Artificial intelligence in Oncology: Where Promise Meets Reality

Artificial intelligence (AI) is increasingly used in oncology to support tasks such as clinical-trial matching, evidence review, mammography, digital pathology, cancer biology, and drug discovery. However, these applications are at different stages of development and should be assessed individually. Some artificial intelligence systems have undergone evaluation in randomized clinical trials. Others have been assessed retrospectively, compared with human annotations, or tested in simulated clinical scenarios. Some remain limited to computational or laboratory research stages. For clinicians and researchers, the key question is not whether artificial intelligence “works” in oncology, but rather what a specific system has been tested to accomplish, how it was evaluated, and whether the results are clinically relevant.

Key takeaways

• Performance depends on the task. A model that accurately extracts information from a clinical report may be considerably less reliable when asked to apply a risk classification or recommend treatment. [1]
• Clinical-trial matching is becoming an important oncology application for LLMs. Several 2026 studies have evaluated AI systems that interpret clinical records and trial eligibility criteria, with some reporting substantial reductions in screening time while retaining human review. [16-18]
• The strongest randomized evidence remains in medical imaging. In the MASAI trial, AI-supported mammography screening produced non-inferior interval-cancer outcomes while reducing the workload associated with screen reading. [3,4]
• AI is also being integrated into cancer biology. Projects such as Microsoft’s Project Ex Vivo combine computational models with single-cell and spatial biology, experimental systems, and clinical oncology rather than relying on a general-purpose language model alone. [19]
• AI-generated drug candidates have entered clinical development, but early clinical experience does not establish improved therapeutic efficacy. [7]
• Human review remains part of the process. Across the strongest examples, clinicians, scientists, expert annotators, or experimental systems are still needed to verify AI-generated information and predictions.

The reality of LLMs in oncology

Large language models (LLMs) have been tested in oncology for information retrieval, evidence synthesis, clinical-trial matching, and clinical reasoning. Their performance varies considerably by task. One study compared four general-purpose models—ChatGPT-4-turbo, Claude-3-opus, Gemini-Pro-1.0, and ChatGPT-3.5-turbo—using simulated stage IV prostate cancer work-ups based on CT, MRI, bone scans, and biopsy pathology reports. The authors used 350 simulated cases to evaluate seven information-retrieval tasks and three risk-assessment tasks involving LATITUDE, CHAARTED, and TwNHI criteria.* Three adjudicators established the reference standard. [1]

For information retrieval, including TNM staging, accuracy ranged from 87.4% to 94.2%. Performance was less consistent when the models were asked to apply risk-assessment criteria. ChatGPT-4-turbo achieved the highest reported accuracy across the three risk-assessment tasks, at 90.1%, 90.7%, and 91.6%. Repeated queries and ensemble voting also improved performance. [1] The difference between these tasks is clinically relevant. Extracting a documented finding from a report is not the same as interpreting that finding within a formal risk classification. A model’s performance on one task should not be assumed to carry over to another.

The model version is important

Many of the oncology studies now being cited were conducted with models that have since been replaced or substantially updated. The prostate cancer study, for example, evaluated ChatGPT-4-turbo, ChatGPT-3.5-turbo, Claude-3-opus, and Gemini-Pro-1.0. [1]

Researchers are now evaluating newer versions of GPT, Claude, Gemini, and other models in oncology and other clinical settings. This makes historical accuracy figures difficult to use as measures of current model performance. Model capabilities, prompting methods, retrieval systems, and evaluation procedures have changed. The more useful finding from these studies is that performance varies by clinical task and the information being extracted or interpreted.

High review accuracy, uneven extraction

A 2026 breast cancer pilot evaluated a human-supervised living oncology evidence platform using GPT-4.1, o3, and Claude Sonnet-4. The system was designed to support systematic literature review and evidence extraction using a human-developed annotation manual. The investigators worked with 29,236 clinical trial abstracts covering breast, lung, and prostate cancer. The system was assessed against 1,997 human annotations involving four review variables and 32 extraction variables. [2]

Review accuracy ranged from 95.1% to 97.2%. Extraction accuracy varied considerably, ranging from 51.1% to 99.4%, depending on the variable. Three variables were still undergoing refinement. [2] Those results illustrate an important limitation of broad AI performance claims. A system can be highly accurate at determining whether a publication belongs in a review while performing much worse at extracting a particular clinical variable. For oncology evidence platforms, validation therefore needs to be performed at the level of the specific task and variable. The investigators also noted that the platform’s effect on physician workflow and clinical decision-making still requires further study. [2]

An emerging oncology application

Clinical-trial matching is one area in which AI is moving from general-purpose question answering toward a defined oncology workflow. A 2026 prospective evaluation of a neuro-symbolic, multi-agent AI system examined clinical-trial matching in 3,804 patients. The platform combined LLM agents with an oncology knowledge graph, a recommendation engine, and human review. The investigators reported an F1 score of 0.82 and reduced screening time from about 120 minutes to about 30 minutes per patient. They also reported no demographic group with a performance gap greater than 10 percentage points. [16]

The study compared the system with GPT-4, chain-of-thought prompting, and GPT-4o on the study’s matching tasks. [16] Other 2026 research has examined using LLMs to determine whether patients meet individual trial eligibility criteria. An AACR 2026 study evaluated six models across 19 eligibility questions drawn from oncology trials. Performance was strongest for criteria that were explicitly documented in the patient record, such as age, ECOG performance status, measurable disease, and tissue availability. More difficult cases arose when information was incomplete or required interpretation. [17]

A separate 2026 study described TrialMatchAI, an end-to-end system using fine-tuned open-source LLMs, retrieval-augmented generation, hybrid search, and criterion-level eligibility assessment. In its reported validation, 92% of oncology patients had at least one relevant trial identified within the top 20 recommendations, while expert assessment supported more than 90% accuracy for criterion-level eligibility classification. [18]

These systems are notable because they do more than ask a language model to answer a question. They combine language models with structured oncology information, trial databases, retrieval systems, and human review. That approach may be more clinically useful than relying on a general-purpose chatbot alone.

How AI is moving into oncology and life sciences

The 2026 landscape includes general-purpose AI developers, oncology-specific platforms, academic medical centers, scientific-computing companies, and pharmaceutical companies.

OpenAI’s January 2026 healthcare announcement identified Memorial Sloan Kettering Cancer Center as one of the institutions introducing ChatGPT for Healthcare. The announcement described applications across clinical and administrative workflows rather than a specific oncology treatment application.

Microsoft and the Broad Institute, with support from Dana-Farber Cancer Institute, are taking a different approach through Project Ex Vivo. The program combines AI with high-throughput biology and clinical oncology, initially focusing on pancreatic ductal adenocarcinoma. The stated goals include understanding tumor-cell states, studying the tumor microenvironment, improving patient stratification, and identifying potential therapeutic candidates. [19]

Mount Sinai’s Tisch Cancer Center launched PRISM in January 2026, an oncology-specific AI platform developed by Triomics for clinical-trial matching.

NVIDIA’s BioNeMo platform represents another category. Rather than serving primarily as a conversational system, it provides infrastructure for AI-driven biology and drug discovery, including molecular modeling and laboratory workflows.

These examples show why artificial intelligence in oncology should not be defined solely by the performance of general-purpose language models. Some systems are designed for clinical records, while others analyze imaging, molecular data, biological sequences, or experimental outcomes.

The strongest prospective evidence

The MASAI trial provides a different level of evidence because AI was evaluated within a randomized clinical screening study. The Swedish trial randomized women aged 40 to 80 years at four screening sites to AI-supported mammography reading or standard double reading by two radiologists. The prespecified safety analysis found non-inferior cancer detection with AI-supported screening and substantially reduced screen-reading workload. [3]

A subsequent 2026 analysis examined interval cancers—cancers diagnosed between screening rounds. It included more than 105,000 randomized participants and reported 1.55 interval cancers per 1,000 participants in the AI-supported group compared with 1.76 per 1,000 in the standard-reading group. AI-supported screening had higher sensitivity, while specificity was similar. [4] The results should not be summarized as evidence that artificial intelligence “reduces interval cancers.” The study demonstrated non-inferior interval-cancer outcomes alongside differences in sensitivity, specificity, and workload. Whether similar findings will occur with other artificial intelligence systems, screening populations, or clinical workflows remains unknown.

An earlier retrospective evaluation of mammography datasets from the United Kingdom and United States also reported reductions in false-positive and false-negative findings with an AI system compared with individual radiologists. [5] MASAI is particularly useful as a point of comparison because it demonstrates what evidence from a defined clinical workflow looks like. That is different from testing a language model against a benchmark or simulated clinical cases.

Promising prediction, not replacement testing

AI has also been evaluated in digital pathology. A 2019 study showed that deep learning applied to routine hematoxylin and eosin histology could predict microsatellite instability in gastrointestinal cancers directly from tissue morphology. [6] MSI status can influence eligibility for immune checkpoint blockade, making the ability to identify patients who may have MSI clinically relevant.

The study, however, did not establish that histology-based AI prediction could replace validated molecular or immunohistochemical testing. An AI-generated prediction can help identify patients who may warrant further evaluation, but it is not equivalent to a validated diagnostic assay on its own.

From cancer-cell states to precision oncology

Integrating AI with experimental cancer biology is another area to watch. Project Ex Vivo, involving Microsoft, the Broad Institute, and Dana-Farber Cancer Institute, is developing models that integrate genetic markers, expression states, and tumor-microenvironment interactions. The initial work focuses on pancreatic ductal adenocarcinoma and uses experimental systems to study tumor-cell states and potential therapeutic vulnerabilities. [19]

This approach differs substantially from using an LLM to summarize a clinical trial. The objective is to connect computational predictions with biological experiments. If a model identifies a potentially important cell state or therapeutic vulnerability, researchers can test that finding in experimental systems. The laboratory work remains essential. A computational prediction does not establish biological significance, and a finding in an experimental model does not establish clinical efficacy.

From investigational drug to clinical benefit

AI is also being used earlier in the drug-development process to identify molecular candidates. An analysis of AI-discovered molecules entering clinical development reported Phase I success rates of approximately 80%–90%, compared with historical industry averages of approximately 40%–65%. Phase II success was approximately 40%, although the Phase II sample was limited. [7]

These numbers require careful interpretation. Phase 1 studies primarily evaluate safety, tolerability, pharmacokinetics, and dose-related characteristics. Phase 2 begins to address whether a candidate has clinical activity or efficacy in the intended disease. The available analysis therefore shows that AI-discovered compounds can progress into human studies. It does not establish that AI has substantially improved the probability of developing an effective cancer drug.

For oncology, this distinction matters. A molecule reaching a clinical trial is an intermediate development milestone, not evidence of therapeutic benefit. The broader field has progressed beyond purely computational predictions. Molecules identified through artificial intelligence approaches are now entering clinical development in several therapeutic areas. Whether these methods ultimately improve the probability, speed, or cost of developing effective cancer treatments remains uncertain.

An example beyond oncology

A September 2026 report from Anthropic provides an example of AI being used several steps upstream of clinical research. Anthropic reported that its life-sciences research group had identified a previously undescribed enzyme system in bacteriophages, called array-associated reverse transcriptases, or ART. The finding was accompanied by a preprint and had not undergone peer review at the time of the report.

According to the company, hundreds of model agents searched large DNA sequence databases for unusual reverse transcriptases. The company assembled and narrowed candidate systems for further analysis before human scientists conducted experimental work. The ART finding should not be presented as evidence that AI has discovered a new therapeutic mechanism. It shows how computational agents can help researchers search large biological datasets and identify patterns for further investigation. Whether ART has a useful biological function remains to be determined. For oncology, the broader relevance is the workflow: AI can help generate and prioritize biological hypotheses, but laboratory validation is still required.

Risks, limitations, and open questions

Task-dependent reliability

The prostate cancer study found relatively high performance across four models for information retrieval but greater variation in risk assessment. [1] The Living-OEP study likewise found high review accuracy but a much wider range of extraction accuracy by variable. [2] For clinical applications, an overall model accuracy figure is therefore of limited value. What matters is how the system performs on the specific task it is used for.

Fluent output can still be wrong

A 2023 evaluation of an LLM chatbot found that a substantial proportion of cancer-treatment recommendations were not concordant with National Comprehensive Cancer Network guidance. Correct and incorrect recommendations could occur within the same response. [13] That study evaluated an earlier generation of chatbot technology, so its numerical results should not be treated as a measure of current models. Its relevance is the underlying failure mode: an incorrect medical answer can be presented clearly and confidently.

Algorithmic bias

A widely cited study of a commercial health-risk algorithm showed how the choice of an outcome proxy can produce systematic racial differences in the patients identified for additional care. The algorithm used health-care costs as a proxy for health needs. Correcting the target variable increased the proportion of Black patients identified for additional care from 17.7% to 46.5%. [14] The study demonstrates why the target variable matters. An algorithm can accurately predict the variable it was trained to predict and still produce inequitable results if that variable poorly substitutes for the clinical outcome of interest. The same concern applies to oncology models developed from observational, registry, or claims data when access to care influences the underlying data.

General-purpose versus oncology-specific

The growing number of oncology-specific systems raises another question: does an AI model need to be built specifically for cancer care to be clinically useful? There is no single answer. General-purpose models can perform well on selected oncology information-retrieval and reasoning tasks. Oncology-specific systems can incorporate cancer terminology, trial criteria, structured knowledge graphs, and disease-specific workflows. The clinical-trial matching studies suggest that the system surrounding the language model can be as important as the language model itself. [16-18] The prospective Massive Bio system, for example, combined LLM agents with an oncology knowledge graph and human review rather than relying on an LLM alone. [16] That distinction will matter more as health systems move from experimenting with general-purpose chatbots to integrating AI into clinical workflows.

Company announcements and preprints

Company announcements can identify important developments, but they are not equivalent to randomized clinical trials or independently replicated peer-reviewed studies. The ART finding is one example. The work may ultimately prove scientifically important, but its current evidentiary status is different from that of a randomized oncology trial. The same principle applies to announcements about AI deployment in health systems. Deployment demonstrates adoption. It does not, by itself, demonstrate improved patient outcomes.

Dual use

AI systems developed for drug discovery can have dual-use implications because capabilities used to optimize desirable molecular properties can potentially be redirected toward harmful properties. [15] For biomedical research, this makes safeguards an important part of developing and deploying increasingly capable AI systems.

Generalizability

The MASAI investigators noted that their findings arose from a specific Swedish screening setting involving particular radiologists and one AI system. [3] The prostate cancer investigators similarly called for validation in other cancers. [1] Results can change across institutions, patient populations, imaging systems, documentation practices, and clinical workflows. Validation in one setting should therefore not automatically be generalized to another.

Human review remains necessary

The strongest examples discussed in this article continue to rely on human expertise. The prostate cancer study used a three-adjudicator consensus standard. The Living-OEP platform was developed around a human annotation manual and benchmarked against 1,997 human annotations. The clinical-trial matching systems incorporated human review or expert assessment. Project Ex Vivo uses experimental biology to test computational predictions. [1,2,16,19]

Artificial intelligence can reduce some forms of manual work, but it does not eliminate the need for clinical judgment or scientific validation. In many cases, the nature of the work shifts—from searching and sorting information to reviewing, adjudicating, interpreting, and testing the results generated by artificial intelligence systems.

A bright future to watch

Will task-level validation become standard?

The most informative studies define the task, reference standard, patient population, and endpoint being measured. For clinical implementation, those details are more useful than a general statement that a model performs well.

Will prospective evidence follow?

MASAI provides randomized clinical evidence in mammography screening. Several 2026 clinical-trial-matching studies have also moved beyond retrospective benchmarks toward prospective or real-world evaluation. [4,16] Much of the remaining literature is retrospective, simulated, or based on human annotation. For AI systems intended to influence clinical care, prospective studies need to measure the clinical or workflow outcomes the technology is meant to change.

Will AI-enabled discovery translate into better cancer treatments?

AI is increasingly being used to study cancer-cell states, identify potential therapeutic vulnerabilities, and generate molecular candidates. [7,19] The important endpoint, however, remains clinical benefit. Identifying a target, generating a molecule, or prioritizing a biological hypothesis is an intermediate step. For patients, the relevant questions are whether an intervention improves response, progression-free survival, overall survival, quality of life, safety, or another meaningful clinical endpoint.

What the evidence supports today

AI is already useful in several areas of oncology, but the strength of evidence differs substantially by application.

The distinction between technical performance and actual clinical utility is fundamental. A model may extract information accurately without improving patient care. An AI system may reduce the time required to identify clinical trials without increasing enrollment. A computational model may identify a potential drug target without producing an effective therapy. Those outcomes need to be measured separately.

Clinical trials

The principal randomized clinical trial discussed in this article is the Mammography Screening with Artificial Intelligence trial (MASAI; ClinicalTrials.gov identifier NCT04838756). [3,4] The prostate cancer LLM comparison used simulated reports rather than patients receiving AI-assisted clinical care. [1] The Living-OEP study was validated against human annotations rather than evaluated as a prospective clinical decision-support intervention. [2]

The 2026 oncology clinical-trial-matching literature includes prospective and real-world evaluations, but these systems remain focused primarily on identifying potentially eligible patients and trials rather than demonstrating improved cancer outcomes. [16-18]

What’s Next?

AI is becoming part of oncology research and clinical practice, but the evidence does not support treating all AI applications as equivalent. The strongest clinical evidence currently comes from areas such as mammography screening, where artificial intelligence has been evaluated in randomized trials. Other applications, such as clinical-trial matching and evidence synthesis, are progressing toward prospective workflow evaluation. Cancer biology and artificial intelligence–enabled drug discovery remain earlier in development, with computational findings still requiring experimental and clinical validation.

The technology landscape is also diversifying. General-purpose language models are being used alongside oncology-specific models, knowledge graphs, imaging algorithms, multimodal systems, scientific foundation models, and experimental platforms. For oncology, the relevant question is therefore not whether AI is “good” or “bad,” or whether one model is better than another in the abstract. The more useful questions are very specific:

• What task is the system performing?
• How was it validated?
• In which patients?
• Against what standard?
• And does using it change something that matters to patients?

Those questions will determine which AI applications become part of routine oncology care—and which remain useful research tools.

References

[1] Yuan LH, Huang SW, Chou D, Tsai CY. The In-depth Comparative Analysis of Four Large Language AI Models for Risk Assessment and Information Retrieval from Multi-Modality Prostate Cancer Work-up Reports. World J Mens Health. 2025 Oct;43(4):918-933. doi: 10.5534/wjmh.240173. Epub 2024 Dec 2. PMID: 39743220; PMCID: PMC12505481.
[2] Rugo H, Forsythe A, Flora D, Glück S, Grieve S, Campden R, Liu RJ, Musat M, Rege JM, Kaufman PA, Coleman RL, Tarantino P, O’Shaughnessy J, Schwartzberg L. Development and validation of a human-supervised AI-augmented living oncology evidence platform: a breast cancer pilot study. ESMO Real World Data Digit Oncol. 2026 Sep 1;13:100770. doi: 10.1016/j.esmorw.2026.100770. PMID: 42729682; PMCID: PMC13563553.
[3] Lång K, Josefsson V, Larsson AM, Larsson S, Högberg C, Sartor H, Hofvind S, Andersson I, Rosso A. Artificial intelligence-supported screen reading versus standard double reading in the Mammography Screening with Artificial Intelligence trial (MASAI): a clinical safety analysis of a randomised, controlled, non-inferiority, single-blinded, screening accuracy study. Lancet Oncol. 2023 Aug;24(8):936-944. doi: 10.1016/S1470-2045(23)00298-X. PMID: 37541274.
[4] Gommers J, Hernström V, Josefsson V, Sartor H, Schmidt D, Hjelmgren A, Larsson AM, Hofvind S, Andersson I, Rosso A, Hagberg O, Lång K. Interval cancer, sensitivity, and specificity comparing AI-supported mammography screening with standard double reading without AI in the MASAI study: a randomised, controlled, non-inferiority, single-blinded, population-based, screening-accuracy trial. Lancet. 2026 Jan 31;407(10527):505-514. doi: 10.1016/S0140-6736(25)02464-X. PMID: 41620232.
[5] McKinney SM, Sieniek M, Godbole V, Godwin J, Antropova N, Ashrafian H, Back T, Chesus M, Corrado GS, Darzi A, Etemadi M, Garcia-Vicente F, Gilbert FJ, Halling-Brown M, Hassabis D, Jansen S, Karthikesalingam A, Kelly CJ, King D, Ledsam JR, Melnick D, Mostofi H, Peng L, Reicher JJ, Romera-Paredes B, Sidebottom R, Suleyman M, Tse D, Young KC, De Fauw J, Shetty S. International evaluation of an AI system for breast cancer screening. Nature. 2020 Jan;577(7788):89-94. doi: 10.1038/s41586-019-1799-6. Epub 2020 Jan 1. Erratum in: Nature. 2020 Oct;586(7829):E19. doi: 10.1038/s41586-020-2679-9. PMID: 31894144.
[6] Kather JN, Pearson AT, Halama N, Jäger D, Krause J, Loosen SH, Marx A, Boor P, Tacke F, Neumann UP, Grabsch HI, Yoshikawa T, Brenner H, Chang-Claude J, Hoffmeister M, Trautwein C, Luedde T. Deep learning can predict microsatellite instability directly from histology in gastrointestinal cancer. Nat Med. 2019 Jul;25(7):1054-1056. doi: 10.1038/s41591-019-0462-y. Epub 2019 Jun 3. PMID: 31160815; PMCID: PMC7423299.
[7] Kp Jayatunga M, Ayers M, Bruens L, Jayanth D, Meier C. How successful are AI-discovered drugs in clinical trials? A first analysis and emerging lessons. Drug Discov Today. 2024 Jun;29(6):104009. doi: 10.1016/j.drudis.2024.104009. Epub 2024 Apr 30. PMID: 38692505.
[8] Anthropic. Claude discovers a novel enzyme system with CRISPR-like repeats. September 2026. Company announcement and linked preprint; Online. Last accessed on September 27, 2026
[9] Genmab A/S. Genmab partners with Anthropic to accelerate research and development through agentic artificial intelligence. Press release. January 7, 2026. Online. Last accessed on September 27, 2026.
[10] Maddox W. ICON inks Anthropic partnership to deploy Claude into clinical trials. Fierce Biotech. August 2026. Online. Last accessed on September 27, 2026.
[11] Anthropic. Life sciences research and AI-enabled biological discovery. Company materials. 2026.
[12] Novo Nordisk / Anthropic. Novo and Anthropic will collaborate to advance drug discovery with Claude. Press release. September 2026. Online. September 27, 2026
[13] Chen S, Kann BH, Foote MB, Aerts HJWL, Savova GK, Mak RH, Bitterman DS. Use of Artificial Intelligence Chatbots for Cancer Treatment Information. JAMA Oncol. 2023 Oct 1;9(10):1459-1462. doi: 10.1001/jamaoncol.2023.2954. PMID: 37615976; PMCID: PMC10450584.
[14] Obermeyer Z, Powers B, Vogeli C, Mullainathan S. Dissecting racial bias in an algorithm used to manage the health of populations. Science. 2019 Oct 25;366(6464):447-453. doi: 10.1126/science.aax2342. PMID: 31649194.
[15] Urbina F, Lentzos F, Invernizzi C, Ekins S. Dual Use of Artificial Intelligence-powered Drug Discovery. Nat Mach Intell. 2022 Mar;4(3):189-191. doi: 10.1038/s42256-022-00465-9. Epub 2022 Mar 7. PMID: 36211133; PMCID: PMC9544280.
[16] Loaiza-Bonilla A, Yost C, Kurnaz S, Tuysuz E, Thaker NG, Giritlioglu D, Noel Meza JP. Transforming oncology clinical trial matching through neuro-symbolic, multi-agent AI and an oncology-specific knowledge graph: a prospective evaluation in 3804 patients. ESMO Real World Data Digit Oncol. 2026 Apr 7;12:100706. doi: 10.1016/j.esmorw.2026.100706. PMID: 42004487; PMCID: PMC13091143.
[17]Desai A, McNeeley E, Alhuski S, et al. Evaluation of large language models for automated clinical trial matching in oncology. Cancer Res. 2026;86(7 Suppl):2739. Presented at: AACR Annual Meeting 2026.Online. Last accessed on September 27, 2026.
[18] TrialMatchAI: an end-to-end AI-powered clinical trial recommendation system to streamline patient-to-trial matching. Nat Commun. 2026. doi:10.1038/s41467-026-70509-w.
[19] Microsoft Research, Broad Institute, Dana-Farber Cancer Institute. Project Ex Vivo: defining, engineering, and targeting cell states in cancer. 2026. Online. Last accessed on September 27, 2026

Featured image: © 2926 Ch LLM and AI in Oncology. Used with permission.

DOI