AI Ethics and Governance Frameworks

Capable models carry real risks: discrimination, privacy harm, misinformation, and safety failures. AI governance is the set of policies, processes, and documentation that keep AI systems lawful, fair, and accountable across their lifecycle.

The EU AI Act is a landmark regulation that classifies AI systems by risk and imposes obligations proportional to that risk. It uses a four-tier risk pyramid.

At the top, unacceptable risk systems are banned outright. Examples include government social scoring and manipulative systems that exploit vulnerabilities. These cannot be deployed regardless of safeguards.

High-risk systems (medical devices, hiring, credit scoring, critical infrastructure) are allowed but heavily regulated: they require risk management, high-quality data, logging, human oversight, and conformity assessment before market entry.

The lower tiers are lighter touch:

• Limited risk (chatbots, deepfakes) requires transparency, e.g. telling users they interact with AI
• Minimal risk (spam filters, game AI) has essentially no obligations

From top to bottom the EU AI Act pyramid is: unacceptable (banned), high (strict requirements), limited (transparency duties), and minimal (free use). Most everyday AI falls in the bottom two tiers.

A Model Card is a short, standardized document published alongside a model. It communicates what the model is for and where it should not be used, so downstream teams make informed decisions. It is the AI equivalent of a nutrition label.

The intended use section states the primary purpose, target users, and explicitly out-of-scope uses. This prevents misuse, like applying a sentiment model trained on English tweets to legal contracts.

The performance section reports metrics, ideally broken down by subgroup, and the evaluation data used. Disaggregated metrics expose whether the model works well for everyone or only the majority group.

The caveats and limitations section lists known failure modes, biases, and conditions where the model degrades, such as low-resource languages or out-of-distribution inputs. Honest caveats build trust and prevent harm.

Teams operationalize governance with a responsible AI checklist reviewed before launch:

• Is training data representative and consented?
• Have we measured fairness across groups?
• Is there human oversight for high-stakes decisions?
• Are explanations available to affected users?
• Is a Model Card published and a monitoring plan in place?

Test your governance knowledge.

You learned AI ethics and governance:

• The EU AI Act risk pyramid: unacceptable (banned), high (strict), limited (transparency), minimal (free)
• A Model Card documents intended use, performance (disaggregated), and caveats
• A responsible AI checklist turns principles into pre-launch gates