Transparency in artificial intelligence has become a question of provenance. Provenance tells you where a piece of content came from; it does not tell you why what it asserts was decided.
The Apply AI Strategy names healthcare and pharmaceuticals among its priority sectors, and the AI in Science Strategy addresses AI-driven research directly. This contribution concerns a gap that sits between the two, and that current transparency instruments do not close.
What provenance does and does not do. Content Credentials (C2PA) and watermarking establish where an artefact originates. They are necessary and they work. But they answer a different question from the one that matters for scientific and clinical accountability: not where a conclusion came from, but why it was preferred over the alternatives. In its 31 July 2026 statement on responsible AI in Europe, OpenAI describes provenance applied to images and being extended to audio, while for text — the material science is made of — the commitment is stated in the future tense, as standards and tools mature.
The stated reason is not necessarily the operative one. A team at Anthropic inserted hints capable of altering model answers, isolated the cases where the answer moved in the direction the hint indicated, and measured whether the chain of thought acknowledged using it. Overall faithfulness was 25% for Claude 3.7 Sonnet and 39% for DeepSeek R1, falling to 20% and 29% for hints pointing towards misaligned behaviour. The authors note their measure is likely an upper bound.
Scientific record-keeping is reproducing the same gap. In the largest expert assessment of machine-generated hypotheses to date, a University of Chicago team collected 25,139 assessments from 5,259 scientists — novelty, feasibility and likelihood on a one-to-nine scale, adoption on a one-to-seven scale. At no point did the questionnaire ask for the reason behind those figures. A reward model was then trained on those ratings and on pairwise preferences derived from them. Almost nine in ten of the papers involved came from biology and medicine; one fifth was medicine specifically.
Recommendation. Where AI-assisted selection informs research funding, experimental design or clinical investigation, transparency requirements should cover the reason attached to a decision, not only the origin of the artefact that carries it. Concretely: that evaluation instruments used in AI-assisted scientific assessment record a decision-specific justification alongside the score; that systems be permitted a third outcome — admission, refutation, or declared indeterminacy with its cause — rather than being forced into a binary; and that provenance standards for text be treated as a near-term deliverable rather than a maturing aspiration.
A system without a third value ends up asserting where it should withhold judgement, because fluency reads as competence and hesitation reads as failure. Should this form of record-keeping spread to clinical research assisted by artificial intelligence, the patient may inherit decisions whose justification is no longer reconstructible from the system that selected them.
Sources
Bao, H., Wu, S., Liu, X., Li, S., Cao, S. & Evans, J. A. (2026). Contemporary AI lacks the imagination to diverge or negate in science. arXiv:2606.08251v2
Chen, Y., Benton, J., Radhakrishnan, A. et al. (2025). Reasoning Models Don't Always Say What They Think. arXiv:2505.05410
Duede, E., Gross, K., Crockett, M. J. & Bergstrom, C. T. (2026). The unintended consequences of large language models as a labor-augmenting technology in science. arXiv:2607.17397
Glickman, K. (31 July 2026). Scientists using LLMs will 'do more, less well', modelling study predicts. Nature
OpenAI (31 July 2026). Advancing responsible AI across Europe
Full article
English: A number with no recorded reason may end up deciding for a patient. Editorialia — La Biblia de la IA / The Bible of AI, ISSN 2695-6411.
Original en español — Lloret Egea, J. A. (2026). Un número sin razón registrada puede acabar decidiendo por un enfermo. itvia.online, IAeñ™. DOI: 10.21428/39829d0b.55593db8
Resumen en español: la transparencia de la inteligencia artificial se ha convertido en un problema de procedencia. La procedencia dice de dónde viene un contenido; no dice por qué se decidió lo que en él se afirma. Donde la selección asistida por IA orienta financiación, diseño experimental o investigación clínica, la exigencia de transparencia debería alcanzar al motivo registrado de una decisión, y no sólo al origen del contenido que la transporta.
- Login for at skrive kommentarer
Kommentarer
Juan Antonio — thank you for this. You've pointed to something most transparency discussions overlook.
Provenance tells you where content came from. It doesn't tell you why that content was chosen over alternatives. And as you rightly point out, for scientific and clinical accountability, the "why" matters more than the "where."
We see the same gap on the enterprise decision side. An AI recommendation can carry a C2PA label and still lack any auditable rationale — still be impossible for a human to sign off on with confidence.
This is why we have been building a different layer: not provenance, but justifiability. A verification framework that builds auditable decision trails — evidence-graded, fact-anchored, traceable — into every AI output. The question it answers is not "where did this come from," but "why should this be trusted over the alternatives, and who is accountable for it?"
I also strongly agree with your point about a third outcome. Systems forced into binary decisions will assert where they should withhold judgment — fluency reads as competence, hesitation reads as failure. That is a design flaw, not a feature. We have been implementing explicit "falsifiability-bound" declarations for scenarios that cannot be verified within practical time horizons, precisely to avoid this false certainty.
Your editorial captures something important: if this gap propagates into clinical research, patients inherit decisions whose justification is no longer reconstructible. That is not a transparency problem. That is an accountability problem. And it requires a different solution than provenance alone.
Thank you for writing this.
Su Yue Fan (Paul) | Zhenyi AI