Blood's AI Policy: ASH Rules for Hematology's Flagship Journal

ASH publishes two main journals: Blood and Blood Advances. The AI policy text is identical across both, but the editorial context differs:

Blood's lower acceptance rate means that every submission receives intense scrutiny. An AI disclosure issue that might be handled with a quick email at a less selective journal becomes a more formal concern at Blood. The editors aren't looking to catch authors out, they're protecting the journal's integrity because a retraction at Blood carries weight across the entire hematology field.

Blood Advances was launched as a companion journal to handle the volume of quality submissions that Blood can't accommodate. It follows the same ASH rules, but the editorial tone tends to be slightly more flexible on formatting details while maintaining the same ethical standards.

Genomics and bioinformatics

Modern hematology research is deeply computational. Whole-genome sequencing for leukemia characterization, RNA-seq for gene expression profiling, single-cell transcriptomics for clonal architecture mapping, these methods involve extensive code and data analysis pipelines.

If you used GitHub Copilot or ChatGPT to help write Python scripts for your bioinformatics pipeline, that's AI-assisted code generation and it requires disclosure. The analysis tools themselves, GATK, STAR, Seurat, CellRanger, are standard bioinformatics software, not generative AI. They go in regular Methods. But AI assistance in writing the code that runs these tools is a different category.

Disclosure example for a genomics paper:

"Whole-genome sequencing analysis was performed using GATK (v4.4) and Mutect2 for somatic variant calling, as described in Methods. GitHub Copilot (Microsoft) was used to assist with writing custom Python scripts for variant filtering and annotation. All scripts were validated against published benchmark datasets. ChatGPT (GPT-4, OpenAI) was used to improve the readability of the Results and Discussion sections. The authors take full responsibility for the published content."

Flow cytometry and immunophenotyping

Blood publishes extensive flow cytometry data, immunophenotyping panels for leukemia diagnosis, MRD detection, immune subset quantification. Some modern flow cytometry software incorporates machine learning for automated gating. This is a research tool, not a writing tool, and belongs in standard Methods.

But if you used AI to help interpret complex flow cytometry patterns or to draft descriptions of immunophenotypic findings, that crosses into the disclosure territory. The distinction matters because flow cytometry interpretation directly affects diagnostic conclusions.

Clinical trial reports

Blood publishes practice-changing clinical trials, new treatments for acute myeloid leukemia, novel therapies for sickle cell disease, CAR-T cell therapy results, transplant conditioning regimen comparisons. These papers determine treatment protocols at cancer centers worldwide.

For clinical trial reports, keep AI away from efficacy conclusions, safety summaries, and treatment recommendations. You can use AI to polish your language, but the clinical interpretations must be entirely human-generated. Blood's clinical reviewers are hematologists who'll notice if AI-generated text introduces the kind of hedged, noncommittal phrasing that large language models tend to produce when discussing clinical outcomes.

Here's what proper disclosure looks like for different Blood submission types.

For an original research article (e.g., AML genomics study):

"During the preparation of this manuscript, the authors used Claude (Claude 3.5, Anthropic) to improve the clarity and readability of the Introduction and Discussion sections. All AI-generated suggestions were reviewed and edited by the corresponding author (J.K.) and the senior computational biologist (L.M.). No AI tools were used in study design, sequencing analysis, variant interpretation, or clinical correlations. The authors take full responsibility for the content of this article."

For a clinical trial report (e.g., Phase III leukemia trial):

"The authors used ChatGPT (GPT-4o, OpenAI) to improve the English language of the Methods section. No AI tools were used for trial design, data analysis, efficacy assessment, safety evaluation, or interpretation of clinical outcomes. All clinical conclusions were drawn by the study investigators based on the pre-specified statistical analysis plan. The authors take full responsibility for the published content."

For a review article:

"During the preparation of this review, the authors used ChatGPT (GPT-4, OpenAI) to improve the readability of selected sections and to check the consistency of reference citations. The literature search, study selection, data synthesis, and all scientific conclusions were performed entirely by the authors. The authors take full responsibility for the published content."

For a brief report / Blood Advances submission:

"Claude (Claude 3.5, Anthropic) was used to improve the language of the manuscript. All content was reviewed and approved by all authors, who take full responsibility for the published work."

Each of these follows the same structure: name the tool, describe what was and wasn't done, confirm author responsibility. Shorter papers can have shorter disclosures, but the essential elements remain the same.

Blood follows COPE guidelines for handling integrity concerns, and undisclosed AI use falls squarely within that framework. Here's the escalation path:

During peer review. If a reviewer flags text that appears AI-generated, or if the editor suspects undisclosed AI use, you'll receive an email asking for clarification. This isn't automatically a rejection, but it does raise a red flag. Editors may use AI detection tools as a preliminary check, though everyone in the field acknowledges these tools have significant false positive rates.

After conditional acceptance. If AI use surfaces during the revision or production phase, the paper can be held until the issue is resolved. You'll need to add a proper disclosure statement and explain the omission. This delays publication and creates a record that follows the manuscript.

After publication. This is where things get serious:

• Correction: If the AI use was limited to language editing and didn't affect scientific content, a published correction adding the disclosure may suffice. This is the best-case scenario for post-publication discovery.
• Expression of concern: If the scope of AI use raises questions about the paper's scientific validity, for example, if AI was used to draft interpretive conclusions, the editor may publish an expression of concern while the matter is investigated.
• Retraction: If AI involvement was extensive enough to undermine confidence in the scientific content, retraction becomes possible. This is the nuclear option and it's rare, but it's on the table.
• Institutional notification: In serious cases, Blood may notify the authors' institution. For faculty at academic medical centers, this triggers a formal research integrity investigation, a process that can take six months to a year and has career implications far beyond one paper.

The practical reality: most undisclosed AI use gets caught during peer review or shortly after. Reviewers in hematology are increasingly attuned to AI-generated prose patterns, and Blood's reviewer pool includes some of the sharpest editors in the field. It's much better to disclose upfront than to explain after the fact.

Several things are worth noting in this comparison:

Blood and JCO overlap significantly in clinical hematology-oncology content. Both publish leukemia trials, lymphoma studies, and transplant outcomes. JCO's policy comes from ASCO, Blood's from ASH, but the substance is nearly identical. If you're deciding between the two for a clinical paper, the AI policy won't be the differentiating factor.

Leukemia follows Springer Nature's policy, which is generally consistent with what Blood requires but comes from a commercial publisher rather than a medical society. Springer Nature's framework covers thousands of journals, so the language tends to be more generic. Blood's ASH-derived policy has a more clinical tone.

Haematologica is published by the European Hematology Association through Ferrata Storti Foundation. The EHA's AI policy aligns with European standards and COPE guidelines. It's broadly equivalent to ASH's approach.

All five journals agree on the fundamentals: no AI authorship, mandatory disclosure, no AI-generated images, and full author responsibility. The differences are in implementation details, not principles.

Keep a running log

Start tracking AI use from the moment you begin drafting. A simple text file noting "Used ChatGPT for language editing, Introduction, March 5" gives you the raw material for an accurate disclosure statement. Reconstructing your AI use from memory six months later when you're doing final revisions isn't reliable.

Coordinate with co-authors

Multi-center hematology studies can have 15 to 30 co-authors across multiple institutions. Your co-first-author in Munich might be using DeepL Write for translation assistance. Your biostatistician in Boston might be using GitHub Copilot. The corresponding author needs to ask every co-author about their AI tool use before assembling the final disclosure.

A simple email to all co-authors asking "Did you use any AI tools during your work on this manuscript?" should be part of your pre-submission workflow. Don't assume everyone will volunteer this information unprompted.

Don't over-disclose

Blood doesn't need to know that you used Google Translate to look up a word, or that your word processor's built-in grammar checker flagged a comma splice. The disclosure requirement covers generative AI tools, systems that can produce new text, substantially rephrase content, or generate code. Standard reference management software, statistical packages, and basic writing tools don't require disclosure.

Watch the clinical content

If your paper includes treatment recommendations, prognostic assessments, or diagnostic criteria, and many Blood papers do, make absolutely sure these sections were written by the clinical investigators. AI-assisted language editing of clinical conclusions is acceptable if the clinical reasoning came from the authors. AI-generated clinical reasoning is not.