What is reasoning effort?
The Harvard Chan AI Glossary
In our classes, we often pause to explain the ideas behind AI and the language used to describe it. These posts grew out of those conversations. They’re for anyone who would like to understand AI better, with no technical background assumed.
Reasoning effort is a setting in some AI models or products that influences how much computation the system may spend working through a task before or while producing a response. Computation means the calculations performed by the software. The setting is intended to adjust the work allocated to a response, not to measure a human-like experience of thinking.
A model is the trained component that interprets information and generates responses. The same model may offer different effort settings for different tasks. This is separate from choosing a different model, although the available settings and their meanings depend on the particular model and product.
For example, extracting a clearly stated date from a short passage may require a different approach from examining conflicting requirements in a complicated project. More computational work may help with some difficult tasks, but it can also increase response time and cost. The outcome needs to be checked rather than assumed.
Some services describe this additional work in reasoning tokens. Tokens are small units used in processing or generating model content; for visible text they may be words, word fragments or punctuation. Reasoning-token accounting concerns intermediate computation supported by the service, not necessarily a complete explanation displayed to the user.
A longer visible answer does not prove that more useful reasoning occurred. A system can spend more effort and still rely on an incorrect premise. It can also produce a concise final answer after substantial internal work. Response length, effort allocation and correctness are different properties.
The setting does not repair missing information. If a model has not been given the current project requirements, allowing it more time may produce a more elaborate response to the wrong task. Clear context—the information supplied for the task—and independent checks remain important.
In a clinical operations workflow, extracting a clearly stated appointment date from a message and reconciling a complex set of conflicting prior-authorization rules are very different tasks. Where supported, compare effort settings on those tasks and inspect whether extra computation reduces meaningful errors or only increases latency and token cost. More reasoning will not supply a missing lab result or make an outdated guideline current.
Sources: OpenAI — Reasoning Models.
If you would like to find out more about AI and improve your leadership skills, consider the AI in Health Care Certificate of Specialization from the Advanced Learning Academy at Harvard T.H. Chan School of Public Health. You can take an individual course or work toward the full certificate.
AI is a fast-moving field. This explanation reflects our class discussions and related material at the time of writing. Terminology, capabilities and tools may change.
For suggested additions or corrections, please email me at tpanch@hsph.harvard.edu.
Dr. Trishan Panch
Program Director, AI in Health Care
Harvard T.H. Chan School of Public Health — Advanced Learning Academy