, and human oversight Correct answer.
  3. Maximum profit only
  4. Avoiding any documentation

Explanation

Responsible AI frameworks (Microsoft, Google, OECD, NIST AI RMF, IEEE) converge on transparency, accountability, fairness, privacy, safety, security, reliability, and human oversight. Operationalize via governance, model cards, and review boards.

Question 43AI risk, accountability and responsible use

Human-in-the-loop (HITL) means:

  1. Removing all humans from the workflow
  2. Keeping a human in the decision or review chain to validate, override, or guide AI outputs Correct answer.
  3. Letting users train the model
  4. Hiring more developers only

Explanation

HITL keeps humans in critical decisions (medical diagnoses, hiring, lending, content moderation). Benefits: accountability, error correction, and trust. Tradeoffs: latency and cost. Required by many high-risk regulations.

Question 44AI risk, accountability and responsible use

Disclosing AI-generated content to users ("This was drafted with AI assistance…") supports which responsible-AI principle?

  1. Speed
  2. Transparency Correct answer.
  3. Compression
  4. Latency

Explanation

Transparency obliges informing users when they are interacting with AI or consuming AI-generated content. The EU AI Act requires this for chatbots and synthetic content. Many platforms now use content provenance labels.

Question 45AI risk, accountability and responsible use

Anonymizing PII before sending data to a public LLM is a practice of:

  1. Marketing
  2. Privacy-by-design and data minimization Correct answer.
  3. Branding
  4. SEO

Explanation

Removing unnecessary identifying data and validating an appropriate privacy-preserving transformation can support minimization and privacy by design. A superficially anonymized record may still permit re-identification through linkage or context. Actual authority, retention, vendor handling and applicable requirements must still be reviewed before disclosure.

Question 46AI risk, accountability and responsible use

An AI red team is responsible for:

  1. Marketing AI products
  2. Adversarially probing AI systems for misuse, jailbreaks, prompt injection, bias, and unsafe outputs Correct answer.
  3. Hardware procurement
  4. Office facilities

Explanation

Red teaming AI uncovers failure modes before deployment: jailbreaks, prompt injection, bias, harmful outputs, and tool misuse. NIST and the EU AI Act expect adversarial evaluation for high-risk systems.

Question 47AI risk, accountability and responsible use

C2PA (Coalition for Content Provenance and Authenticity) helps with:

  1. Cryptocurrency mining
  2. Standardized content credentials (provenance metadata) to indicate origin, edits, and AI involvement in media Correct answer.
  3. Hardware accelerators
  4. Operating systems

Explanation

C2PA content credentials can provide signed provenance assertions and support checking covered origin/edit history. They do not certify that the depicted event or statement is true, that every prior step was captured or that a use is lawful. Absence of credentials does not alone prove a fake.

Question 48AI risk, accountability and responsible use

Privacy by design with GenAI means:

  1. Adding privacy notices after launch
  2. Embedding privacy controls (data minimization, consent, residency, retention) from the earliest design phase Correct answer.
  3. Skipping consent
  4. Logging all PII forever

Explanation

Privacy by design integrates relevant data minimization, access, retention, vendor and user-impact considerations from the outset. Consent is not the sole possible legal basis in every jurisdiction or use. A generic checklist does not establish lawful processing; qualified review must follow the actual scope and requirements.

Question 49AI risk, accountability and responsible use

Model cards (Mitchell et al. 2019) document:

  1. Server hardware
  2. Intended use, performance across subgroups, training data, limitations, and ethical considerations Correct answer.
  3. Marketing slogans
  4. Office addresses

Explanation

Model cards standardize transparency: who should/should not use, performance disaggregated by demographics, training data summary, known limitations, and ethical considerations. Adopted by Hugging Face, Google, Microsoft, and others.

Question 50AI risk, accountability and responsible use

When a GenAI tool produces a clearly biased recommendation, the appropriate first step is to:

  1. Ship it as-is
  2. Pause/limit the workflow, investigate the cause, retrain or change prompts, and document the incident and remediation Correct answer.
  3. Blame the user
  4. Deny the issue publicly

Explanation

Incident response for AI: contain (limit blast radius), investigate (data, prompt, model), remediate (prompt fix, data, model swap), document, and communicate. Build a post-incident review culture similar to SRE practices.