Kristen Swanson, head of AI literacy research and learning at Anthropic, emphasizes that true AI proficiency is not about using the tool for every task. Instead, the core skill is the ability to judge which specific tasks are suitable for AI delegation.
Swanson introduces the concept of 'judgment cost,' where reviewing AI-generated content can be more demanding than completing the task manually. For tasks where a person is already proficient, evaluating AI output may consume more time than the original work would have.
Technical limitations, such as hallucinations, also dictate when AI should be avoided. For instance, when querying about obscure research papers or researchers not well-represented in training data, the AI is more likely to produce inaccurate information.
As AI capabilities evolve, the boundary between manual and automated tasks shifts. Swanson suggests that tasks currently requiring human effort may become suitable for AI in six months, necessitating regular re-testing of the model's abilities.
This approach prevents 'capability perception lag,' where users rely on outdated experiences and fail to adopt new model improvements. Regular testing helps users understand exactly what current models can and cannot achieve.
Ultimately, Swanson argues that mastery of AI is not defined by the number of prompts or tools used, but by the critical thinking applied to when and how to use the AI until these judgments become a natural habit.
