AI Does Not Remove the Need for a Governed Slide Library. It Makes One Essential.

Retrieval speed and scientific reliability are different properties. AI is improving the first quickly. The second still depends on what the content carries with it.

A small change in wording can be a large change in meaning

Removing the qualifier changes neither the subject nor the direction of the finding. It changes how certain the claim sounds, and a reader can reasonably take the second version as a stronger conclusion than the evidence supports.

A 2026 study presented at the Conference on Empirical Methods in Natural Language Processing gave this effect a name and a measurement. The authors define certainty distortion as a meaningful change in expressed certainty during a transformation that was meant to preserve meaning. The measure is about how confident a statement sounds, not whether it is factually correct.

Across the rewriting tasks measured, certainty distortion affected up to 75% of model outputs, and the shifts were not symmetrical. Most of the five models tested were 1.5 to 2 times more likely to raise expressed certainty than to lower it. The pattern varied by model and by task. Two of the five were roughly balanced on the science texts, and one was the only model to lower certainty more often than it raised it on the medical texts.

Repetition compounded the effect, and only in one of the two domains. On the medical texts, one model raised certainty in 20% of examples after a single paraphrase and 40% after five, and the tendency grew by a further 13 to 20 percentage points by the fifth pass. On the science texts most of the shift happened on the first output instead of accumulating.

The problem was not that the rewritten text became obviously wrong. The language stayed fluent and plausible while the level of certainty moved.

This is why qualifiers are not padding. Words such as may, could, suggests, and associated with carry the boundary of the claim, and a rewrite that drops them has changed what is being asserted even when every fact in the sentence survives.

The study’s texts were scientific findings paired with their news write-ups, from biology, computer science, and psychology, and sentences from the findings sections of radiology reports. Neither is pharmaceutical content, and the study did not examine regulated promotional or Medical Affairs review workflows. The authors kept their prompts deliberately simple to reflect ordinary use, and found that instructing a model to preserve certainty reduced distortion without eliminating it.

What could this mean for a Medical Science Liaison?

That scenario is illustrative rather than a documented incident, and whether a given instance raises a compliance issue depends on the claim, the context, the applicable requirements, and the review process that applies. The practical exposure is simpler. A claim can end up stronger than its evidence. Approved wording can change during summarization or reuse. The altered version travels into later presentations. And a source that was approved does not make a modified version of it approved.

Plausible is not the same as approved

AI systems are good at returning something relevant. That is not the same as returning the answer a particular organization stands behind.

A 2026 study in the Journal of the Medical Library Association put five public chatbots, ChatGPT, Gemini, Copilot, DeepSeek, and Grok-3, against evidence summaries written by medical information scientists at Vanderbilt University Medical Center. Forty-five clinical questions produced 225 chatbot answers, and pairs of information scientists rated each one against their own summary.

• Total alignment: 20.9% (47 of 225)
• Partial alignment: 78.7% (177 of 225)
• No alignment: 0.4% (1 of 225)

Those ratings are easy to misread, so it is worth being exact about what they measured. Alignment here means agreement with one information scientist’s summary of the evidence. Differences in wording, or in which data were cited, still counted as aligned when they expressed the same concept.

Partial alignment covered four situations, and in three of them every shared concept agreed. The rating was partial only because one side had included additional material: the chatbot, the information scientist, or both. Only the fourth situation involved actual disagreement on the elements themselves. The authors note that assessing whether the additional material was valid and relevant remains future work, which is a direct statement that alignment was not a measure of correctness.

The five chatbots did not differ significantly from one another. What did differ significantly was the extra material, which the information scientists supplied more often than the chatbots did.

So four things are related here and none of them are the same thing: whether an answer aligns with a reference answer, whether it is factually correct, whether it is clinically appropriate, and whether it matches the scientific position a company has approved. A system searching broadly for a reasonable answer can satisfy the first and still miss the fourth, because a company position, an approved response or a preferred evidence source is not something it has any way to know. A governed system does. A plausible answer does not.

Approval is a moment in time

Content does not hold still after it is approved. A publication appears. A label changes. A safety analysis is updated. A company position develops. A new version of the source slide replaces the one that was signed off.

So the useful question is not was this slide approved. It is: is this the current approved slide, what evidence supports it, and where is the older version still in use?

That last part is where most systems stop, and it is the question a governed slide library exists to answer.

The industry is already treating this as unfinished work. A phactMI benchmarking survey of 71 questions went to 36 US member companies in May 2024, and 35 of the 36 took part, a 97% response rate. Because every question was optional, the number answering any single question varies, and the figures below are reported against their own denominators.

• AI adoption was early: 3 companies using it actively and 18 exploring it, primarily for writing and paraphrasing.
• 19% (6 of 32) reported modular content strategies, all of them large and centralized.
• 16% (5 of 32) capture analytics on internal use of Medical Information assets, such as scientific response document usage by Field Medical or MSLs.
• 31% (10 of 32) use digital metrics to inform content development or channel strategy.

The first of those connects straight back to the research above: the most common use of AI in these teams was paraphrasing, which is precisely the task the certainty study measured.

The authors’ own conclusion names three priorities for the field, and the third is strengthening governance around AI and content reuse. They also write that reuse of content without sufficient oversight may introduce quality and regulatory risks. Their stated limitation applies to all of the above: phactMI represents 36 US pharmaceutical companies, and the sample may not represent the broader industry.

The unit of governance is getting smaller

Enterprise content management was largely built around documents and files. That model still matters. But the thing people reuse is usually smaller than the file. It may be a scientific statement, a chart, an approved answer, a figure, or a single slide.

Veeva has been moving in that direction. Its 26R3 release, in prerelease at the time of writing and scheduled for general availability on 20 November and 4 December 2026, moves duplicate checking on Scientific Statement records to the point of creation rather than waiting for steady state, adds an action to refresh linked statement annotation cards to the latest reference document versions before review and approval, and introduces a standard report identifying approved documents that have unapproved scientific statements linked to them.

That last one is the telling one. It exists because an approved document can contain a sub-document unit that is not itself approved. Release notes evidence a product direction rather than an outcome, and configurations differ between customers, but the direction is plain: governance is moving below the document level.

For presentations that matters, because the PowerPoint file is often not the unit of reuse. The slide is. A fifty-slide deck may hold three slides another team needs, and those three may then appear in a field deck, an advisory board presentation, and two regional adaptations. The question is whether what gives those slides their meaning travels with them.

File-level controls can say what changed in a deck and when. They cannot say which decks still carry last quarter’s version of one slide. Governing the reusable unit is more precise than governing the container it happened to ship in, and that is the layer SlideSource Library works at: governed slide management for PowerPoint, alongside the systems that hold the approval record.

Several of those relationships are in the product today. A slide carries its own version history and a per-field record of what changed. Approval state is tracked against the slide, with rules that can differ by collection. Citation extraction and evidence links connect a claim to the source behind it. Presentations are assembled from references to library slides rather than detached copies, so a usage view can list the presentations built on a given slide when that slide changes.

For content that has already left the system, vCheck stamps a supported download with a link back to the library. Whoever holds that file can follow it, see that their copy is no longer current, see what has changed, and download the updated content.

The stamp goes on each slide rather than on the file as a whole, which is the part that matters for reuse. Copy a stamped slide into a different deck and the connection travels with it, so that slide can still be checked against its source even when the presentation around it is new and the library has never seen it.

There are limits worth stating plainly. A slide that is retyped or screenshotted rather than copied carries nothing. A stamp can be deleted like any other object on a slide. And nothing can reach into a distributed file to alter or recall it. Telling someone their copy is out of date is solvable, and SlideSource Library solves it. Deciding what they do about it is not, and no system should claim otherwise.

Veeva Vault can serve as the system of record for approved scientific materials, while SlideSource Library supports governed slide management and reuse in the presentation workflows downstream of it. The two sit alongside each other. Neither replaces the other.

What should travel with a scientific slide

The Lost Slide Problem sets out four attributes that let someone decide whether a slide they have found can actually be used. AI adds a fifth.

1. VersionWhich version of the slide is current, and does a newer approved one exist?
2. Approval stateHas this content been reviewed and approved for the use intended here?
3. EvidenceWhich references and data substantiate the claim, and what else depends on them?
4. UsageWhere has the slide been reused, so affected presentations can be found when it changes?
5. Scientific contextWhich indication, population, audience, market, or company position applies?

The first four are what a person needs. The fifth is what a retrieval system needs, because of who is doing the finding. Someone searching the library brings context with them. They know the product, the audience, and the market. A retrieval system may be able to infer some of that from a slide, and sometimes it will infer correctly. But inferred context is not the same as context the organization has verified and approved,, and a slide that is right for one indication can be wrong for another while looking identical. The useful line is between what a system may guess and what has been governed.

Relevant retrieval is not governed retrieval

The first answer finds something relevant. The second finds something an organization can evaluate, which is the only kind of answer that is any use when the content is going in front of a physician.

Preserving meaning as content moves

AI is already part of scientific communication, so the open question is not whether to use it. It is what it will be allowed to work from.

When content moved slowly, teams could lean on people to remember which deck was current and where a number came from. That does not hold when content can be found, rewritten, and redistributed in seconds.

Scientific meaning lives in specifics: the version, the approval, the evidence, and the context that made the wording right in the first place.

Content that carries those things can move between systems, people, and presentations and still say what the science says. Content that does not is only as reliable as the memory of whoever last touched it.

References

1. Belem CG, Wu S, Yao H, Steyvers M, Singh S, Smyth P. From ‘May’ to ‘Is’: Certainty Distortion in Language Model Rewriting. Proceedings of EMNLP 2026 (Main). arXiv:2606.07951.
2. Blasingame MN, Koonce TY, Williams AM, Su J, Giuse DA, Krump PA, Giuse NB. Comparing five generative AI chatbots’ answers to LLM-generated clinical questions with medical information scientists’ evidence summaries. J Med Libr Assoc. 2026;114(2):94-104.
3. Tamburri R, Sendi MK, Glaetzer A, Garas J, Mehrabian M, Wnorowski S, Hermes-DeSantis ER. Benchmark Survey on Content Development and Key Performance Indicators: A phactMI Benchmarking Survey Update. Ther Innov Regul Sci. 2026;60(5):1459-1467.
4. Veeva Vault. What’s New in 26R3: Scientific Statement enhancements. Prerelease 2 November 2026; general availability 20 November and 4 December 2026.