安全与可信 AI
What are hallucinations in LLMs, and how do you mitigate them?
原题What is prompt injection, and what are the different types (direct, indirect)?
原题How do you implement input and output guardrails for AI systems?
原题What is AI alignment, and why is it important?
原题How do you detect and mitigate bias in AI systems?
原题What are the key data privacy considerations (GDPR, CCPA) when building AI applications?
原题How do you handle PII in LLM inputs and outputs?
原题What is explainability in AI, and why does it matter?
原题What is the difference between interpretability and explainability?
原题What is mechanistic interpretability, and why do AI labs invest in it?
原题How do you build trust with users in AI-powered applications?
原题What are adversarial attacks on AI systems, and how do you defend against them?
原题What is data poisoning, and how can it affect AI models?
原题How do you implement content safety filters for AI-generated content?
原题What is responsible AI, and what frameworks exist for implementing it?
原题How do you handle copyright and intellectual property concerns with AI-generated content?
原题What is the EU AI Act, and how does it affect AI engineering?
原题How do you implement audit trails and logging for AI decisions?
原题What is model card documentation, and why is it important?
原题How do you handle misuse and abuse of AI systems in production?
原题What is differential privacy, and how can it be applied during model training?
原题How would you design an AI incident response plan?
原题What is the NIST AI Risk Management Framework (AI RMF)?
原题Your healthcare chatbot gives medical diagnoses it should not make. How do you add safety guardrails?
原题Your AI system is reproducing copyrighted material verbatim. How do you prevent this?
原题Your resume screening AI rejects more female candidates for engineering roles. How do you fix gender bias?
原题Your AI model passes bias checks by gender and race separately, but fails for intersectional groups. How do you handle it?
原题Your AI denied a loan, and the customer demands a GDPR explanation. How do you provide one?
原题A user invokes the right to be forgotten, but their data is in your model weights. How do you comply?
原题The EU AI Act may classify your AI system as high-risk. How do you comply?
原题Your differentially private model lost significant accuracy. How do you balance privacy and utility?
原题One malicious participant is poisoning your federated learning model. How do you defend against it?
原题Your AI hiring model uses proxy features for protected attributes. How do you eliminate proxy discrimination?
原题Your predictive model creates a feedback loop of biased outcomes. How do you break it?
原题Your AI generates fake news images. How do you implement watermarking for AI-generated content?
原题Your AI denies a service, and the user has no way to challenge it. How do you design an appeals process?
原题An auditor asks why your AI rejected a request 6 months ago, and you have no logs. How do you build audit trails?
原题You removed PII, but users were re-identified from anonymized data. How do you prevent re-identification?
原题A pre-trained model from an open-source repo may contain a hidden backdoor. How do you detect it?
原题Your LLM's training data was deliberately poisoned by an adversary. How do you respond?
原题Your AI mental health chatbot gave harmful advice to a user in crisis. How do you mitigate harm?
原题Your AI system caused incorrect critical decisions. How do you run a blameless post-mortem?
原题Radiologists agree with AI 98% of the time, even when it is wrong. How do you prevent human over-reliance on AI?
原题Your content moderation flags normal cultural expressions as offensive in other markets. How do you adapt cross-culturally?
原题Your AI training produces massive carbon emissions. How do you reduce environmental impact?
原题
暂无题目