Article reader Listen + reading controls
Article reader
Preparing the reader…
Reading settings
Measure learning, not just AI use¶
OpenAI has introduced new tools for understanding artificial intelligence and learning outcomes. The work asks a more consequential question than whether students use an assistant or finish an assignment faster: what are they learning over time?
Every organization deploying artificial intelligence should ask the same question about its workforce.
Adoption metrics can reward dependence¶
Artificial intelligence (AI) programs often measure active users, prompts, time saved, or tasks completed. Those indicators reveal reach and activity. They do not show whether people are developing better judgment, losing important practice, or becoming more capable at work without the tool.
An analyst may produce a report faster while becoming less able to interrogate evidence. A developer may resolve more tickets while understanding fewer system interactions. A project manager may automate status synthesis while losing the conversations that reveal weak signals and conflict. Immediate productivity and long-term capability can move in different directions.
This is not an argument against assistance. It is an argument for defining learning as part of the outcome.
Performance and competence are not the same¶
Education research distinguishes what a person can do during supported practice from what they can retain and transfer later. Bjork and Bjork call attention to “desirable difficulties”: conditions that may slow apparent performance while strengthening durable learning.
AI can remove useful difficulty along with useless friction. If it immediately supplies the answer, structure, or interpretation, the user may complete the task without constructing the mental model needed for a novel case. Conversely, an assistant that asks a well-timed question, requests a prediction, or offers feedback after an attempt can improve both performance and learning.
The design choice is not binary delegation versus no delegation. It is deciding when the system should answer, coach, challenge, demonstrate, or remain quiet.
Treat workforce capability as a system requirement¶
For consequential knowledge work, teams can define two sets of outcomes:
- Task outcomes: quality, speed, error, cost, and mission effect with AI support.
- Capability outcomes: what users can explain, detect, decide, and recover from after repeated use.
Capability can be assessed through scenario exercises, delayed tests, peer review, error detection, or performance when the tool is unavailable or degraded. The measure should reflect the work. A cybersecurity analyst needs to recognize anomalous reasoning. A program manager needs to identify a weak dependency. A clinician needs to challenge an unsupported recommendation.
This also changes training. A one-time class on prompting is not enough. Teams need deliberate practice with realistic failure, examples of calibrated reliance, and feedback from experienced practitioners. Expertise grows through participation in the work, not simply through access to information.
Lave and Wenger's situated-learning framework explains why community matters. Knowledge develops through legitimate participation in a practice. If AI quietly removes the junior tasks through which newcomers learn how work fits together, organizations will need new pathways for observation, guided contribution, and increasing responsibility.
Ask what remains after the assistance¶
An adoption program should be able to answer:
- Which human capabilities remain mission-essential?
- Which routine burdens can be removed without eroding those capabilities?
- Where should AI prompt reflection instead of supplying completion?
- How will novices gain exposure to exceptions and recovery?
- What evidence would show that reliance has become unhealthy?
Usage is easy to count. Learning is harder, slower, and more important. The organizations that benefit most from AI will not merely complete today's work more efficiently. They will use the technology to build people who can recognize tomorrow's problem, question the system, and act when the assistance is wrong or absent.
Sources and research trail¶
- OpenAI, “New Tools for Understanding AI and Learning Outcomes” (March 4, 2026).
- Bjork and Bjork, “Making Things Hard on Yourself, but in a Good Way” (2011).
- Lave and Wenger, Situated Learning: Legitimate Peripheral Participation (1991).
- Parasuraman, Sheridan, and Wickens, “A Model for Types and Levels of Human Interaction with Automation” (2000).
- National Academies of Sciences, Engineering, and Medicine, Building America's Skilled Technical Workforce (2017).