Ethics & Security Academy
Navigate the risks of AI: Hallucinations, Bias, and Security Vulnerabilities.
Hallucinations
LLMs can confidently state falsehoods as facts. They are designed to be plausible, not truthful.
Tip: Ask the model to cite sources or explicitly say 'I don't know' if it lacks information.
Bias & Fairness
Models reflect the stereotypes present in their vast training data, including gender, racial, and cultural biases.
Tip: Explicitly prompt for neutrality or distinct perspectives (e.g., 'Analyze this from multiple viewpoints').
Prompt Injection
Attacks where users trick the model into overriding its system instructions to perform malicious actions.
Tip: Use delimiters (like ###) to separate instructions from user input and validate inputs.
Data Privacy (PII)
Never input Personally Identifiable Information (PII) like confidential code, customer lists, or passwords into public LLMs. Your data may be used to train future models.
Tip: Anonymize data before prompting. Replace real names/companies with 'Company A' or 'Customer X'.
Copyright & IP
The legal status of AI-generated content is complex. In many jurisdictions, AI art and text cannot be copyrighted. Be careful when using AI for commercial assets.
Tip: Treat AI output as a draft or inspiration, not finished intellectual property. Human modification is key.
Over-reliance & Model Collapse
Blindly trusting AI erodes critical thinking. 'Model Collapse' suggests that future AIs trained on AI-generated data may become dumber and more homogenous.
Tip: Use AI as a co-pilot, not an autopilot. Always verify, edit, and understand the output.
Agent & tool risks
Agents that can browse, send email, or run code amplify mistakes. A confused plan can delete data or leak secrets at machine speed.
Tip: Least privilege tools, human approval for irreversible actions, and clear stop rules.