Danielle Perszyk

Danielle Perszyk is a cognitive scientist who leads Amazon’s human-computer interaction team, developing AI agents that expand human judgment, creativity, and control. Her work brings developmental psychology and the evolution of language to a practical engineering question: how can people and machines become more capable together?

Perszyk studied cognitive science and neuroscience at Williams College, investigating syntax in birdsong, and worked on autism and social-neuroscience research at the Yale Child Study Center. Her early writing about grammatical structure in birdsong explored how sophisticated communication arises beyond human language.

She completed her Northwestern University doctorate in 2018, studying language development and social cognition with Sandra Waxman. Their research on infant speech perception and object categorization showed how developing sensitivity to language influences early concept formation; their review of language and cognition in infancy examined how communication and thought develop together.

At Google, Perszyk developed research programs gathering human feedback for AI training. At Adept, she applied developmental and evolutionary perspectives to human-AI interaction before joining Amazon’s AGI lab. She also hosts Amazon’s Making a Mind, a podcast about building collaborative AI.

• Useful general intelligence: Perszyk judges AI by whether it strengthens human judgment, strategic thinking, creativity, and agency. Automation qualifies only when saved effort returns meaningful attention and control to the user.

• Shared world models: Human intelligence depends on interpreting intentions, directing attention jointly, and negotiating meaning. Perszyk argues that dependable agents likewise need shared grounding in an environment and interfaces that keep human and machine understandings aligned.

• Reliable computer-use agents: Amazon Nova Act combines a specialized model and Python SDK for browser automation. Perszyk emphasizes reliable elementary actions and granular developer control; an apartment-search example combines Pydantic schemas, Google Maps, parallel browser sessions, and pandas to compare listings and commutes.

• Human agency as an evaluation target: Capability benchmarks and engagement metrics cannot establish whether an agent actually improves someone’s life. Perszyk advocates measuring productivity, creativity, strategic reasoning, and user control, extending that concern to agency in education.

Her approach treats alignment as an ongoing relationship: useful products generate human-agent interactions, those interactions reveal what people need, and that feedback guides more capable, responsive systems.