CONSORT 2025 and SPIRIT 2025: Open Science yes, Artificial Intelligence no
In April 2025, CONSORT and SPIRIT published their first coordinated update since 2010 and 2013. Major advance: Open Science is now integrated into the base standards. But artificial intelligence? Absent. Despite CONSORT-AI existing since 2020, the new guidelines mention neither AI, nor existing extensions, nor algorithmic transparency issues.
Source analysée
https://www.bmj.com/content/389/bmj-2025-082647A historic publication
On 14 April 2025, CONSORT 2025 was published simultaneously in JAMA, The Lancet, The BMJ, Nature Medicine and PLoS Medicine. Two weeks later, on 28 April, SPIRIT 2025 followed the same path. It was the first time that the two complementary guidelines — CONSORT for reporting randomised controlled trials (RCTs), SPIRIT for writing trial protocols — were updated in a coordinated manner and published in the five most influential medical journals in the world.
The institutional message was clear: these standards are foundational for clinical research.
CONSORT 2025 replaces CONSORT 2010 with a revised checklist of 30 items (7 new, 3 revised, 1 removed). SPIRIT 2025 replaces SPIRIT 2013 with 34 items (2 new, 5 revised, 5 removed). The development process involved an international survey of 317 participants followed by a 30-expert consensus meeting over two days.
The advances are significant: a new Open Science section (pre-registration, data sharing, analysis code transparency), strengthened patient/public involvement in study design, and detailed assessment of harms.
“CONSORT 2025 is the first major reporting guideline to formally integrate the principles of Open Science, recognising that transparency is essential to the reproducibility and verifiability of clinical trial results.”
— Hopewell et al. (2025), CONSORT 2025 Statement, BMJ
A methodological advance long awaited since the reproducibility crisis. But there is a blind spot: artificial intelligence.
The documented omission
Neither CONSORT 2025 nor SPIRIT 2025 contains a single mention of artificial intelligence. The guidelines do not recommend disclosing the use of machine learning algorithms in the design, conduct or analysis of trials. They reference neither CONSORT-AI nor SPIRIT-AI, despite their publication in 2020 in Nature Medicine, The BMJ and The Lancet Digital Health.
This omission is documented in a dedicated commentary published in The Lancet in 2025: “AI in clinical trials is missing from CONSORT and SPIRIT 2025 guidelines”. The authors highlight:
“Neither CONSORT 2025 nor SPIRIT 2025 recommends that researchers disclose AI involvement in any component of a clinical trial. The use of proprietary and non-transparent algorithms further exacerbates concerns about bias and reproducibility.”
A randomised clinical trial using an LLM for patient screening, deep learning image analysis, or a therapeutic chatbot could therefore fully comply with CONSORT 2025 without ever mentioning the use of AI.
What CONSORT 2025 integrates (and does not)
Here are the major advances of CONSORT 2025, and why they do not cover AI:
Open Science section (new)
Pre-registration of studies, sharing of anonymised data, statistical analysis code transparency. But: if the analysis code uses a proprietary AI model (GPT-4, Claude, Gemini), this transparency is illusory — the algorithm remains a black box.
Item 1b - Patient/public involvement (new)
Document patient participation in study design. But: no mention of participation in the design of AI interventions — who decides on the prompts, the safeguards, the chatbot’s tone?
Item 27 - Data sharing (new)
Declare whether data will be shared and how. But: if the intervention is a chatbot, sharing transcripts raises consent and privacy concerns. CONSORT 2025 specifies nothing about conversational data.
No mention of AI
No reference to CONSORT-AI/SPIRIT-AI extensions. No requirement to document algorithm use. No mention of algorithmic bias or human-AI interaction. A trial could use GPT-4 for screening without ever disclosing it.
The paradox of normative fragmentation
CONSORT and SPIRIT are not monoliths. They are fragmented ecosystems: there are now 32 CONSORT extensions and 10 SPIRIT extensions covering specific contexts (acupuncture, pragmatic trials, pilot trials, Chinese herbal medicine, social and psychological interventions, etc.).
CONSORT-AI is one of them. Published in 2020 with a rigorous Delphi process (103 stakeholders, 31 experts in consensus), it defines 14 AI-specific items: algorithm version, code accessibility, error management, human-AI interaction.
The problem: adoption is voluntary and peripheral. As documented in our analysis of CONSORT-AI, only 3 out of 52 journals (6%) mention CONSORT-AI in their author guidelines. Median adherence is 90% on generic items, but drops to 20% for algorithm version and 42% for code accessibility — the two most critical items for reproducibility.
Worse still: adherence is declining. A study across 15 oncology journals shows a drop from 96% (2021) to 79% (2023).
The illusion of the overall score: A trial can achieve a 90% CONSORT-AI score while omitting the only two items that would allow the study to be reproduced — the exact algorithm version and code accessibility.
Why such low adoption? Because CONSORT-AI is an optional extension in an already complex ecosystem. Researchers consult CONSORT 2010 (then CONSORT 2025), journals cite CONSORT in their guidelines, and most don’t even know CONSORT-AI exists.
CONSORT 2025 could have changed this. A simple mention in the base items (“If the intervention includes an AI component, see CONSORT-AI”) would have been enough to make the extension visible. But this bridge was not created.
Comparison: 5 frameworks for evaluating AI in healthcare
This article series has analysed five methodological frameworks for the evaluation of AI interventions in healthcare:
| Framework | Objective | Key items | Adoption | Status 2026 |
|---|---|---|---|---|
| Hua (2022) | Evidence level classification for AI in mental health | T1 (proof-of-concept), T2 (clinical benchmark), T3 (ecological RCT) | Low (1 academic citation) | Theoretical |
| Choudhury (2024) | Ecological validity of LLM health studies | Transparency, Context-Centered Design, real patient pathways | Academic (12 citations) | Conceptual |
| CHART (2024) | Reporting for health chatbot evaluations | 21 transparency items (design, data, evaluation) | Very low (2 citations) | Proposed |
| CONSORT-AI (2020) | Reporting for AI clinical trials | 14 AI items (algorithm version 20%, code 42% adherence) | Low (3/52 journals, declining 96%→79%) | Declining |
| CONSORT/SPIRIT 2025 | Reporting/protocols for clinical trials (all types) | 30/34 items including Open Science — 0 AI items | Universal (top 5 journals, massive adoption expected) | Base standard |
The finding: The first four frameworks finely analyse AI issues (evidence levels, ecological validity, chatbot transparency, algorithmic reporting). But they remain peripheral — low academic citations, voluntary adoption, invisible in journal guidelines.
CONSORT/SPIRIT 2025, on the other hand, will be massively adopted as it was published in the 5 most influential journals and integrated into global editorial requirements. But it contains no mention of AI.
The result: the most widely adopted framework is the one that ignores AI. The frameworks that address AI remain confidential.
What this reveals: the epistemic double standard
When a pharmacological intervention is tested, CONSORT 2025 will now require protocol transparency, data sharing, and analysis code accessibility. Open Science becomes a baseline standard.
When an intervention involves AI — whether a diagnostic algorithm, a therapeutic chatbot, or a risk prediction system — none of these requirements apply automatically. The algorithm version? Optional (CONSORT-AI item 5i, adopted at 20%). Code accessibility? Optional (item 25, adopted at 42%). Documentation of training biases? Not covered. Analysis of model errors? CONSORT-AI recommends it (item 19), but it is an extension nobody consults.
Epistemic double standard: Traditional interventions are subject to increasing demands for transparency and reproducibility. AI interventions — though more opaque, more dynamic, and with less comprehensible mechanisms — escape these requirements because they fall under optional extensions not referenced by the base standards.
This is not an oversight. It is a normative architecture that treats AI as a special case while it is becoming ubiquitous in clinical research.
Terminological clarification
CONSORT (Consolidated Standards of Reporting Trials) and SPIRIT (Standard Protocol Items: Recommendations for Interventional Trials) are two complementary but distinct guidelines:
-
CONSORT governs the reporting of results from completed randomised clinical trials. It says: “Here is how you should write your scientific article so that readers can understand what you did and assess the validity of your conclusions.”
-
SPIRIT governs the writing of trial protocols before they begin. It says: “Here is what your protocol must contain so that ethics committees, funders and peers can evaluate the methodological rigour of your project.”
Both have AI extensions:
- CONSORT-AI (2020): 14 items for reporting trials involving AI
- SPIRIT-AI (2020): items for protocols of trials involving AI
The 2025 updates mention neither. A researcher consulting CONSORT 2025 or SPIRIT 2025 to prepare a trial on a therapeutic chatbot will have no indication that they should also consult CONSORT-AI or SPIRIT-AI.
What a clinician should take away
When you read a clinical trial published after April 2025:
-
Check whether CONSORT 2025 is cited — If so, it’s a good baseline methodological sign (Open Science, patient involvement, statistical transparency).
-
Check whether the intervention includes an AI component — Chatbot, diagnostic algorithm, prediction system, clinical decision support.
-
Look for CONSORT-AI in the references — If absent while the intervention involves AI, the critical items (algorithm version, accessible code, error analysis, human-AI interaction) are probably omitted.
-
Identify the blind spots:
- Is the exact model version specified? (e.g. “GPT-4 turbo 2024-04-09” vs “GPT-4”)
- Is the code accessible? Or only “available upon request”?
- Are the model’s errors analysed? Or only average performance?
- Is the human-AI interaction described? Who has the final clinical say?
A trial can be CONSORT 2025-compliant AND methodologically opaque on AI. The two standards are not contradictory, they are orthogonal. One covers general transparency, the other algorithmic transparency. But only the former is required by journals.
Sources and references
-
Hopewell S et al. (2025). CONSORT 2025 Statement: Updated guideline for reporting randomised trials. JAMA, The Lancet, BMJ, Nature Medicine, PLoS Medicine. https://www.bmj.com/content/389/bmj-2025-082647
-
Chan AW et al. (2025). SPIRIT 2025 Statement: Updated guideline for protocols of randomised trials. JAMA, The Lancet, BMJ, Nature Medicine, PLoS Medicine.
-
Lancet Commentary (2025). AI in clinical trials is missing from CONSORT and SPIRIT 2025 guidelines. The Lancet. https://www.thelancet.com/journals/lancet/article/PIIS0140-6736(25)01268-1/fulltext
-
Liu X et al. (2020). Reporting guidelines for clinical trial reports for interventions involving artificial intelligence: the CONSORT-AI extension. Nature Medicine, BMJ, The Lancet Digital Health. https://www.nature.com/articles/s41591-020-1034-x
-
Cruz Rivera S et al. (2024). Concordance of Artificial Intelligence-Specific Extensions to the CONSORT Statement With Published AI Intervention Trials: The AI-CONSORT Concordance Study. JAMA Network Open.
Series: AI Evaluation Frameworks in Healthcare
- Hua: three tiers of evidence for AI in mental health
- Choudhury: ecological validity of LLM studies
- CHART: health chatbot transparency
- CONSORT-AI: transparency of AI clinical trials
- CONSORT/SPIRIT 2025: Open Science yes, AI no (this article)
- PROBAST+AI: quality of AI prediction models
Mots-clés