AI Copilot, Not AI Author

A few weeks ago I wrote about why KISHO deleted its AI content pipeline and replaced it with deterministic code. The short version: a language model confidently put the wrong treatment on the wrong disease page. The data feeding it was clean. The model blended. The fact-checker missed it because the fact-checker was also a language model. We deleted the whole thing.

That post was about what we stopped doing. This one is about what we built instead.

The right division of labor

The problem with AI-generated disease content isn't that AI is useless in rare disease. It's that most tools ask AI to do the wrong job.

Writing clinical content from scratch, deciding what's accurate, determining what a family with a newly diagnosed child needs to understand first — those are expert jobs. The experts for any given rare condition might number in the dozens globally. Genetic counselors who have managed hundreds of cases. PAG medical directors who know the natural history better than most published papers capture it. Researchers who have spent careers on a single pathway. No model trained on general medical literature closes that gap.

But those same experts are stretched thin. A PAG science director might be responsible for maintaining disease content across multiple conditions, writing advocacy materials, reviewing research updates, and fielding questions from families. The bottleneck isn't expertise. It's time.

That's where AI belongs in this field. Not as author. As the tool that makes expert review faster.

What AI is doing in KISHO right now

Behind the scenes, AI classification and enrichment pipelines run continuously. KISHO aggregates data across 12 authoritative sources covering 10,888+ conditions: ClinicalTrials.gov, PubMed, FDA, NIH RePORTER, Congress.gov, MONDO, HPO, Orphanet, HGNC, ClinVar, ClinGen, and GeneReviews. Every day, trial statuses change. Drug designations get updated. Policy bills move. Research gets published.

No human team monitors that at scale. AI classification handles the noise reduction, the tagging, the signal detection. When a trial status change is meaningful, it surfaces. When a bill relevant to a specific disease area advances in Congress, it gets flagged. That layer runs 24 hours a day and it earns its cost because the task is bounded and the outputs are verifiable.

That's the infrastructure side. The content side works differently.

How the KISHO Workspace tools are designed

The Workspace is where PAG teams and medical affairs organizations do their actual content work: disease pages, patient materials, intelligence reports, advocacy briefs. The inline AI tools available there are built around one principle: every result requires a human decision before anything changes.

Validate checks a selected passage against PubMed evidence. It pulls abstracts and returns a confidence score from 0 to 100. It does not rewrite the claim. It gives the editor the evidence to decide whether the claim holds.

Simplify rewrites selected text for an 8th-grade reading level, then checks its own work. After rewriting, it extracts the medical terms from the original and flags any that went missing in the output. A genetic counselor reviewing a patient-facing summary can see in seconds whether the simplification lost something clinically meaningful. The Flesch-Kincaid reading score displays in the interface.

Cite searches PubMed for relevant citations. No AI model involved. Pure API lookup against a known, structured source. Zero hallucination risk by design.

Elaborate expands a selected passage using KISHO's structured data as a reference block. The model cannot invent. When the reference data is empty, the tool returns the original text unchanged rather than filling the gap with synthesis.

None of these tools auto-apply anything. Results appear in a panel. The editor accepts, modifies, or rejects. Every action is logged: what was selected, what the AI returned, what the editor decided. The audit trail is complete.

The editor is the author. The AI is a fast research assistant that shows its work.

Why "patient-friendly" still requires an expert

Patient-facing content for rare diseases carries specific risks that don't exist in other health content contexts. A simplification that drops a gene-specific treatment consideration might seem fine to a general reviewer but be clinically meaningful for a subset of patients with a particular variant. A list of symptoms that omits an early-onset presentation can delay diagnosis for families who don't recognize what they're seeing.

These aren't edge cases. In rare disease, the edge is often the entire population.

The Simplify tool's term-checking step exists because of exactly this failure mode. An 8th-grade rewrite that reads clearly but silently drops the word "autosomal recessive" from a section on inheritance creates a different document than the original. The editor needs to know that happened and decide whether it matters for their audience.

Speed and accuracy are not in tension if the tools are designed correctly. What creates tension is asking AI to do the accuracy work instead of the speed work. KISHO's tools are built to do the latter so experts can focus on the former.

The broader architecture

The content tools are one layer. The data underneath them matters too.

GeneReviews, the 938 expert-authored clinical chapters from NCBI, are treated as ground truth in KISHO's knowledge pipeline. Not as one signal among many to blend. When a genetic counselor has written a peer-reviewed chapter on a specific condition, that chapter earns a different trust status than a PubMed abstract pulled at inference time. The Elaborate tool grounds its expansions in KISHO's structured data, which includes GeneReviews content, before reaching for anything else.

This is what "AI copilot" actually means in practice. The model operates inside a reference frame defined by authoritative sources. It cannot synthesize freely. It can help an expert work faster within a structure the expert controls.

What this means for organizations

For PAG teams: the Workspace tools are built for people who already know the disease. They are not a substitute for that knowledge. A genetic counselor who manages content for three conditions can validate a claim in seconds instead of minutes, produce a readable patient summary without losing clinical precision, and build a properly cited advocacy document without a separate research pass. The expertise requirement does not change. The time cost does.

For medical affairs teams: the same tools support the content workflows that currently eat hours of subject matter expert time on review cycles. Validate surfaces the PubMed evidence so the reviewer is confirming, not searching. Cite builds the reference list as the document gets written. The SME's judgment is still the output. The inputs arrive faster.

AI that accelerates expert review is a real productivity gain in a field where experts are scarce. That's the version worth building. The version that attempts to route around expert judgment is not a productivity gain. It is a different kind of risk showing up on a disease page that a family trusts.

KISHO is built around the first version. The word we are trying to earn is trusted. That requires keeping the expert in the chain, every time.

KISHO is a rare disease intelligence platform covering 10,888+ conditions across 12 authoritative data sources. The Workspace is available to patient advocacy organizations, pharma and biotech medical affairs teams, and research institutions. Learn more at kishomed.io.