Why English? | Reporting CONSORT-AI Trials Through Intended Use, Input Data, Human Interaction and Error Analysis

CONSORT-AI clinical trial reporting matters because an artificial-intelligence intervention is more than a model score. People searching for AI clinical trials, medical AI evidence or human–AI interaction need to know the intended use, algorithm version, input requirements, user expertise, output, decision pathway and error consequences.

The open Nature Medicine CONSORT-AI statement, published on 9 September 2020 and checked on 5 November 2026, adds 14 AI-specific items to the core CONSORT framework. The open full-text version explains 11 extensions and three elaborations concerning intended use, versioning, input and output data, human–AI interaction and performance errors.

English does the linking work. A report must move from who was enrolled to which inputs were usable, what the system produced, how a person acted and what outcome followed. This article explains those links, their limits and practical questions for students, parents and future researchers. It is educational information, not medical advice.

Use this guide with eduKate’s CONSORT 2025 trial-reporting guide and GRADE certainty guide. The SETC critical-reading guide adds a general claims-and-evidence routine.

---

Identify the AI Trial in the Title and Abstract

Make the AI component and trial design findable before readers reach the methods.

What the trial report must make visible

CONSORT-AI adds 14 AI-specific reporting items to the core CONSORT framework, comprising 11 extensions and three elaborations. For identify the ai trial in the title and abstract, connect the technical description to the decision it changes. Name the participant, input, user, algorithm version, output and action in ordinary sentences before discussing performance. This chain prevents the word “AI” from becoming a substitute for an intervention description.

Worked trial reading

For identify the ai trial in the title and abstract, a radiology system accepts an enrolled patient’s scan only after a technologist selects a region and rejects motion-corrupted images. Mark what happened before the system, inside the system and after the output. Then identify the denominator: enrolled participants, usable inputs, generated outputs or completed decisions. The denominator matters because a polished accuracy percentage can exclude failed inputs or missing outputs unless the report keeps them in view.

Error, consequence and uncertainty

Ask which failures are random, which repeat under particular conditions and what consequence follows. Under identify the ai trial in the title and abstract, report estimates with uncertainty and describe error cases without implying that every error has equal clinical importance. A false reassurance, an unnecessary follow-up and a temporary processing failure are not interchangeable outcomes.

Practical boundary and next move

For identify the ai trial in the title and abstract, do not turn trial reporting into personal medical advice or assume one trial proves that a system is ready for every hospital. Finish the section with one check a reader can perform: inspect the protocol, participant flow, input exclusions, version history, error table, subgroup analysis or decision pathway. English is valuable here because it converts an opaque technology claim into an auditable sequence of people, data, actions and limits.

---

State the Intended Use

Name the task, target users, participants and decision the intervention is intended to support.

What the trial report must make visible

The guideline asks authors to state the AI system’s intended use, version, input requirements, output and place in the clinical pathway. For state the intended use, connect the technical description to the decision it changes. Name the participant, input, user, algorithm version, output and action in ordinary sentences before discussing performance. This chain prevents the word “AI” from becoming a substitute for an intervention description.

Worked trial reading

For state the intended use, a risk score appears as a percentage, but clinicians act only above a stated threshold and can override the recommendation after reviewing other evidence. Mark what happened before the system, inside the system and after the output. Then identify the denominator: enrolled participants, usable inputs, generated outputs or completed decisions. The denominator matters because a polished accuracy percentage can exclude failed inputs or missing outputs unless the report keeps them in view.

Error, consequence and uncertainty

Ask which failures are random, which repeat under particular conditions and what consequence follows. Under state the intended use, report estimates with uncertainty and describe error cases without implying that every error has equal clinical importance. A false reassurance, an unnecessary follow-up and a temporary processing failure are not interchangeable outcomes.

Practical boundary and next move

For state the intended use, a high average metric cannot substitute for patient outcomes, error consequences, workflow fit and uncertainty. Finish the section with one check a reader can perform: inspect the protocol, participant flow, input exclusions, version history, error table, subgroup analysis or decision pathway. English is valuable here because it converts an opaque technology claim into an auditable sequence of people, data, actions and limits.

---

Place the AI in the Clinical Pathway

Show what happens immediately before and after the system rather than presenting it as a detached model.

What the trial report must make visible

Input-data eligibility is different from participant eligibility: an enrolled person can still produce an image, signal or record that the system cannot process. For place the ai in the clinical pathway, connect the technical description to the decision it changes. Name the participant, input, user, algorithm version, output and action in ordinary sentences before discussing performance. This chain prevents the word “AI” from becoming a substitute for an intervention description.

Worked trial reading

For place the ai in the clinical pathway, the model gives no output for a subset of participants; the report keeps those failures in the flow and explains how the primary analysis handled them. Mark what happened before the system, inside the system and after the output. Then identify the denominator: enrolled participants, usable inputs, generated outputs or completed decisions. The denominator matters because a polished accuracy percentage can exclude failed inputs or missing outputs unless the report keeps them in view.

Error, consequence and uncertainty

Ask which failures are random, which repeat under particular conditions and what consequence follows. Under place the ai in the clinical pathway, report estimates with uncertainty and describe error cases without implying that every error has equal clinical importance. A false reassurance, an unnecessary follow-up and a temporary processing failure are not interchangeable outcomes.

Practical boundary and next move

For place the ai in the clinical pathway, fairness requires appropriate subgroup evidence and context; merely listing demographic variables is not a fairness analysis. Finish the section with one check a reader can perform: inspect the protocol, participant flow, input exclusions, version history, error table, subgroup analysis or decision pathway. English is valuable here because it converts an opaque technology claim into an auditable sequence of people, data, actions and limits.

---

Define Participant Eligibility

Explain who could enter the trial and why those criteria match the intended use.

What the trial report must make visible

The report should describe acquisition, selection and preprocessing before the AI system receives an input. For define participant eligibility, connect the technical description to the decision it changes. Name the participant, input, user, algorithm version, output and action in ordinary sentences before discussing performance. This chain prevents the word “AI” from becoming a substitute for an intervention description.

Worked trial reading

For define participant eligibility, nurses receive structured training before using an alert, while the comparator arm follows the existing escalation pathway without the alert. Mark what happened before the system, inside the system and after the output. Then identify the denominator: enrolled participants, usable inputs, generated outputs or completed decisions. The denominator matters because a polished accuracy percentage can exclude failed inputs or missing outputs unless the report keeps them in view.

Error, consequence and uncertainty

Ask which failures are random, which repeat under particular conditions and what consequence follows. Under define participant eligibility, report estimates with uncertainty and describe error cases without implying that every error has equal clinical importance. A false reassurance, an unnecessary follow-up and a temporary processing failure are not interchangeable outcomes.

Practical boundary and next move

For define participant eligibility, cONSORT-AI complements the applicable core CONSORT statement and does not replace protocol, ethics, regulatory or technical documentation. Finish the section with one check a reader can perform: inspect the protocol, participant flow, input exclusions, version history, error table, subgroup analysis or decision pathway. English is valuable here because it converts an opaque technology claim into an auditable sequence of people, data, actions and limits.

---

Define Input-Data Eligibility

State which images, signals, records or other inputs the system could accept.

What the trial report must make visible

Human–AI interaction includes the expertise needed to prepare inputs, interpret outputs and act on a recommendation. For define input-data eligibility, connect the technical description to the decision it changes. Name the participant, input, user, algorithm version, output and action in ordinary sentences before discussing performance. This chain prevents the word “AI” from becoming a substitute for an intervention description.

Worked trial reading

For define input-data eligibility, average performance looks strong, yet error analysis finds repeated misses in low-quality inputs and one under-represented subgroup. Mark what happened before the system, inside the system and after the output. Then identify the denominator: enrolled participants, usable inputs, generated outputs or completed decisions. The denominator matters because a polished accuracy percentage can exclude failed inputs or missing outputs unless the report keeps them in view.

Error, consequence and uncertainty

Ask which failures are random, which repeat under particular conditions and what consequence follows. Under define input-data eligibility, report estimates with uncertainty and describe error cases without implying that every error has equal clinical importance. A false reassurance, an unnecessary follow-up and a temporary processing failure are not interchangeable outcomes.

Practical boundary and next move

For define input-data eligibility, missing information should remain visible rather than being filled with assumptions about how the software probably worked. Finish the section with one check a reader can perform: inspect the protocol, participant flow, input exclusions, version history, error table, subgroup analysis or decision pathway. English is valuable here because it converts an opaque technology claim into an auditable sequence of people, data, actions and limits.

---

Name the Algorithm Version

Identify the evaluated version and report whether it changed during recruitment or follow-up.

What the trial report must make visible

CONSORT-AI item 19 asks for analysis of performance errors so systematic failures and their consequences are not hidden by an average score. For name the algorithm version, connect the technical description to the decision it changes. Name the participant, input, user, algorithm version, output and action in ordinary sentences before discussing performance. This chain prevents the word “AI” from becoming a substitute for an intervention description.

Worked trial reading

For name the algorithm version, a software update changes calibration midway through recruitment, so the report separates versions rather than treating the trial as one unchanged intervention. Mark what happened before the system, inside the system and after the output. Then identify the denominator: enrolled participants, usable inputs, generated outputs or completed decisions. The denominator matters because a polished accuracy percentage can exclude failed inputs or missing outputs unless the report keeps them in view.

Error, consequence and uncertainty

Ask which failures are random, which repeat under particular conditions and what consequence follows. Under name the algorithm version, report estimates with uncertainty and describe error cases without implying that every error has equal clinical importance. A false reassurance, an unnecessary follow-up and a temporary processing failure are not interchangeable outcomes.

Practical boundary and next move

For name the algorithm version, a transparent report can still describe a weak trial, while a strong trial may remain difficult to judge if crucial details are omitted. Finish the section with one check a reader can perform: inspect the protocol, participant flow, input exclusions, version history, error table, subgroup analysis or decision pathway. English is valuable here because it converts an opaque technology claim into an auditable sequence of people, data, actions and limits.

---

Describe Data Acquisition

Report how inputs were produced, collected and transferred before analysis.

What the trial report must make visible

Reporting a fixed algorithm version matters because changes during a trial can alter the intervention being evaluated. For describe data acquisition, connect the technical description to the decision it changes. Name the participant, input, user, algorithm version, output and action in ordinary sentences before discussing performance. This chain prevents the word “AI” from becoming a substitute for an intervention description.

Worked trial reading

For describe data acquisition, an AI recommendation is displayed beside the source image, and the study records whether users accept, modify or ignore it. Mark what happened before the system, inside the system and after the output. Then identify the denominator: enrolled participants, usable inputs, generated outputs or completed decisions. The denominator matters because a polished accuracy percentage can exclude failed inputs or missing outputs unless the report keeps them in view.

Error, consequence and uncertainty

Ask which failures are random, which repeat under particular conditions and what consequence follows. Under describe data acquisition, report estimates with uncertainty and describe error cases without implying that every error has equal clinical importance. A false reassurance, an unnecessary follow-up and a temporary processing failure are not interchangeable outcomes.

Practical boundary and next move

For describe data acquisition, human override is not automatically good or bad; the report must show when, why and with what consequences it occurred. Finish the section with one check a reader can perform: inspect the protocol, participant flow, input exclusions, version history, error table, subgroup analysis or decision pathway. English is valuable here because it converts an opaque technology claim into an auditable sequence of people, data, actions and limits.

---

Describe Selection and Preprocessing

Explain cropping, cleaning, normalisation, feature preparation or other transformations.

What the trial report must make visible

A reporting checklist improves visibility but does not by itself establish methodological quality, fairness, safety or clinical benefit. For describe selection and preprocessing, connect the technical description to the decision it changes. Name the participant, input, user, algorithm version, output and action in ordinary sentences before discussing performance. This chain prevents the word “AI” from becoming a substitute for an intervention description.

Worked trial reading

For describe selection and preprocessing, a student reading a headline about an AI trial reconstructs the participants, data inputs, users, output and decision pathway before accepting the conclusion. Mark what happened before the system, inside the system and after the output. Then identify the denominator: enrolled participants, usable inputs, generated outputs or completed decisions. The denominator matters because a polished accuracy percentage can exclude failed inputs or missing outputs unless the report keeps them in view.

Error, consequence and uncertainty

Ask which failures are random, which repeat under particular conditions and what consequence follows. Under describe selection and preprocessing, report estimates with uncertainty and describe error cases without implying that every error has equal clinical importance. A false reassurance, an unnecessary follow-up and a temporary processing failure are not interchangeable outcomes.

Practical boundary and next move

For describe selection and preprocessing, transfer to education or careers should preserve the reading habit, not import clinical claims into a different field. Finish the section with one check a reader can perform: inspect the protocol, participant flow, input exclusions, version history, error table, subgroup analysis or decision pathway. English is valuable here because it converts an opaque technology claim into an auditable sequence of people, data, actions and limits.

---

Report Poor-Quality Input Handling

Show whether weak inputs were rejected, repeated, repaired, imputed or passed to a human.

What the trial report must make visible

CONSORT-AI adds 14 AI-specific reporting items to the core CONSORT framework, comprising 11 extensions and three elaborations. For report poor-quality input handling, connect the technical description to the decision it changes. Name the participant, input, user, algorithm version, output and action in ordinary sentences before discussing performance. This chain prevents the word “AI” from becoming a substitute for an intervention description.

Worked trial reading

For report poor-quality input handling, a radiology system accepts an enrolled patient’s scan only after a technologist selects a region and rejects motion-corrupted images. Mark what happened before the system, inside the system and after the output. Then identify the denominator: enrolled participants, usable inputs, generated outputs or completed decisions. The denominator matters because a polished accuracy percentage can exclude failed inputs or missing outputs unless the report keeps them in view.

Error, consequence and uncertainty

Ask which failures are random, which repeat under particular conditions and what consequence follows. Under report poor-quality input handling, report estimates with uncertainty and describe error cases without implying that every error has equal clinical importance. A false reassurance, an unnecessary follow-up and a temporary processing failure are not interchangeable outcomes.

Practical boundary and next move

For report poor-quality input handling, do not turn trial reporting into personal medical advice or assume one trial proves that a system is ready for every hospital. Finish the section with one check a reader can perform: inspect the protocol, participant flow, input exclusions, version history, error table, subgroup analysis or decision pathway. English is valuable here because it converts an opaque technology claim into an auditable sequence of people, data, actions and limits.

---

Describe Human Interaction With Inputs

Name who selected or prepared inputs and which expertise that work required.

What the trial report must make visible

The guideline asks authors to state the AI system’s intended use, version, input requirements, output and place in the clinical pathway. For describe human interaction with inputs, connect the technical description to the decision it changes. Name the participant, input, user, algorithm version, output and action in ordinary sentences before discussing performance. This chain prevents the word “AI” from becoming a substitute for an intervention description.

Worked trial reading

For describe human interaction with inputs, a risk score appears as a percentage, but clinicians act only above a stated threshold and can override the recommendation after reviewing other evidence. Mark what happened before the system, inside the system and after the output. Then identify the denominator: enrolled participants, usable inputs, generated outputs or completed decisions. The denominator matters because a polished accuracy percentage can exclude failed inputs or missing outputs unless the report keeps them in view.

Error, consequence and uncertainty

Ask which failures are random, which repeat under particular conditions and what consequence follows. Under describe human interaction with inputs, report estimates with uncertainty and describe error cases without implying that every error has equal clinical importance. A false reassurance, an unnecessary follow-up and a temporary processing failure are not interchangeable outcomes.

Practical boundary and next move

For describe human interaction with inputs, a high average metric cannot substitute for patient outcomes, error consequences, workflow fit and uncertainty. Finish the section with one check a reader can perform: inspect the protocol, participant flow, input exclusions, version history, error table, subgroup analysis or decision pathway. English is valuable here because it converts an opaque technology claim into an auditable sequence of people, data, actions and limits.

---

Define the AI Output

State the form, scale, categories and uncertainty of the output.

What the trial report must make visible

Input-data eligibility is different from participant eligibility: an enrolled person can still produce an image, signal or record that the system cannot process. For define the ai output, connect the technical description to the decision it changes. Name the participant, input, user, algorithm version, output and action in ordinary sentences before discussing performance. This chain prevents the word “AI” from becoming a substitute for an intervention description.

Worked trial reading

For define the ai output, the model gives no output for a subset of participants; the report keeps those failures in the flow and explains how the primary analysis handled them. Mark what happened before the system, inside the system and after the output. Then identify the denominator: enrolled participants, usable inputs, generated outputs or completed decisions. The denominator matters because a polished accuracy percentage can exclude failed inputs or missing outputs unless the report keeps them in view.

Error, consequence and uncertainty

Ask which failures are random, which repeat under particular conditions and what consequence follows. Under define the ai output, report estimates with uncertainty and describe error cases without implying that every error has equal clinical importance. A false reassurance, an unnecessary follow-up and a temporary processing failure are not interchangeable outcomes.

Practical boundary and next move

For define the ai output, fairness requires appropriate subgroup evidence and context; merely listing demographic variables is not a fairness analysis. Finish the section with one check a reader can perform: inspect the protocol, participant flow, input exclusions, version history, error table, subgroup analysis or decision pathway. English is valuable here because it converts an opaque technology claim into an auditable sequence of people, data, actions and limits.

---

Explain How Outputs Changed Decisions

Report how users interpreted, accepted, modified or ignored the output.

What the trial report must make visible

The report should describe acquisition, selection and preprocessing before the AI system receives an input. For explain how outputs changed decisions, connect the technical description to the decision it changes. Name the participant, input, user, algorithm version, output and action in ordinary sentences before discussing performance. This chain prevents the word “AI” from becoming a substitute for an intervention description.

Worked trial reading

For explain how outputs changed decisions, nurses receive structured training before using an alert, while the comparator arm follows the existing escalation pathway without the alert. Mark what happened before the system, inside the system and after the output. Then identify the denominator: enrolled participants, usable inputs, generated outputs or completed decisions. The denominator matters because a polished accuracy percentage can exclude failed inputs or missing outputs unless the report keeps them in view.

Error, consequence and uncertainty

Ask which failures are random, which repeat under particular conditions and what consequence follows. Under explain how outputs changed decisions, report estimates with uncertainty and describe error cases without implying that every error has equal clinical importance. A false reassurance, an unnecessary follow-up and a temporary processing failure are not interchangeable outcomes.

Practical boundary and next move

For explain how outputs changed decisions, cONSORT-AI complements the applicable core CONSORT statement and does not replace protocol, ethics, regulatory or technical documentation. Finish the section with one check a reader can perform: inspect the protocol, participant flow, input exclusions, version history, error table, subgroup analysis or decision pathway. English is valuable here because it converts an opaque technology claim into an auditable sequence of people, data, actions and limits.

---

Describe User Training and Expertise

Make the skills and support needed to reproduce use visible.

What the trial report must make visible

Human–AI interaction includes the expertise needed to prepare inputs, interpret outputs and act on a recommendation. For describe user training and expertise, connect the technical description to the decision it changes. Name the participant, input, user, algorithm version, output and action in ordinary sentences before discussing performance. This chain prevents the word “AI” from becoming a substitute for an intervention description.

Worked trial reading

For describe user training and expertise, average performance looks strong, yet error analysis finds repeated misses in low-quality inputs and one under-represented subgroup. Mark what happened before the system, inside the system and after the output. Then identify the denominator: enrolled participants, usable inputs, generated outputs or completed decisions. The denominator matters because a polished accuracy percentage can exclude failed inputs or missing outputs unless the report keeps them in view.

Error, consequence and uncertainty

Ask which failures are random, which repeat under particular conditions and what consequence follows. Under describe user training and expertise, report estimates with uncertainty and describe error cases without implying that every error has equal clinical importance. A false reassurance, an unnecessary follow-up and a temporary processing failure are not interchangeable outcomes.

Practical boundary and next move

For describe user training and expertise, missing information should remain visible rather than being filled with assumptions about how the software probably worked. Finish the section with one check a reader can perform: inspect the protocol, participant flow, input exclusions, version history, error table, subgroup analysis or decision pathway. English is valuable here because it converts an opaque technology claim into an auditable sequence of people, data, actions and limits.

---

Show Workflow Integration

Explain timing, interface, alerts and responsibilities in the actual trial setting.

What the trial report must make visible

CONSORT-AI item 19 asks for analysis of performance errors so systematic failures and their consequences are not hidden by an average score. For show workflow integration, connect the technical description to the decision it changes. Name the participant, input, user, algorithm version, output and action in ordinary sentences before discussing performance. This chain prevents the word “AI” from becoming a substitute for an intervention description.

Worked trial reading

For show workflow integration, a software update changes calibration midway through recruitment, so the report separates versions rather than treating the trial as one unchanged intervention. Mark what happened before the system, inside the system and after the output. Then identify the denominator: enrolled participants, usable inputs, generated outputs or completed decisions. The denominator matters because a polished accuracy percentage can exclude failed inputs or missing outputs unless the report keeps them in view.

Error, consequence and uncertainty

Ask which failures are random, which repeat under particular conditions and what consequence follows. Under show workflow integration, report estimates with uncertainty and describe error cases without implying that every error has equal clinical importance. A false reassurance, an unnecessary follow-up and a temporary processing failure are not interchangeable outcomes.

Practical boundary and next move

For show workflow integration, a transparent report can still describe a weak trial, while a strong trial may remain difficult to judge if crucial details are omitted. Finish the section with one check a reader can perform: inspect the protocol, participant flow, input exclusions, version history, error table, subgroup analysis or decision pathway. English is valuable here because it converts an opaque technology claim into an auditable sequence of people, data, actions and limits.

---

Describe Intervention and Comparator Fairly

Give enough detail about both pathways to understand the causal contrast.

What the trial report must make visible

Reporting a fixed algorithm version matters because changes during a trial can alter the intervention being evaluated. For describe intervention and comparator fairly, connect the technical description to the decision it changes. Name the participant, input, user, algorithm version, output and action in ordinary sentences before discussing performance. This chain prevents the word “AI” from becoming a substitute for an intervention description.

Worked trial reading

For describe intervention and comparator fairly, an AI recommendation is displayed beside the source image, and the study records whether users accept, modify or ignore it. Mark what happened before the system, inside the system and after the output. Then identify the denominator: enrolled participants, usable inputs, generated outputs or completed decisions. The denominator matters because a polished accuracy percentage can exclude failed inputs or missing outputs unless the report keeps them in view.

Error, consequence and uncertainty

Ask which failures are random, which repeat under particular conditions and what consequence follows. Under describe intervention and comparator fairly, report estimates with uncertainty and describe error cases without implying that every error has equal clinical importance. A false reassurance, an unnecessary follow-up and a temporary processing failure are not interchangeable outcomes.

Practical boundary and next move

For describe intervention and comparator fairly, human override is not automatically good or bad; the report must show when, why and with what consequences it occurred. Finish the section with one check a reader can perform: inspect the protocol, participant flow, input exclusions, version history, error table, subgroup analysis or decision pathway. English is valuable here because it converts an opaque technology claim into an auditable sequence of people, data, actions and limits.

---

Report Outcomes and Harms

Separate technical performance, user behaviour, clinical outcomes and adverse consequences.

What the trial report must make visible

A reporting checklist improves visibility but does not by itself establish methodological quality, fairness, safety or clinical benefit. For report outcomes and harms, connect the technical description to the decision it changes. Name the participant, input, user, algorithm version, output and action in ordinary sentences before discussing performance. This chain prevents the word “AI” from becoming a substitute for an intervention description.

Worked trial reading

For report outcomes and harms, a student reading a headline about an AI trial reconstructs the participants, data inputs, users, output and decision pathway before accepting the conclusion. Mark what happened before the system, inside the system and after the output. Then identify the denominator: enrolled participants, usable inputs, generated outputs or completed decisions. The denominator matters because a polished accuracy percentage can exclude failed inputs or missing outputs unless the report keeps them in view.

Error, consequence and uncertainty

Ask which failures are random, which repeat under particular conditions and what consequence follows. Under report outcomes and harms, report estimates with uncertainty and describe error cases without implying that every error has equal clinical importance. A false reassurance, an unnecessary follow-up and a temporary processing failure are not interchangeable outcomes.

Practical boundary and next move

For report outcomes and harms, transfer to education or careers should preserve the reading habit, not import clinical claims into a different field. Finish the section with one check a reader can perform: inspect the protocol, participant flow, input exclusions, version history, error table, subgroup analysis or decision pathway. English is valuable here because it converts an opaque technology claim into an auditable sequence of people, data, actions and limits.

---

Explain Sample Size and Analysis

Connect the calculation and model to the trial’s participant and clustering structure.

What the trial report must make visible

CONSORT-AI adds 14 AI-specific reporting items to the core CONSORT framework, comprising 11 extensions and three elaborations. For explain sample size and analysis, connect the technical description to the decision it changes. Name the participant, input, user, algorithm version, output and action in ordinary sentences before discussing performance. This chain prevents the word “AI” from becoming a substitute for an intervention description.

Worked trial reading

For explain sample size and analysis, a radiology system accepts an enrolled patient’s scan only after a technologist selects a region and rejects motion-corrupted images. Mark what happened before the system, inside the system and after the output. Then identify the denominator: enrolled participants, usable inputs, generated outputs or completed decisions. The denominator matters because a polished accuracy percentage can exclude failed inputs or missing outputs unless the report keeps them in view.

Error, consequence and uncertainty

Ask which failures are random, which repeat under particular conditions and what consequence follows. Under explain sample size and analysis, report estimates with uncertainty and describe error cases without implying that every error has equal clinical importance. A false reassurance, an unnecessary follow-up and a temporary processing failure are not interchangeable outcomes.

Practical boundary and next move

For explain sample size and analysis, do not turn trial reporting into personal medical advice or assume one trial proves that a system is ready for every hospital. Finish the section with one check a reader can perform: inspect the protocol, participant flow, input exclusions, version history, error table, subgroup analysis or decision pathway. English is valuable here because it converts an opaque technology claim into an auditable sequence of people, data, actions and limits.

---

Report Randomisation and Blinding

State who was assigned, what could be masked and where masking was impossible.

What the trial report must make visible

The guideline asks authors to state the AI system’s intended use, version, input requirements, output and place in the clinical pathway. For report randomisation and blinding, connect the technical description to the decision it changes. Name the participant, input, user, algorithm version, output and action in ordinary sentences before discussing performance. This chain prevents the word “AI” from becoming a substitute for an intervention description.

Worked trial reading

For report randomisation and blinding, a risk score appears as a percentage, but clinicians act only above a stated threshold and can override the recommendation after reviewing other evidence. Mark what happened before the system, inside the system and after the output. Then identify the denominator: enrolled participants, usable inputs, generated outputs or completed decisions. The denominator matters because a polished accuracy percentage can exclude failed inputs or missing outputs unless the report keeps them in view.

Error, consequence and uncertainty

Ask which failures are random, which repeat under particular conditions and what consequence follows. Under report randomisation and blinding, report estimates with uncertainty and describe error cases without implying that every error has equal clinical importance. A false reassurance, an unnecessary follow-up and a temporary processing failure are not interchangeable outcomes.

Practical boundary and next move

For report randomisation and blinding, a high average metric cannot substitute for patient outcomes, error consequences, workflow fit and uncertainty. Finish the section with one check a reader can perform: inspect the protocol, participant flow, input exclusions, version history, error table, subgroup analysis or decision pathway. English is valuable here because it converts an opaque technology claim into an auditable sequence of people, data, actions and limits.

---

Keep Participant and Input Flow Visible

Show exclusions and failures at both the person and input levels.

What the trial report must make visible

Input-data eligibility is different from participant eligibility: an enrolled person can still produce an image, signal or record that the system cannot process. For keep participant and input flow visible, connect the technical description to the decision it changes. Name the participant, input, user, algorithm version, output and action in ordinary sentences before discussing performance. This chain prevents the word “AI” from becoming a substitute for an intervention description.

Worked trial reading

For keep participant and input flow visible, the model gives no output for a subset of participants; the report keeps those failures in the flow and explains how the primary analysis handled them. Mark what happened before the system, inside the system and after the output. Then identify the denominator: enrolled participants, usable inputs, generated outputs or completed decisions. The denominator matters because a polished accuracy percentage can exclude failed inputs or missing outputs unless the report keeps them in view.

Error, consequence and uncertainty

Ask which failures are random, which repeat under particular conditions and what consequence follows. Under keep participant and input flow visible, report estimates with uncertainty and describe error cases without implying that every error has equal clinical importance. A false reassurance, an unnecessary follow-up and a temporary processing failure are not interchangeable outcomes.

Practical boundary and next move

For keep participant and input flow visible, fairness requires appropriate subgroup evidence and context; merely listing demographic variables is not a fairness analysis. Finish the section with one check a reader can perform: inspect the protocol, participant flow, input exclusions, version history, error table, subgroup analysis or decision pathway. English is valuable here because it converts an opaque technology claim into an auditable sequence of people, data, actions and limits.

---

Report Missing Outputs and System Failures

Do not let cases without an AI result disappear from the analysis narrative.

What the trial report must make visible

The report should describe acquisition, selection and preprocessing before the AI system receives an input. For report missing outputs and system failures, connect the technical description to the decision it changes. Name the participant, input, user, algorithm version, output and action in ordinary sentences before discussing performance. This chain prevents the word “AI” from becoming a substitute for an intervention description.

Worked trial reading

For report missing outputs and system failures, nurses receive structured training before using an alert, while the comparator arm follows the existing escalation pathway without the alert. Mark what happened before the system, inside the system and after the output. Then identify the denominator: enrolled participants, usable inputs, generated outputs or completed decisions. The denominator matters because a polished accuracy percentage can exclude failed inputs or missing outputs unless the report keeps them in view.

Error, consequence and uncertainty

Ask which failures are random, which repeat under particular conditions and what consequence follows. Under report missing outputs and system failures, report estimates with uncertainty and describe error cases without implying that every error has equal clinical importance. A false reassurance, an unnecessary follow-up and a temporary processing failure are not interchangeable outcomes.

Practical boundary and next move

For report missing outputs and system failures, cONSORT-AI complements the applicable core CONSORT statement and does not replace protocol, ethics, regulatory or technical documentation. Finish the section with one check a reader can perform: inspect the protocol, participant flow, input exclusions, version history, error table, subgroup analysis or decision pathway. English is valuable here because it converts an opaque technology claim into an auditable sequence of people, data, actions and limits.

---

Analyse Performance Errors

Describe the nature, pattern and consequence of mistakes rather than only their count.

What the trial report must make visible

Human–AI interaction includes the expertise needed to prepare inputs, interpret outputs and act on a recommendation. For analyse performance errors, connect the technical description to the decision it changes. Name the participant, input, user, algorithm version, output and action in ordinary sentences before discussing performance. This chain prevents the word “AI” from becoming a substitute for an intervention description.

Worked trial reading

For analyse performance errors, average performance looks strong, yet error analysis finds repeated misses in low-quality inputs and one under-represented subgroup. Mark what happened before the system, inside the system and after the output. Then identify the denominator: enrolled participants, usable inputs, generated outputs or completed decisions. The denominator matters because a polished accuracy percentage can exclude failed inputs or missing outputs unless the report keeps them in view.

Error, consequence and uncertainty

Ask which failures are random, which repeat under particular conditions and what consequence follows. Under analyse performance errors, report estimates with uncertainty and describe error cases without implying that every error has equal clinical importance. A false reassurance, an unnecessary follow-up and a temporary processing failure are not interchangeable outcomes.

Practical boundary and next move

For analyse performance errors, missing information should remain visible rather than being filled with assumptions about how the software probably worked. Finish the section with one check a reader can perform: inspect the protocol, participant flow, input exclusions, version history, error table, subgroup analysis or decision pathway. English is valuable here because it converts an opaque technology claim into an auditable sequence of people, data, actions and limits.

---

Examine Subgroups and Fairness

Report prespecified variation with denominators, uncertainty and context.

What the trial report must make visible

CONSORT-AI item 19 asks for analysis of performance errors so systematic failures and their consequences are not hidden by an average score. For examine subgroups and fairness, connect the technical description to the decision it changes. Name the participant, input, user, algorithm version, output and action in ordinary sentences before discussing performance. This chain prevents the word “AI” from becoming a substitute for an intervention description.

Worked trial reading

For examine subgroups and fairness, a software update changes calibration midway through recruitment, so the report separates versions rather than treating the trial as one unchanged intervention. Mark what happened before the system, inside the system and after the output. Then identify the denominator: enrolled participants, usable inputs, generated outputs or completed decisions. The denominator matters because a polished accuracy percentage can exclude failed inputs or missing outputs unless the report keeps them in view.

Error, consequence and uncertainty

Ask which failures are random, which repeat under particular conditions and what consequence follows. Under examine subgroups and fairness, report estimates with uncertainty and describe error cases without implying that every error has equal clinical importance. A false reassurance, an unnecessary follow-up and a temporary processing failure are not interchangeable outcomes.

Practical boundary and next move

For examine subgroups and fairness, a transparent report can still describe a weak trial, while a strong trial may remain difficult to judge if crucial details are omitted. Finish the section with one check a reader can perform: inspect the protocol, participant flow, input exclusions, version history, error table, subgroup analysis or decision pathway. English is valuable here because it converts an opaque technology claim into an auditable sequence of people, data, actions and limits.

---

Did You Know? Automation Can Change Human Decisions

A system can influence behaviour even when its output is later overridden.

What the trial report must make visible

Reporting a fixed algorithm version matters because changes during a trial can alter the intervention being evaluated. For did you know? automation can change human decisions, connect the technical description to the decision it changes. Name the participant, input, user, algorithm version, output and action in ordinary sentences before discussing performance. This chain prevents the word “AI” from becoming a substitute for an intervention description.

Worked trial reading

For did you know? automation can change human decisions, an AI recommendation is displayed beside the source image, and the study records whether users accept, modify or ignore it. Mark what happened before the system, inside the system and after the output. Then identify the denominator: enrolled participants, usable inputs, generated outputs or completed decisions. The denominator matters because a polished accuracy percentage can exclude failed inputs or missing outputs unless the report keeps them in view.

Error, consequence and uncertainty

Ask which failures are random, which repeat under particular conditions and what consequence follows. Under did you know? automation can change human decisions, report estimates with uncertainty and describe error cases without implying that every error has equal clinical importance. A false reassurance, an unnecessary follow-up and a temporary processing failure are not interchangeable outcomes.

Practical boundary and next move

For did you know? automation can change human decisions, human override is not automatically good or bad; the report must show when, why and with what consequences it occurred. Finish the section with one check a reader can perform: inspect the protocol, participant flow, input exclusions, version history, error table, subgroup analysis or decision pathway. English is valuable here because it converts an opaque technology claim into an auditable sequence of people, data, actions and limits.

---

Finish With a CONSORT-AI Audit

Verify intended use, version, data flow, interaction, failures, errors, outcomes and limits together.

What the trial report must make visible

A reporting checklist improves visibility but does not by itself establish methodological quality, fairness, safety or clinical benefit. For finish with a consort-ai audit, connect the technical description to the decision it changes. Name the participant, input, user, algorithm version, output and action in ordinary sentences before discussing performance. This chain prevents the word “AI” from becoming a substitute for an intervention description.

Worked trial reading

For finish with a consort-ai audit, a student reading a headline about an AI trial reconstructs the participants, data inputs, users, output and decision pathway before accepting the conclusion. Mark what happened before the system, inside the system and after the output. Then identify the denominator: enrolled participants, usable inputs, generated outputs or completed decisions. The denominator matters because a polished accuracy percentage can exclude failed inputs or missing outputs unless the report keeps them in view.

Error, consequence and uncertainty

Ask which failures are random, which repeat under particular conditions and what consequence follows. Under finish with a consort-ai audit, report estimates with uncertainty and describe error cases without implying that every error has equal clinical importance. A false reassurance, an unnecessary follow-up and a temporary processing failure are not interchangeable outcomes.

Practical boundary and next move

For finish with a consort-ai audit, transfer to education or careers should preserve the reading habit, not import clinical claims into a different field. Finish the section with one check a reader can perform: inspect the protocol, participant flow, input exclusions, version history, error table, subgroup analysis or decision pathway. English is valuable here because it converts an opaque technology claim into an auditable sequence of people, data, actions and limits.

---

Frequently Asked Questions

What is CONSORT-AI?

It is a reporting extension for clinical trials evaluating interventions with an artificial-intelligence component.

How many AI-specific items does it add?

The 2020 extension adds 14 items: 11 extensions and three elaborations.

Does complete reporting prove an AI system works?

No. Reporting makes methods and results visible; effectiveness, safety, bias and applicability still require appropriate study design and appraisal.

Why report input-data eligibility separately?

A participant can be eligible while the image, signal or record supplied to the system is unusable. Both levels affect denominators and generalisability.

Why does human–AI interaction matter?

Users select or prepare inputs, interpret outputs and decide whether to act. Their expertise, interface and workflow can change the intervention’s effect.

Can students use the same method?

Yes. They can trace a tool claim through intended use, data, user, output, action, error and limit, while respecting the original clinical scope.

---

Useful Next Reading

• Read the primary CONSORT-AI statement.
• Use the open full text and checklist explanations.
• Connect the extension to CONSORT 2025 reporting.
• Read trial results through GRADE certainty.

Closing thought: The practical value of English is the ability to turn an AI trial from a polished claim into a traceable account of people, data, decisions, errors and uncertainty.