The Real Benchmark of Behavior-Change AI Is Human Benefit—and Harm

Introducing Manager · November 5, 2026 · P07-6

An app doubles someone's meditation minutes. Its engagement dashboard looks excellent. Does that mean the person is better off?

Not necessarily. A longer practice period can coincide with poor sleep, guilt, work impairment, unwanted experiences, or loss of autonomy. The product goal must be what the person actually wanted, accompanied by independently measured benefits and harms—not time spent with the product.

Five separate evaluation layers

Reuse, don't reinvent, validated measures

The WHO-5 measures self-reported well-being over the past two weeks with five response values; it scores from 0 to 100 by multiplying their 0–25 sum by four. But the 2024 WHO published instrument carries a CC BY-NC-SA 3.0 IGO document license. A commercial app must check permissions rather than copying the questionnaire into product screens.

WHOQOL-BREF, PROMIS Global Health, FFMQ mindfulness dimensions, PTQ repetitive negative thinking, and MedEx-I meditation-related experiences answer different measurement questions. Their instruments, languages, validity evidence and licensing cannot be treated as interchangeable.

Manager's initial 0–10 internal burden and autonomy fields are unvalidated monitoring variables, not substitutes for authorized questionnaires.

The risk studies that product teams should read

A 2020 systematic review of 83 studies and 6,703 participants reported a pooled adverse-event estimate of 8.3%, with very different estimates for experimental and observational studies (3.7% versus 33.2%).

Lindahl et al. 2017 catalogued 59 kinds of meditation-related experiences across seven domains; experiences with similar descriptions could be helpful, distressing, or functionally impairing depending on person and context.

In a 2022 MedEx-I study, active questioning of 96 participants found negative functional impacts for 37%. A 2025 survey of 886 U.S. meditators reported meditation-related functional impairment in 9.1%.

These numbers cannot be averaged into a single universal risk. Definitions, sampling, intensity, follow-up, and active versus passive questioning differ. A product should not conclude “no harm” just because users did not spontaneously submit complaints.

The code we built

P07-6 adds a strictly nonclinical, read-only OutcomeObservation and assess_welfare decision boundary to Manager. It treats consent, revocation, self-chosen behavior, functioning, autonomy, reported burden, distress, sleep interruption, and impairment as distinct variables.

It can return MEASURE_ONLY, NO_ACTION, STOP, or STOP_AND_REFER. The last status is a need to stop and present appropriate support options, not an autonomous contact with a clinician. It makes no intervention, and thresholds are engineering safeguards—not validated diagnostic cutoffs.

A WHO-5 numeric scoring helper refuses use by default until appropriate rights are confirmed. No copyrighted items are embedded.

And in response to a real failure in the first HeartSteps audit, the code explicitly separates randomized assignment from post-assignment delivery status, preserving a NOT_ITT label when missing outcomes or the estimand require greater care.

A scientifically credible comparison

A model predicting tomorrow's behavior is not the same as a model predicting the increment caused by an intervention.

In randomized trials, intention-to-treat (ITT) compares people or decision points according to their original random assignment, not simply whether a message was delivered or opened. Missing outcomes, interference among people, changes in the outcome definition and departures from the protocol must all be examined.

Manager should preregister the primary user-chosen outcome, comparison, follow-up and safety guardrails, and publish negative findings and attrition alongside wins. Evidence that earlier researchers' interventions worked does not show that Manager's own product helps users.

Terms and FAQ

Outcome: a measurement following an intervention.

Adverse event: an unwanted experience; causal attribution requires separate assessment.

Functional impairment: disruption to everyday life.

ITT: analysis based on original randomized allocation.

Autonomy: the person's real ability to decline and revoke support.

Psychometric validity: evidence that a scale measures its intended construct.

Q: If WHO-5 improves, has a user attained Buddhist liberation? No. Self-reported well-being and traditional enlightenment are different constructs.

Q: Is all temporary discomfort a harm? No, but duration, distress, functional impact and the person's own preference matter. Harm should never be dismissed as necessary progress.

Q: Does a passing software test prove Manager makes people healthier? No. It verifies only the tested boundary under synthetic inputs.

See the detailed P07-6 source guide.

The standard for Manager is usefulness in life outside the benchmark, with uncertainty and harm measured honestly.