Right Path for AI: Karen Hao and Paola Ricaurte on Ethics

Karen Hao and Paola Ricaurte spoke at an MIT panel on March 20, 2026, laying out competing visions for the future of artificial intelligence and concrete steps to steer the technology toward public benefit. The session focused on policy and design choices for large language models, with an emphasis on governance, accountability and meeting real human needs.

Key arguments

Hao framed the debate around practical decisions engineers and policymakers face now: which model capabilities to prioritize, what data practices to require and how to measure social harms. She emphasized that choices about training data, evaluation benchmarks and deployment protections determine whether systems amplify existing inequalities or help mitigate them. Hao urged metrics that go beyond narrow benchmarks to capture harms such as misinformation spread, labor displacement and uneven access to benefits.

Ricaurte stressed the political and institutional dimensions of those same choices. She argued that governance cannot be limited to technical fixes or voluntary industry standards, because models concentrate power and shape public discourse. Ricaurte called for legal and regulatory guardrails that preserve democratic accountability, including clearer avenues for civil society to request audits and stronger antitrust scrutiny of dominant model providers.

Both speakers flagged environmental costs as a rising policy concern. They recommended lifecycle accounting for energy and resource use tied to model size and deployment intensity, and encouraged procurement policies that favor lower-footprint alternatives for public-sector use.

Policy and design recommendations

The panel offered overlapping, practical recommendations rather than a single prescriptive roadmap. They recommended the following actions for governments, research labs and companies:

• Mandate independent audits for high-risk deployments, with public summaries that disclose methodology and key findings, while protecting legitimately sensitive data.
• Expand evaluation suites to include social impact tests tailored to sectors such as health, education and labor, alongside traditional accuracy and safety checks.
• Require tiered licensing or certification for models based on risk profile, so higher-risk systems face stricter oversight and operational constraints.
• Strengthen data provenance standards so downstream users and regulators can trace and verify critical training sources.
• Support community-led research and participatory design processes to surface impacts on marginalized groups before large-scale deployments.

Both speakers urged caution about binary choices such as full model bans or unrestricted openness. Instead they recommended calibrated interventions that align incentives across industry, academia and civil society. The panel also discussed the role of public procurement in shaping markets, noting that government purchasing power can reward models designed for robustness and fairness rather than raw capability.

Why it matters

The session reframes AI policy debates as a set of concrete trade-offs about design, governance and distribution of benefits. Decisions now about audits, licensing and evaluation will determine whether large models entrench existing power imbalances or are steered toward serving diverse public needs. Policymakers, corporate leaders and researchers who ignore those trade-offs risk creating systems that are technically impressive but socially harmful.