At ISMPP Annual 2026, no one was debating whether to use AI. The conversation had moved firmly into harder territory: who is accountable, what gets disclosed, how best to use it and how do we ensure that our data within publications is AI discoverable.
The discoverability risk is real – AI is becoming our most important ‘stakeholder’.
The keynote from James Phimister of Open Evidence delivered a direct challenge to publication strategy. AI clinical decision support platforms, like Open Evidence, now reaches over 40% of US physicians, delivering more than 20 million monthly consultations. Traffic from these platforms to published content now rivals PubMed. This clinical decision-making tool is not just a LLM, it is a retrieval-augmented generation system which retrieves information from a curated medical evidence base (such as NEJM, JAMA, Cochrane and NCCN). Although sites such as OPEN Evidence are transparent about its sources, it is unclear on the hierarchy of evidence for retrieval purposes. Also given it is closed to US physicians, how do we know if data within our publications is ‘retrieved’ appropriately? Journals can of course track where traffic to their articles are coming from, assuming of course HCPs check data sources, which may not be the case, as more than half of HCPs trust what they get back from AI searches. Moreover, given this is only for US physicians, what about other geographies? These types of clinical decision-making retrieval platforms are likely to be developed in different countries in local language with relevant guidelines, the question will be again is any clinical validation included in these. The stakes of getting this wrong were illustrated during the conference week itself, when Nature reported that a fictitious disease planted by researchers on preprints.org began appearing as a clinical indication in ChatGPT and Perplexity results. Platforms grounded in curated, licensed content could not surface it. Content locked behind paywalls or not structured for AI faces the same invisibility risk, precisely when clinical decisions are being made. This is a commercial consequence, not just a technical one, and it needs to be built into journal targeting and publication planning now. The key question everyone was asking was ‘what determines whether a publication surfaces in general AI systems?’. The answer is not impact factor: it is LLM discoverability. Generative engine optimisation (GEO) is evolving; from a publication perspective ‘The AI or Nay?’ session was explicit on this point: open access, machine-readable formats, and reputable licensing are what get publications in front of HCPs. Practical considerations like publishing in journals that use AI accessible figure formats (SVG, XML, HTML) is more important than ever. More and more patients are also trusting AI more than their HCPs. So, with this in mind the need to have plain-language summaries outside of the paywall becomes critical.The governance gap is real
The AI or Nay? plenary reported that over 95% of publication professionals actively use AI. Yet most acknowledged they do not know whether their external authors are using approved, compliant platforms. Data moves between sponsors, agencies, authors, freelancers, and publishers, and at each handoff a different AI system could be in play. Disclosure standards are equally disparate: there is still no uniform position across publishers on where AI use should be declared. An industry effort to align on a standard is underway, but agreement has not yet been reached.Questions to ask
Governance can no longer be informal. When reviewing your AI approach, you should be asking:- Which tools are used at which workflow stages?
- Is AI disclosure aligned with the evolving standards of target journals?
- Is LLM discoverability factored into journal selection alongside traditional impact metrics?
- What validation is in place before AI-assisted content reaches authors and submission?
