Published
The CONSORT-AI checklist is the reporting standard that journals look for when a randomised trial tests an intervention with an artificial intelligence component. Published in 2020 as an extension to CONSORT 2010, it adds 14 items that ask authors to describe the algorithm version, the input data, the output, and the way clinicians acted on it. This guide explains each item in plain language, shows where published AI trials still leave gaps, and applies the checklist row by row to a real open-access trial so you can see what a complete answer looks like. One section covers SPIRIT-AI, the companion for trial protocols.
In Simple Words
CONSORT-AI is an add-on checklist for reporting randomised trials of AI tools. You still follow the core CONSORT checklist (CONSORT-AI was written against the 25 items of CONSORT 2010, which CONSORT 2025 has since replaced with 30 items), then you answer 14 extra questions about the AI itself, such as which version you used, what data went in, what came out, and how clinicians acted on it. If you cannot answer an item, say so and say why.
What the CONSORT-AI Checklist Is and How It Was Built
CONSORT-AI stands for Consolidated Standards of Reporting Trials, Artificial Intelligence extension. Liu and colleagues published it in September 2020 in BMJ, Nature Medicine and The Lancet Digital Health, alongside its protocol companion SPIRIT-AI [1][2]. CONSORT 2010 already gives a 25-item checklist for every randomised trial. CONSORT 2025 has since replaced it with a 30-item checklist. The CONSORT-AI items were written against CONSORT 2010, so their numbering follows that version. The CONSORT-AI checklist keeps all 25 and adds 14 items for what is new about an AI intervention. The authors classify 11 of the 14 as extensions and three as elaborations of existing CONSORT items [1].
The group followed the EQUATOR Network framework. A literature review and expert consultation produced 29 candidate items, of which 26 applied to both guidelines and three applied only to trial reports. A two-stage Delphi survey drew 103 responses in round one and 91 in round two, a two-day consensus meeting in January 2020 brought together 31 stakeholders, and a pilot with 34 participants refined the wording. An item needed 80% agreement to enter the checklist. Item 5(iii), on poor quality input data, scored 77% on its first wording and passed at 97% after rewording [1]. For an author, that history carries a message. The panel treated bad input data as important enough to rework rather than drop.
Scope matters before you open the checklist. The CONSORT-AI checklist is for trial reports. It is not meant for studies that develop or validate a prediction or diagnostic model, and the guideline authors pointed those studies to TRIPOD-ML and STARD-AI, both still under development when they wrote [1]. Our guide to STARD versus TRIPOD helps with that choice. The extension also leaves out continuously learning systems, which the consensus group judged too early in their healthcare use to cover [1].
When the CONSORT-AI Checklist Applies to Your Study
Apply a simple working rule. If the output of the AI changes what a clinician or patient does, the AI is part of the intervention, and the trial report needs the CONSORT-AI checklist. The table below maps common study types to the guideline that fits.
| Your study | Guideline to follow | Note |
|---|---|---|
| Randomised trial where an AI component is part of the intervention | CONSORT (2025, or 2010 where the journal still asks for it) plus the CONSORT-AI checklist | Answer every core item and the 14 AI items |
| Protocol for that kind of trial | SPIRIT 2013 plus SPIRIT-AI | Covered in the protocol section below |
| Developing or validating a prediction model | The TRIPOD family | See our prediction model reporting guide |
| Diagnostic accuracy study of an AI test | STARD | See our STARD versus TRIPOD comparison |
| Randomised trial with no AI component | CONSORT 2025 only | Start with our CONSORT guidelines guide |
Do not confuse this with AI disclosure rules. Journals also ask authors to state when they used generative AI to write or edit a manuscript, which is a separate policy question covered in our post on the AI watermark and disclosure policies. The CONSORT-AI checklist concerns AI that is being tested as a clinical intervention.
The 14 Items of the CONSORT-AI Checklist Explained
The table paraphrases each item from the guideline [1]. Item numbers follow CONSORT 2010, with Roman numerals marking the AI additions. Use the right-hand column as a quick scan for the most common way each item goes wrong.
| CONSORT-AI item | Section | What to report | Common trap |
|---|---|---|---|
| 1a,b (i) Elaboration | Title and abstract | Say that the intervention involves AI or machine learning, and name the model type. Keep the title broad and put the precise term in the abstract. | A title that says “digital tool” with no AI term, which weakens indexing. |
| 1a,b (ii) Elaboration | Title and abstract | State the intended use of the AI and the disease context. | Intended use shifts over time, so the abstract should fix it as of the trial. |
| 2a (i) Extension | Introduction | Explain where the AI sits in the clinical pathway, its purpose and its intended users. | Never saying whether the AI replaces, augments or adjudicates a clinician decision. |
| 4a (i) Elaboration | Methods, participants | Give inclusion and exclusion criteria at the participant level. | Mixing participant criteria with data criteria. |
| 4a (ii) Extension | Methods, participants | Give inclusion and exclusion criteria at the level of the input data, such as minimum image quality, with when and by whom this was judged. | Omitting what happened to participants whose input data were unusable. |
| 4b Extension | Methods, setting | Describe how the AI was integrated at the site, including hardware, software, cloud needs and any local fine-tuning. | Describing the algorithm but not the IT environment it ran in. |
| 5 (i) Extension | Methods, intervention | State the algorithm version, whether it matches earlier studies, and whether it changed during the trial. | Citing the original model paper with no version identifier. |
| 5 (ii) Extension | Methods, intervention | Describe how input data were acquired, selected and pre-processed. | Leaving pre-processing to a reference. |
| 5 (iii) Extension | Methods, intervention | Describe how poor quality or unavailable input data were assessed and handled, report the amount missing, and do the same for the comparator. | No figure for unusable inputs. |
| 5 (iv) Extension | Methods, intervention | Say whether humans interacted with the input data, what expertise was needed, and what training was given. | Not saying whether an error came from the user or the AI. |
| 5 (v) Extension | Methods, intervention | Specify the output, for example a probability, a class, an alarm, a recommended action or a closed-loop action. | Describing the model but not what the clinician actually saw. |
| 5 (vi) Extension | Methods, intervention | Explain how the output fed into decisions, including thresholds, pathways and any departure from the protocol. | No threshold for acting on a risk score. |
| 19 Extension | Results | Report any analysis of performance errors and how errors were identified. If none was planned or done, explain why not. | Silence, where a one-line explanation would do. |
| 25 Extension | Other information | State whether and how the AI and its code can be accessed, including licence and restrictions. | A data availability statement that covers datasets only. |
Items 5(i) to 5(vi) carry the intervention description
Six of the 14 items sit under CONSORT item 5, which describes the intervention. The guideline authors explain why. Health outcomes depend on how humans use the output, so the report must describe the downstream actions as well as the model [1]. Their example is a skin cancer system that returns a percentage. The report should state both the pathway, such as excision when the diagnosis is positive, and the threshold for entering it, such as a probability above 80%. The same description should be given for the comparator arm where relevant [1].
Input data exclusions after randomisation are analysis exclusions
Item 4a(ii) has a practical consequence for your flow diagram. Where input data are acquired after randomisation, the guideline treats any exclusion as an exclusion from the analysis rather than from enrolment [1]. That affects how you draw the participant flow and which population you analyse, so decide it alongside your intention-to-treat and per-protocol plan.
Item 19 rewards honesty
Item 19 is the one authors skip most readily. The guideline asks for performance error analysis and, if none was planned, a reason. The authors tie this to safety, noting that AI systems can fail in ways humans do not predict, and they encourage checking error rates across population subgroups [1]. A one-sentence explanation of why no error analysis was done is better than silence.
What Published AI Trials Still Leave Out of the CONSORT-AI Checklist
Three meta-research studies show where completeness drops. Martindale and colleagues assessed 65 randomised trials of AI interventions published after the guideline appeared. Median concordance with CONSORT-AI was 90% (interquartile range 77% to 94%), yet only 10 of the trials said they had used the guideline. Algorithm version, accessibility of the AI or its code, and reference to a protocol were consistently under-reported, and only 3 of 52 journals explicitly endorsed or mandated CONSORT-AI [4].
Zhong and colleagues examined 19 published trials and 21 protocols in primary care. Adequately reported items reached 65% (172 of 266) for trials and 68% (214 of 315) for protocols, with gaps in input data, performance error analysis and access to the AI or its code [5]. Chen and colleagues reviewed 57 oncology trials and found a median concordance of 82% with CONSORT 2010 and CONSORT-AI combined. Input data inclusion and exclusion, algorithm version, low quality data handling, performance error assessment and data accessibility were all under-reported. Trials at serious risk of bias (57%) were less concordant than trials at moderate (71%) or low (84%) risk [6].
The pattern is consistent across all three. The items an author finds easiest are the ones that resemble ordinary trial reporting. The items that go missing are the AI-specific ones in the table above. If you check only those items before submission, you cover the gaps reviewers keep finding. Reviewers who score trials with a tool such as RoB 2 will also read the intervention description closely, so clear reporting helps with risk of bias judgements as well.
CONSORT-AI Checklist Worked Example on an Open-Access AI Trial
To show what the items look like in practice, the table below applies the CONSORT-AI checklist to a published trial. Tsai and colleagues ran a pragmatic randomised trial at an academic medical center in Taiwan, testing an electrocardiogram-based AI alert for low ejection fraction among 13,631 inpatients under non-cardiologist care. The article states that the trial was registered (NCT05117970) and “followed CONSORT-AI Extension checklist guidelines” [3]. It is published in BMC Medicine under a CC BY 4.0 licence, confirmed from the Crossref version-of-record licence field, the Europe PMC licence field and the article XML.
The main result was a higher rate of newly diagnosed low ejection fraction at 30 days in the alert arm, 1.5% versus 1.1% (hazard ratio 1.50, 95% CI 1.11 to 2.03), with no overall change in echocardiogram use [3]. This post is not judging the clinical result. It is reading the article against the 14 reporting items, using the main text only.
| CONSORT-AI item | Where to look | What the article reports | Verdict |
|---|---|---|---|
| 1a,b (i) | Title; Abstract | The title names AI assistance and electrocardiograms. The abstract calls the tool an “electrocardiogram (ECG)-based artificial intelligence (AI)-assisted clinical decision support tool” but does not name the model type. Deep learning appears in the Background and Methods. | Judgement call Whether “AI-assisted” is enough for the model type is a matter of reading. |
| 1a,b (ii) | Abstract, Background | The aim is to improve early diagnosis of low ejection fraction among inpatients under non-cardiologist care. That gives the purpose and the disease context. | Reported |
| 2a (i) | Background, final two paragraphs | Sets out limited ECG interpretation skill among general practitioners and positions the algorithm as help for generalists and non-cardiology departments. The intended users are non-cardiologist physicians. | Reported |
| 4a (i) | Methods, Trial design; Figure 1 | Inpatients with at least one ECG between November 2022 and May 2023 under non-cardiologist care. Patients under 18 years or with LVEF of 50% or less in the previous 180 days were excluded, with counts per arm in Figure 1. | Reported |
| 4a (ii) | Methods, Trial design | The Methods list six routine ECG indications and say these were not recorded as data points. The text gives no minimum signal quality requirement for the ECG input. | Not located in the article |
| 4b | Methods, AI-ECG model | The hospital information system generated an on-screen alert in real time. Hardware, software and local adaptation needs are not described. The trial ran at one centre, so multi-site integration did not arise. | Judgement call |
| 5 (i) | Methods, AI-ECG model | The trial used “a previously developed deep learning model” and cites an earlier publication for architecture, training and validation. No version identifier appears in the text. | Judgement call Reference to the earlier paper may be intended to stand for the version. |
| 5 (ii) | Methods; Data collection process | Standard 12-lead ECGs from routine inpatient care, with index time rules explained for high-risk and low-risk patients. Pre-processing before the model is not described in this article. | Judgement call |
| 5 (iii) | Methods; Results | The text has no statement on how noisy, missing or unusable ECGs were handled in either arm. | Not located in the article |
| 5 (iv) | Methods, AI-ECG model | The system flags results automatically and no step is described where a person selects or edits the ECG input. I read this as no human-AI interaction at the input stage, but the text does not say so. | Judgement call |
| 5 (v) | Methods, AI-ECG model | The model gives a continuous probability that EF is 50% or lower. A threshold chosen by ROC analysis splits patients into high and low risk, with internal validation sensitivity of 72.4% and specificity of 89.1% at that threshold. | Reported |
| 5 (vi) | Methods; Figure 1; Results, Post-hoc analysis | “Clinicians could then decide whether to pursue additional diagnostic evaluations (such as echocardiogram) based on both the AI classification and their clinical judgment.” The control arm received no alert. Figure 1 shows the high-risk alarm as a short message and the low-risk report available in the record, while the text calls it an on-screen alert. Both can be true. | Reported |
| 19 | Results; Discussion | The article reports positive predictive value for low EF among patients who had an echocardiogram (34.2% versus 20.2%) and cites internal validation figures. The text has no analysis of false positive or false negative cases and no statement on why none was done. | Judgement call |
| 25 | Declarations | The Data availability statement reads “No datasets were generated or analysed during the current study.” It covers datasets only. The text has no statement on access to the model or its code. | Not located in the article |
Five rows are reported, six are judgement calls and three could not be located in the main text. Rows marked “Not located in the article” mean the information was not found in the article body, and the supplementary file was not reviewed. Rows marked “Judgement call” are ones where a reasonable reader could score the item either way.
What the worked example shows
The trial does what many do not. It states that it followed the CONSORT-AI checklist, which only 10 of 65 trials did in the Martindale review [4]. Even so, the unanswered rows are the same ones the meta-research flagged, namely version, input data quality, performance errors and access to the AI. That is the practical lesson. An author who knows those four weak spots can close them in an hour of drafting, usually with one sentence each.
A clear answer for each missing row is short. For item 5(i), name the model build and state whether it is the build validated earlier. For 5(iii), report how many ECGs were unusable and what happened to those patients in each arm. For 19, report any false positive review, or state that none was planned and why. For 25, say whether the model is proprietary, available on request or open, and under what licence.
Beyond the 14 items, the core CONSORT items still apply
The CONSORT-AI checklist does not replace CONSORT 2010. The same article shows two core items at work. The abstract lists positive predictive value and cardiology consultation rates among the secondary outcomes, while the Methods list echocardiogram use and 30-day all-cause mortality, and say “we have designated these analyses as post-hoc” for the consultation analysis. That is a Judgement call, because the authors do flag the post-hoc status, but a reader would want the outcome list to match between abstract and Methods. The authors also disclose that they “updated our trial registry in October 2024” after changing the primary outcome and target sample size, which is the kind of change CONSORT 2010 asks authors to report with reasons.
SPIRIT-AI Checklist as the Protocol Companion
SPIRIT-AI covers the protocol stage of the same trials. It was developed in parallel with CONSORT-AI, and the SPIRIT-AI extension adds 15 new items to the core SPIRIT 2013 checklist [2]. Most items mirror CONSORT-AI in future tense. The protocol states which algorithm version will be used, how input data will be acquired, what the output will be, how outputs will feed decisions, how performance errors will be analysed, and how the AI and its code can be accessed [2]. One protocol item has no direct partner in the trial report list. SPIRIT-AI 6a(ii) asks authors to describe pre-existing evidence for the AI intervention [2]. For a fuller walk-through of the protocol companion, see our SPIRIT-AI guide.
| Stage | Guideline pair | AI-specific items | Practical point |
|---|---|---|---|
| Protocol | SPIRIT 2013 plus SPIRIT-AI | 15 new items | Fix the algorithm version and the action thresholds before enrolment |
| Trial report | CONSORT 2010 plus CONSORT-AI | 14 new items | Report any change from the protocol and the reason |
The two lists work best as a pair. Writing item 5(i) and 5(vi) answers into the protocol makes the later report an exercise in confirming what happened. It also makes departures visible, which the guideline asks you to report for decision-making that differed from the plan [1]. If you are still designing the study, our clinical trial design consultancy covers endpoints, sample size and protocol writing support.
How to Complete a CONSORT-AI Checklist Before Submission
- Download the current CONSORT checklist (CONSORT 2025, or CONSORT 2010 if your journal still asks for it) and add the 14 CONSORT-AI rows beneath the matching core items.
- Write the answer to each row as a page or line pointer in your own manuscript, so an editor can verify it in seconds.
- For any row that does not apply, write a short reason in the row instead of leaving it blank.
- Draw the participant flow with input data exclusions shown separately from participant exclusions, using a generator such as Flowchart Pro below.
- State in the Methods that you followed the CONSORT-AI checklist. Only 10 of 65 trials did in the Martindale review [4].
- Read the journal author instructions. Only 3 of 52 journals in the same review endorsed or mandated CONSORT-AI, so a journal may accept it without naming it [4].
Common Mistakes When Reporting an AI Trial
- Describing the model but not the workflow. Items 5(v) and 5(vi) ask what the clinician saw and what they did with it.
- Citing the model paper in place of a version. Item 5(i) wants the build used in this trial and whether it changed.
- Treating the CONSORT-AI checklist as a replacement. The core CONSORT items still apply in full.
- Skipping item 19. A short reason for no error analysis is acceptable. Silence is not.
- Using the wrong guideline. A model development paper needs TRIPOD, and a diagnostic accuracy study needs STARD.
Frequently Asked Questions About the CONSORT-AI Checklist
How many items does the CONSORT-AI checklist add?
It adds 14 items to the core CONSORT checklist (written against the 25 items of CONSORT 2010), made up of 11 extensions and three elaborations [1].
Is the CONSORT-AI checklist the same as SPIRIT-AI?
No. The CONSORT-AI checklist is for the completed trial report. SPIRIT-AI is for the trial protocol and adds 15 new items to SPIRIT 2013 [2].
Does CONSORT-AI apply if I only used AI to help write the manuscript?
No. The CONSORT-AI checklist applies when an AI component is part of the trial intervention. Use of AI to draft text falls under journal disclosure policies instead.
Can I use the CONSORT-AI checklist for a model validation study?
Not as the main guideline. The authors directed development and validation studies to TRIPOD-ML and STARD-AI, which were under development when the extension appeared [1].
Free Research Tools from Editverse
- GraphPro: build publication-ready charts, 25 types, with built-in t-test, ANOVA, correlation, and chi-square p-values. Free, no signup.
- StatPro: free statistical calculators and a test-selector tool covering group comparisons, categorical outcomes, time-to-event analysis, and diagnostic accuracy.
- Flowchart Pro: generate PRISMA 2020, PRISMA-ScR, CONSORT, STROBE, STARD, MOOSE, and ARRIVE flow diagrams directly in your browser.
Professional Support for Your AI Trial Manuscript
The AI-specific items sit mostly in the Methods and the abstract, which is where reviewers look first. Editverse supports authors at both ends. Our work follows COPE, ICMJE and GPP 2022 guidance, and subject-matter experts review your reporting against the checklist.
Before you submit
$150
Pre-submission peer review simulating a journal referee, with a detailed reviewer report and follow-up consultation. 7 to 10 business days.
Methods section
$225
A complete, reproducible Methods section written from your protocols and notes, with missing details listed for you to fill.
Not sure which fits your trial report? Send us your manuscript and the journal you are targeting and we will recommend a starting point.
Conclusion
The CONSORT-AI checklist asks 14 questions that ordinary trial reporting never raised, and meta-research shows the same four areas going unanswered again and again. Those areas are algorithm version, input data quality, performance errors and access to the AI. Answer them with one clear sentence each, keep the core CONSORT items intact, state that you followed the extension, and pair the trial report with a SPIRIT-AI protocol where you can. A manuscript built that way is easier for editors to check and easier for readers to trust.
References
- Liu X, Cruz Rivera S, Moher D, Calvert MJ, Denniston AK; SPIRIT-AI and CONSORT-AI Working Group. Reporting guidelines for clinical trial reports for interventions involving artificial intelligence: the CONSORT-AI extension. BMJ. 2020;370:m3164. doi:10.1136/bmj.m3164. PMID 32909959. Also published in Nat Med 2020;26(9):1364-1374 and Lancet Digit Health 2020;2(10):e537-e548.
- Cruz Rivera S, Liu X, Chan AW, Denniston AK, Calvert MJ; SPIRIT-AI and CONSORT-AI Working Group. Guidelines for clinical trial protocols for interventions involving artificial intelligence: the SPIRIT-AI extension. BMJ. 2020;370:m3210. doi:10.1136/bmj.m3210. PMID 32907797.
- Tsai DJ, Lin C, Liu WT, Lee CC, Chang CH, Lin WY, et al. Artificial intelligence-assisted diagnosis and prognostication in low ejection fraction using electrocardiograms in inpatient department: a pragmatic randomized controlled trial. BMC Med. 2025;23:342. doi:10.1186/s12916-025-04190-z. PMID 40484925. PMCID PMC12147261. Registered at ClinicalTrials.gov NCT05117970.
- Martindale APL, Llewellyn CD, de Visser RO, Ng B, Ngai V, Kale AU, et al. Concordance of randomised controlled trials for artificial intelligence interventions with the CONSORT-AI reporting guidelines. Nat Commun. 2024;15:1619. doi:10.1038/s41467-024-45355-3. PMID 38388497.
- Zhong J, Zhu T, Huang Y. Reporting quality of AI intervention in randomized controlled trials in primary care: systematic review and meta-epidemiological study. J Med Internet Res. 2025;27:e56774. doi:10.2196/56774. PMID 39998876.
- Chen D, Arnold K, Sukhdeo R, Farag Alla J, Raman S. Concordance with CONSORT-AI guidelines in reporting of randomised controlled trials investigating artificial intelligence in oncology: a systematic review. BMJ Oncol. 2025;4(1):e000733. doi:10.1136/bmjonc-2025-000733. PMID 40923022.