Millions of people are turning to generative AI for medical and fitness advice, opening new possibilities for more accessible, personalized health guidance. However, AI-generated recommendations aren’t always accurate or safe. In exercise programming, inappropriate advice could lead to overexertion, inadequate recovery, or injury. As AI takes on a greater role in supporting physical activity, the question of who or what determines safe limits becomes increasingly important.
When Health Advice Sounds Right But May Not Be Safe
A 2026 study in npj Digital Medicine evaluated 888 chatbot responses to patient medical questions and found that 22% to 43% of responses were problematic, with 5% to 13% classified as potentially unsafe, depending on the AI model. Similarly, a JAMA review of 519 studies evaluating large language models in healthcare found that only 5% used real patient care data.
While these studies didn’t specifically evaluate fitness coaching, they highlight a fundamental vulnerability. Generative AI can produce recommendations that sound reasonable without knowing whether they are safe for a specific individual’s health or physical abilities.
Separating AI Coaching from Safety Decisions
One way to make AI-powered exercise guidance safer is to separate personalized coaching from decisions about how much activity someone can safely handle. Rather than giving generative AI complete control, predefined rules establish boundaries on what the system can recommend.
This dual architecture is the approach behind Training Intelligence®, which uses two complementary components:
- Generative AI guides the coaching direction: By interpreting goals, recent activity, and subjective feedback, it determines whether an individual should continue as planned, reduce intensity, or prioritize recovery.
- Deterministic software enforces safety limits: By following predefined, testable rules grounded in exercise science, it produces consistent recommendations for appropriate exercise intensity and duration.
What Determines Safe Exercise Limits?
Safe exercise guidance requires understanding more than just someone’s baseline fitness level. It also requires recognizing how much activity they can reasonably handle on any given day.
Pear’s architecture considers four dynamic factors when establishing daily exercise limits:
- Daily readiness: Sleep, soreness, energy, mood, and stress are evaluated against the individual’s usual patterns.
- Training capacity: Recent activity and overall workload help establish appropriate limits for intensity and duration.
- Recovery: The intensity of previous workouts and the time needed for the body to recover influence current exercise recommendations.
- Program history: Individuals starting or returning to exercise receive more conservative recommendations than those who have been exercising regularly.
Controlling the Pace of Progression
Safety also means controlling how quickly exercise demands increase. The Physical Activity Guidelines for Americans recommend gradual progression because increasing activity too quickly can increase the risk of musculoskeletal injury.
Pear builds evidence-based progression limits into its system to help ensure that exercise intensity and duration increase gradually. Rather than advancing someone’s workout plan simply because they’re motivated or consistently exercising, the system considers whether they’re physically ready to do more.
Setting a Higher Standard for Agentic AI Safety
For organizations deploying AI across diverse populations, safety must be built into the system’s architecture through clear safeguards and consistent rules. As AI takes on a larger role in guiding health behaviors, success will depend not just on how personalized its recommendations are, but on how well they reflect what each individual can safely do.






