Governance Flashcards
(37 cards)
What is governance in machine learning?
A framework for ensuring transparency, accountability, and traceability across the ML lifecycle.
Why is governance important in ML systems?
It ensures decisions and model behaviors can be tracked, justified, and audited.
What are examples of artifacts produced during the ML lifecycle?
Datasets, model weights, evaluation metrics, deployment configs, logs.
What are data cards used for?
To describe a dataset’s origin, structure, use cases, and ethical limitations.
What are model cards used for?
To document a model’s purpose, performance, risks, and intended usage.
What does ML governance require for artifacts?
That they are versioned, documented, and attributable to responsible parties.
Why is feature traceability important in ML?
To know how data and models influence decisions and detect where failures occur.
What is the trolley problem used to illustrate in ML ethics?
Moral dilemmas where algorithms must make tradeoffs between harmful outcomes.
What kind of ethical risk does weather prediction pose?
Biased data coverage can lead to poor outcomes for underrepresented communities.
Who might be responsible for a flawed prediction model?
The developer, data provider, or deploying organisation.
What are ethical concerns raised by ChatGPT deployment?
Fair wages, content moderation harm, energy usage, data ownership, inequality.
What ethical question does ChatGPT raise?
Is the tool beneficial to society, and is it being developed responsibly?
What is the UK’s approach to AI regulation?
Applies existing laws like the Data Protection Act and Equality Act to AI systems.
What laws apply to AI in the UK?
Data Protection Act, Equality Act, and Intellectual Property law.
What is the limitation of the UK’s AI legal framework?
It lacks AI-specific rules on transparency, explainability, and accountability.
What is the US AI Bill of Rights?
A draft set of principles for ethical and fair use of AI technologies.
What are key principles in the US AI Bill of Rights?
Safety, bias protection, privacy, transparency, and human alternatives.
Is the US AI Bill of Rights legally binding?
No, it is currently a policy proposal without legal enforcement.
What is the EU AI Act?
The first formal legislation regulating AI systems based on risk levels.
What types of AI are banned under the EU AI Act?
Cognitive manipulation, social scoring, real-time biometric surveillance.
What AI systems are considered high-risk under the EU AI Act?
Systems used in healthcare, law enforcement, education, or employment.
What are the rules for high-risk AI under the EU AI Act?
Mandatory documentation, audits, registration, and human oversight.
What do general-purpose AI systems require under the EU AI Act?
They must meet transparency and disclosure requirements.
What must models trained on copyrighted data do under the EU AI Act?
Disclose the sources of training data.
