AI & Maatschappij

Benefits and harms are not necessarily shared equally. Evaluation needs to consider access, power, accountability, and the choices made around the technology, as well as technical performance.

Key takeaways

• Include affected non-users in the analysis.
• Assess access and the ability to challenge outcomes.
• Measure real impacts and revise the system.

Identify who is affected, including people who do not directly use the system. An automated service can change access to assistance, the information people see, or the work expected from staff. Those effects may be missed by measuring only the operator’s efficiency. Consider distribution and participation. Language coverage, disability access, connectivity, cost, and the ability to challenge an outcome can determine who benefits. Consult affected communities when defining the problem and evaluating the proposed solution. Information environments require particular care. Generated content can increase access to explanations while also amplifying misleading material. Provenance, corrections, and transparent evidence help readers evaluate claims without requiring blind trust in a platform. Treat governance as a continuing process. Track actual outcomes, investigate complaints, and revise systems when uses or impacts change. Avoid assuming that broad social outcomes follow automatically from one model improvement. Institutions and policy choices shape how capabilities are deployed and who bears their consequences.

1. Imagine a service moving important assistance to an AI-only chat interface.
2. Test the experience for people with unreliable connectivity, uncommon questions, and assistive technology.
3. Measure whether a staffed or alternative route remains available and whether unresolved cases receive effective follow-up.

The constructed example expands evaluation from average interaction speed to access and accountability.