The Sentence Both Sides Already Agree On: Why Regulatory AI Needs a Shared Standard Before It Needs Better Models

Ask ten medical writers what a Clinical Study Report should say about a given finding, and you'll likely get ten different sentences. Ask which is "correct," and the honest answer is: it depends who you ask.

Companies interpret the same ICH guidelines differently. Regulatory guidelines often lack prescriptiveness, leaving the industry without clear guidance on how much information is "just enough." There is no single, universally agreed answer to what a good sentence about a given finding looks like.

That's been a manageable problem for as long as humans have been writing and reading these documents. It stops being manageable the moment AI is doing the writing on one side and becoming part of the review environment on the other.

What AI does with an undefined standard

A model has to learn "good" from somewhere. Absent a clear, agreed target, it defaults to whatever is statistically common in the examples it has been trained on - which includes every bad habit already baked into the industry: over-writing, repetition, hedging, and restating tables in prose out of fear rather than necessity. An AI model with no standard does not fix the over-writing problem medical writers have been fighting for a decade. It scales it, at the speed of generation.

Now push the timeline forward. Sponsors are increasingly using AI to draft regulatory documents. On the agency side, AI is no longer hypothetical. FDA launched its internal generative AI tool, Elsa, agency-wide in 2025 and in May 2026 released Elsa 4.0 alongside HALO, a platform consolidating more than 40 application and submission data sources, systems and portals. (Source: FDA) FDA says the integration is designed so that staff can query agency data and build workflows without manually uploading documents into each chat. In FDA's own words, the direction is from staff bringing data to Elsa to Elsa sitting "on top of" the agency's data.

FDA is not alone. In March 2026, EMA expanded its AI-enabled Scientific Explorer so EMA and national competent authority assessors can retrieve regulatory scientific information relating to initial marketing-authorisation applications and regulatory precedents. (Source: EMA) That is not the same as an AI independently reviewing a CSR. But it does mean AI is already becoming part of the information environment on both sides of the regulatory exchange.

Now consider what happens as these systems become more capable. A document may be generated with AI on the sponsor side and interrogated, searched or assessed with AI-assisted workflows on the agency side, with no guarantee that either system is working from the same definition of "correct." The sharper risk is not simply AI checking AI. It is two probabilistic systems operating without a shared external ground truth to fall back on when they disagree.

The underlying models can also change without any change in the regulatory standard they are meant to serve. Without an external content standard, there is no independent anchor against which the behaviour of successive models can be compared. Industry legal analysis is already advising sponsors to make submissions more "AI-ready" through consistent terminology, stronger cross-referencing and conclusions that trace cleanly to underlying data. That is not the same as standard sentences, but it points in the same direction: regulatory content designed not only for human readability, but for consistent machine interpretation.

Worth being precise about what the January 2026 FDA/EMA Guiding Principles of Good AI Practice actually solve, because the distinction sharpens this problem. (Source: EMA/FDA) The principles address areas such as human-centric design, risk-based validation, data governance, performance assessment, documented limitations and lifecycle management. They do not define a shared external standard for what a particular clinical finding should mean or how it should be expressed. Importantly, however, FDA and EMA explicitly identify international harmonisation and consensus standards as areas for collaboration, and say that as AI use evolves, good practice and consensus standards must evolve with it.

FDA's draft guidance on AI in regulatory decision-making excludes AI used to draft submissions where that use does not affect patient safety, product quality or the reliability of results. (Source: FDA) That leaves an interesting boundary question for regulatory writing. When AI merely renders an already-established finding into prose, is it drafting? And when AI determines the finding itself, has it begun producing information that supports regulatory decision-making? The distinction matters because the hard regulatory problem is not language generation alone. It is how the clinical conclusion was established.

An old answer to a new version of the problem

Dr. Barry Drees, TriloDocs' founding advisor, has been proposing a strikingly literal fix to a version of this problem for years, long before generative AI made it urgent: standard sentences.

The idea is simple. For a defined set of conditions, agree the sentence in advance - jointly, across regulators and industry, not decided by any single sponsor or vendor. If there is no difference between treatment groups that would affect interpretation of safety and efficacy, say exactly that, in a form both sides already recognise. If the primary objective was met, say so, in the form both sides expect. Once the sentence itself is not in question, everything downstream gets simpler: a writer can check the finding against the source data and select the pre-agreed language; a reviewer can move more quickly to verification instead of interpretation; and an AI system on either side has a fixed target to write toward or check against, rather than a statistical average of whatever has been written before.

The building blocks already exist. Several ICH guidelines speak directly to what a section should communicate, but many stop short of prescribing the precise content or expression of an individual clinical finding. Others are now making the same point about AI quality more broadly. A peer-reviewed article in ACRP's Clinical Researcher in April 2026 argued that regulatory writing still lacks a standardised quantitative way to define or measure the quality of AI-generated text. (Source: ACRP Clinical Researcher, 2026) Internal benchmarks and gold-standard documents are useful, but an organisation-specific benchmark still proves only that a system meets that organisation's definition of good. It does not establish that a sponsor and an agency share one.

Structure and sentence are the same fix, at different resolutions

This connects to something larger already underway in the field. A 2026 framework published in Therapeutic Innovation & Regulatory Science, led by regulatory medical writing professionals from eight major pharmaceutical companies, describes a parallel discipline: mapping ICH E3 against the wider set of guidelines governing efficacy and safety reporting - E9, M4E, E2C, CIOMS III and V - and against the eCTD structure those documents ultimately populate, so writers know where a given piece of content belongs before drafting starts. (Source: Ther Innov Regul Sci, 2026)

The same framework goes further. It describes structured authoring as a writing standard built around consistent terms and sentence structures to support reuse and automation. It also points to Novo Nordisk's CSR automation approach, which combines study results and protocol information with a predefined, pre-approved text library. That is not yet the shared standard proposed here - one company's approved language remains that company's language. But it demonstrates something important: agreeing language and content rules before generation is not an obstacle to automation. It can be what makes reliable automation possible. The authors also suggest that cross-company collaboration could allow the lean CSR to develop into an industry standard, and explicitly invite regulatory reviewers to comment on the approach.

ICH M11 shows the same instinct already turned into adopted regulatory guidance, one document upstream. The CeSHarP protocol template standardises structure, controlled terminology and machine-readable data elements. M11 does not standardise clinical judgement in a CSR, and that distinction matters. But it is a live example of regulators moving toward greater structural and semantic specificity where consistency and machine readability matter.

Content mapping standardises where evidence belongs. The next challenge is standardising what that evidence means. Standard sentences are the most literal expression of that idea, but the deeper requirement is a shared content model: given the same evidence and the same defined conditions, two systems should resolve to the same regulatory finding before either one decides how to phrase it.

That distinction matters. The goal is not to eliminate language variation for its own sake. It is to stop clinical interpretation from varying simply because the model, vendor or organisation changed. The standard therefore has to sit outside any one vendor's product. A vendor whose AI is internally consistent and well documented has shown that its own system is reproducible. It has not shown that a sponsor's system and an agency's system share the same clinical interpretation.

Where this leaves AI vendors, TriloDocs included

None of this is fully solved yet, by TriloDocs or anyone else in this space. What is becoming clearer is the direction. The architecture that matters is not simply the one that generates the most fluent language, or even the one with the best internal audit trail. It is one that can be anchored to an external, shared standard as the industry converges on one - deterministic where clinical rules can be defined, traceable back to a specific finding and a specific piece of evidence, and explicit about where interpretation begins.

That is the principle TriloDocs is built around: establish the clinically relevant finding first, through explicit rules a reviewer can inspect, before language is generated. It is the same instinct behind Barry's standard sentences, applied at the level of clinical findings rather than individual phrases - and it is the foundation an eventual shared standard would need to build on, whichever direction the industry takes it.

AI can write. The harder, more urgent work is agreeing - across companies, and eventually across the table from regulators - on what it should mean before we argue about how it should say it.