Preregistration: Would US Adults on an Online Panel Donate Tissue for AI Training? A Four-Version Split-Ballot Survey

What this is A preregistration of a survey, published before any data were collected. It fixes the questions, hypotheses, exclusions, analysis code and decision rules in public before the pilot launches. Survey privacy notice: privacy-notice.pdf. This is the notice the survey introduction links to; the same text is part 4 of preregistration.pdf. Purpose Some consent forms for medical research, such as the Human Cell Atlas template, say donated tissue data may be used "for any purpose". The survey asks US adults whether they would agree to donate a small sample of tissue or cells for medical research when told the data could be used to train artificial intelligence (AI) models, by whom, and with a one-year period of first use by companies. Design A four-version split ballot. Each respondent is assigned at random to one description of how the data could be used: by university and hospital researchers by university and hospital researchers to train AI models by companies to train AI models by companies to train AI models, with only those companies able to use the data for the first year, after which it is made available to the public The question that follows is word for word the same in all four versions. An attention check comes first; four follow-up questions and four background questions come after. 10 questions in all. Sample US adults 18 and over in the Pollfish online panel (panel vendor: Pollfish LLC). An opt-in panel, not a probability sample. Planned: a 40-complete pilot, never pooled; then 1,000 completes, about 250 per version. Budget Hard cap of $2,000 for pilot and main survey together, paid by SuperTruth Inc. If the price is higher than planned, the survey buys fewer completes. Preregistered contrasts Outcome: the share answering "Definitely would" or "Probably would". H1 (primary): version 4 lower than version 1. H2: version 3 lower than version 1. H3: version 4 lower than version 3. H2 and H3 are tested only if H1 is a finding, with Holm's correction. Version 2 against 1 and version 3 against 2 are exploratory, with intervals only. Where a comparison changes more than one thing at once, every result names all of the changes. Results Reported whichever way they fall, using the sentences written in advance for each outcome (preregistration.pdf, part 1, section 10), including a result against our hypotheses. Privacy The authors receive no names, contact details or device IDs. No question asks about health. The public response file holds only the version seen, answers to the first six questions and the exclusion reason, with no ID. The raw export stays on one SuperTruth computer and is deleted within 90 days of the report and public file being published, and by 31 December 2027 at the latest. Ethics No IRB review. In the authors' judgment this is a minimal-risk opinion survey with no federal funding. Funding and conflicts SuperTruth Inc. pays for the fieldwork. SuperTruth develops data-trust technology; a finding that people tell research uses apart would support that business. The authors wrote the questions and the directional hypotheses. MedSync supplies no money, staff, data or facilities for the survey. Files preregistration.pdf: the full preregistration. privacy-notice.pdf: the survey privacy notice. analyze.py, test_analyze.py, make_synthetic.py: analysis code and tests. README.txt: file list with SHA-256 checksums. Companion paper "Data Truth Is AI Truth: Provenance and Consent in the Data Behind the $1.8 Billion Virtual Cell" (in preparation).