How AI can actually be helpful

Unlocking Helpfulness in Generative AI

A hiring platform powered by generative AI begins screening resumes and quietly downgrades applications from candidates with non-Western names or career paths outside conventional tech hubs, quietly narrowing opportunity before any human ever reviews the files.

Timnit Gebru's talk "How AI can actually be helpful" centers on the ethical challenges inherent in developing AI systems and stresses that diverse perspectives are essential for reducing bias and harm. She argues that AI must be built to prioritize fairness, minimize damage, and produce genuine societal benefits instead of simply scaling existing prejudices.

Gebru frames these issues as design choices rather than inevitable side effects. By insisting on inclusive teams and rigorous scrutiny of training data and objectives, developers can steer systems away from outcomes that amplify disparities. The talk positions ethical reflection not as an afterthought but as the foundation for any claim that AI is truly helpful.

Applied to the hiring platform, Gebru's emphasis on diverse perspectives would require involving people from varied backgrounds in every stage of model development. Such involvement could surface the ways training data encodes historical exclusion, prompting adjustments that prevent the system from systematically disadvantaging certain groups. Fairness would shift from a metric to an ongoing practice of listening and correction.

The result is AI that supports broader access to opportunity rather than reinforcing narrow definitions of merit. This approach directly addresses the central question of how generative AI can be guided toward beneficial outcomes while confronting its core ethical risks.

What lasting safeguards will we build so that the next generation of AI tools expands possibility for everyone they touch?