Results, Reports & Data

AI-Assisted Text: Mandatory Review and Limitations

July 27, 2026 Marius Schober 2 min read

Outcome

Review automatically generated AI text safely and remove any statement that is incorrect, unsupported, or unsuitable for the recorded context.

Applies to

In the current workflow, AI-assisted text is generated automatically as part of the report. It can be wrong. A qualified practitioner must review it before the report is used or shared.

Warning that automatically generated AI text must be checked against the measured trial
Mandatory AI text review
Sanitized report pages used to compare AI-assisted text with tables, graphs, and examination context
Report evidence for human review

Before you begin

  • Confirm the trial itself is usable.
  • Open the saved examination, metrics table, graphs, notes, and generated report together.
  • Make sure the reviewer understands the protocol and the person’s relevant context.
  • Do not infer the report service’s transfer, hosting, storage, retention, deletion, or anonymisation behaviour from this guide; those details are excluded until separately verified.

Steps

  1. Review the measured values, units, graphs, and task notes before reading the AI text.
  2. Check every value mentioned in the AI text against the metrics table.
  3. Check every statement about timing, side, phase, or movement against the selected configuration and saved trace.
  4. Remove any formula, reference range, threshold, or comparison that is not present in approved documentation for that report.
  5. Remove causal claims, clinical conclusions, treatment instructions, risk classifications, and unsupported recommendations.
  6. Add necessary context about protocol quality, setup, deviations, and repeat-trial variation.
  7. If the text cannot be verified, exclude it rather than rewriting it as a fact.
  8. The qualified reviewer makes and records the final decision about what may be communicated.

You should now see

Any retained AI-assisted text agrees with the measured values, trace, protocol, configuration, and reviewer judgment, with unsupported content removed.

Stop or repeat when

  • The AI text conflicts with a value, unit, graph, event sequence, note, or observed task.
  • The reviewer cannot verify a statement from the available evidence.
  • The report would be shared before qualified review is complete.

If it does not work

Do not rely on the AI wording. Use the reviewed measured content and factual protocol notes only, and preserve the original generated report when investigating an error.