Human Affordance
Human Affordance addresses the mismatch between the operational speed of AI agents and the speed at which humans can meaningfully interact with them. While AI systems can process and execute tasks at rates orders of magnitude faster than human cognition and reaction time, this speed advantage often fails to translate into proportional improvements for human users. The concept suggests that the bottleneck lies not in AI capability but in the infrastructure and interface design that mediates human-AI interaction.
The Speed Disparity Problem
Current web infrastructure was designed primarily for human-paced interaction, creating friction when AI agents attempt to operate at machine speeds. Traditional web interfaces—forms, navigation, visual feedback loops—assume human decision-making timelines measured in seconds to minutes. When AI agents encounter these same interfaces, the inherent latency of DOM parsing and sequential interaction models negates their computational advantages.
Structured Interaction Models
To bridge this gap, modern HCI practices are shifting from raw prompting to structured interaction frameworks that separate intent verification from environment execution. A prominent example is Andrej Karpathy’s three-layer method for interacting with AI models, which optimizes the throughput and accuracy of agent outputs by decoupling the process into distinct stages:
- Spec (Specification): Explicitly defining the task constraints and desired output format before execution begins, reducing ambiguity in prompt engineering.
- Verifier: An intermediate layer that critiques or validates the AI’s draft output against the spec before finalization, acting as a quality control gate.
- Environment: The final execution stage where the verified action is applied to the target system, ensuring that only high-confidence operations modify state.
This structured approach mitigates the hallucination risk and aligns AI capabilities with human oversight requirements. See detailed breakdown in Karpathy’s Three-Layer AI Interaction Method: Spec, Verifier, Environment.