AI Tools Oncologists Use in 2026: Evidence and the Tumour Board Gap

Five categories of clinical AI oncologists run today

Ask an oncologist which AI tools they use and the honest answer is usually a stack rather than a product. As of April 2026 more than 1,500 AI-enabled medical devices hold FDA marketing authorisation, and roughly 76% of them are radiology devices, with cardiovascular at 10% and neurology at 4%. Most cleared through the 510(k) pathway, which asks for substantial equivalence to a predicate device rather than evidence of clinical benefit.

That distinction matters for how oncologists read vendor claims. Authorisation is a floor, not a result. The table below separates what each category has actually been shown to do from what it is often assumed to do.

Note the last row. Five categories have vendors, clearances and in one case randomised evidence. The sixth, the literature reconciliation a molecular tumour board does between a sequencing report and a treatment decision, has none of that, and it is the step this post is ultimately about.

Imaging AI in cancer screening: what the MASAI randomised trial proved

MASAI is the strongest evidence any AI tool in oncology has produced. Gommers, Hernström and colleagues, with Kristina Lång as senior author at Lund University, randomised more than 100,000 Swedish women between April 2021 and December 2022 to AI-supported mammography screening or standard double reading. Final results published in The Lancet in 2026 (407:505-514, PMID 41620232).

In the AI-supported arm, 338 of 420 cancers (81%) were detected at screening, against 262 of 355 (74%) under double reading. Interval cancers fell from 1.76 per 1,000 screened (93/52,872) to 1.55 (82/53,043). The AI arm found 16% fewer invasive cancers presenting later, 21% fewer large tumours and 27% fewer aggressive subtypes. False positive rates were 1.5% and 1.4%. The earlier interim safety analysis reported a 44% reduction in screen-reading workload.

This is what a validated oncology AI claim looks like: randomised, powered on interval cancer, published with denominators. It is also narrow. MASAI tells a breast unit how to read screening mammograms. It says nothing about the patient whose tumour has already been sequenced.

Digital pathology AI in prostate and breast cancer diagnosis

Paige Prostate received the first FDA authorisation for an AI product in digital pathology in September 2021, as an assistive tool flagging areas of a digitised prostate biopsy slide suspicious for cancer. The regulatory framing has held: the pathologist signs out, the software points.

Adoption follows scanner deployment rather than clinical enthusiasm, which is why uptake is uneven between academic centres and community laboratories. For the tumour board, digital pathology AI changes when a diagnosis is confirmed. It does not change what happens when the diagnosis is confirmed and standard therapy has already failed.

Ambient AI scribes in the oncology clinic: relief that is real and bounded

Ambient documentation is the fastest-spreading AI category in oncology, and the one with the clearest positioning from the profession itself. Debra Patt, MD, PhD, MBA, Executive Vice President of Texas Oncology, President of the Community Oncology Alliance and Chair of ASCO's AI Task Force, framed the current boundary plainly in an April 2026 interview: "Where we are today is that we want administrative burdens to be relieved, and we want our people to be able to work at the top of their license." She added that AI used appropriately "doesn't threaten [the doctor-patient relationship], it can enhance it, because doctors and nurses can have more meaningful interactions with patients when they are less burdened by administrative tasks."

Her first question about any oncology AI deployment is the right one to carry into the rest of this post: "What is the use case?" Scribes have a clean answer. They give back clinic time. They do not read a paper, weigh an ERBB2 exon 20 insertion against a phase I signal, or tell you whether the evidence for a repurposed agent holds up.

General-purpose LLMs at the molecular tumour board: what ChatGPT got wrong

This is where most oncologists have quietly experimented, and where the published evidence is least reassuring.

Schmutz and colleagues ran ChatGPT 4.0 against 20 consecutive real molecular tumour board cases from the Comprehensive Cancer Center Augsburg, published in The Oncologist in 2025 (PMID 40973166). The model produced a median of 3 therapeutic recommendations per case against 1 from the human board (P = .005). It drew on level 3 to 4 evidence in roughly 15% of cases where the human board used none (P = .0019). Internal consistency across triplicate runs was moderate, median Fleiss kappa 0.51 with a range of 0.12 to 1.0. Run the same case three times, get materially different boards.

The broader picture is the same. Carl, Schramm, Kather and colleagues reviewed 34 studies of LLM use in clinical oncology in npj Precision Oncology in 2024 (PMID 39443582). Treatment recommendations were the dominant test case (30 of 34 studies), but only 12 of 34 assessed test-retest reliability and only 9 of 34 reported prompt strategy. Performance variance across studies was substantial and largely methodological.

The unclosed gap: somatic variant interpretation and the 14-day window

A molecular tumour board convenes when standard treatment has failed, to decide whether a mutation points to another therapy. Three findings define why that decision is still hard.

Classification is inconsistent between trained specialists. In the Association for Molecular Pathology multi-laboratory assessment, 134 participants classified 11 variants across four cancer cases against the AMP/ASCO/CAP guidelines. 86% classified variants correctly overall, range 54% to 94%. But only 59% of responses matched the working group consensus on both tier and category of clinical significance, range 39% to 84%.

The knowledge bases do not close the gap. Pallarz, Benary and colleagues at Charité and Humboldt University compared the public precision oncology knowledge bases and found substantial divergence in gene and variant coverage between CIViC, ClinVar, OncoKB and their peers, with no single resource complete on its own (PMID 32914021). Reconciliation is therefore manual, and it happens under time pressure.

Most of what sequencing returns is uninterpretable at the point of decision. Across the TCGA PanCancer Atlas, 10,967 samples over 32 studies, the proportion of identified mutations already classified as variants of uncertain significance runs at 62% for ATM (474/766), 68% for CHEK2 (101/148), 70% for BRCA1 (215/307) and 75% for BRCA2 (513/683) (PMID 38847368).

Against that, the 2025 IASLC consensus statement on molecular tumour boards, authored by Aldea, Rotow, Arcila and 30 colleagues, advises that regular meetings be held "to avoid delays beyond 14 days from result availability to discussion."

Where OncoSkepsis fits, and what it does not claim

OncoSkepsis is the free clinical tier of BioSkepsis, built for verified oncologists and molecular tumour boards. It searches the primary biomedical literature and returns the relevant evidence with graded citations, retraction screening and a trust index on every answer. Board briefs, variant evidence summaries, country-specific literature checks, and literature-backed reimbursement dossiers, which are needed for roughly one in three patients according to the clinicians who asked for the feature.

Three of those features came directly from oncologists in Cyprus, Greece, Portugal and the United Kingdom, not from a product roadmap. Country-specific literature was the most insistent request: a variant treated as benign on the basis of one population dataset may not be benign in another, and the databases do not carry that nuance.

What OncoSkepsis does not claim: it does not classify variants, it does not make treatment recommendations, and it is not a device. It does the literature step, with every claim traceable to a source you can open. The board decides. That boundary is deliberate, and it is the one Patt and Flora are both describing from inside the profession.

The engine underneath is BioSkepsis: over 40 million curated biomedical papers from 1931 onwards, analysed full text rather than by abstract, with preprints from bioRxiv and medRxiv treated as first-class content.

Why the clinical tier is free, and what leaves your tumour board

Free tools in oncology usually mean one of two things: a trial that will end, or a product whose real customer is not you. Neither is the case here, and the honest version is more reassuring than a mission statement, so here it is in full.

One engine, three surfaces. BioSkepsis is a single citation-grounded literature engine. What differs is the account type it serves. Academic and biotech researchers get the Research Workspace. Verified clinicians and tumour boards get OncoSkepsis, free. Pharmaceutical Medical Affairs teams get Insights, the paid dashboard sold under the name LeaderSkepsis. OncoSkepsis is not a stripped-down clinical edition of a research tool. It is the same engine, exposed to a different account type.

What pharma actually buys. Not your searches. Every existing expert-intelligence platform infers influence from artifacts an expert has already produced: publications, trials, grants. That is retrospective by construction, and the lag is long. Across 165,135 registered trials, the median time from first patient enrolled to published paper is 4.8 years, and only 53% are ever published in full (PMID 39601300). Bibliometric scoring is least accurate for early-career researchers (PMID 24165898), which is exactly where identifying someone first has value. Sustained, specific enquiry into a mechanism is an earlier marker of expertise than a paper about it. Insights receives aggregated, derived signals of that kind. It does not receive queries, cases or users.

The boundary is architectural, not a policy promise. The oncology workspace operates under an institutional licence, and clinicians opt in explicitly after seeing what is recorded, what is not, and what reaches pharmaceutical customers. Four guarantees follow from how the system is built rather than from what we undertake to do:

Saying this plainly is a deliberate choice. A free clinical tier presented as charity invites the question of what is really being funded, and the answer is better than the suspicion.

Oncologists shaping AI in cancer care

The people setting the terms for clinical AI in oncology are not vendors. Debra Patt chairs ASCO's AI Task Force while running a community oncology network. Douglas Flora edits AI in Precision Oncology from a community cancer centre. Kristina Lång built the randomised evidence base for screening AI at Lund. Sanjay Juneja, a triple board-certified haematologist-oncologist, now leads clinical AI strategy at Tempus. Aldea, Rotow, Arcila, Sholl, Rolfo, Subbiah, Leighl, Drilon and their co-authors wrote the IASLC consensus that defines what a molecular tumour board owes a patient.

The pattern is consistent. Every functional piece of oncology AI was specified by clinicians who described the failure mode first. The literature step at the tumour board has not had that treatment yet, which is the reason for the invitation below.

If you sit on a molecular tumour board, OncoSkepsis is free and we want the criticism. Specifically: where the retrieved evidence misses a paper you knew about, where the country-specific check fails for your population, and where a board brief would not survive your own board's scrutiny. Those three failure reports are worth more to us than any feature request.

Questions oncologists ask about clinical AI tools