Heme/onc fellowship moves faster than any tool can keep up with. New approvals land every few weeks, tumor boards expect trial level recall, and the half life of what you memorized last year keeps shrinking. AI has become genuinely useful for this, but only if you pick the right tools for the right jobs. Most fellows I talk to are using one general chatbot for everything, which is exactly how you end up with a confident wrong answer in front of an attending.
I am a practicing oncologist who builds AI tools for medical education, so I see both the promise and the failure modes. This is an honest guide to the AI tools worth using as a heme/onc fellow in 2026, organized by what you actually need them to do: learn for boards, answer questions on rounds, get through journal club and trials, turn dense guidelines into study material, and document faster in clinic.
What makes an AI tool actually useful for a fellow
Before the names, a quick filter that saves a lot of wasted subscriptions.
First, is it grounded in real evidence? A tool built on cited literature or physician reviewed content is far safer for clinical facts than a raw chatbot. In oncology, where a dosing or sequencing error matters, grounding is not optional.
Second, can you see the source? Good tools link to the trial or the guideline. If a tool gives you a confident paragraph about a regimen with no citation, treat it as a hypothesis, not an answer.
Third, does it fit fellowship reality? You do not have four hour study blocks. The best tools work in the gaps: between clinic patients, on the drive home, during a slow stretch on call.
Fourth, is it honest about being wrong? Hallucinations in oncology are not rare, and they read exactly like the truth. Tools that show their work and admit uncertainty are the ones you can trust.
Board prep and learning
This is where AI has changed studying the most.
MeDucation AI is built specifically for hematology-oncology, which is the key difference from general medicine question banks that treat heme/onc as one chapter. The question bank covers malignant hematology, benign hematology, and solid tumors at the depth the ITE and ABIM boards actually test. The standout is the AI explanation on every question: instead of a generic rationale, it walks through why each wrong answer is wrong the way a senior fellow would at the whiteboard, and it uses spaced repetition so the questions you miss come back at the right interval. It is a newer platform, so the question count is still growing and it does not replace ASCO-SEP or ASH-SAP for sheer breadth, but for active learning per question with a heme/onc focus it is one of the best fits for this audience. Disclosure: I founded MeDucation.
For breadth, ASCO-SEP and ASH-SAP remain the reference content for oncology and hematology respectively. Neither has AI features or spaced repetition, so many fellows pair one of them for coverage with an AI driven question bank for active practice.
Evidence at the point of care
Once you are on the wards, the question shifts from what is the answer to what does the evidence say.
OpenEvidence has become the standout, and it is free for verified clinicians. It answers clinical questions by synthesizing current literature from sources like NEJM and JAMA and shows cited, clickable sources. On a busy consult service it is genuinely useful for dosing, drug interactions, and second or third line options when the first plan fails. It is built for point of care decision support, not structured board study, so use it as a fast evidence lookup and still verify anything high stakes against the primary trial or the NCCN guideline.
Research, journal club, and trials
Fellowship runs on trials, and AI has made it much faster to find and digest them.
Consensus searches across a large body of papers and shows you the weight of scientific agreement on a focused question, which is handy when you need a quick read on whether the literature actually supports a practice. Elicit is built for real literature reviews: it pulls relevant papers and extracts structured data like sample sizes and outcomes into a table you can export, which is a real time saver for a research project or a systematic review. Perplexity is a solid cited search engine for quick, sourced answers when you want links rather than a single synthesized paragraph. All three are freemium, with heavier use behind paid tiers.
Turning dense trials and guidelines into study material
Google NotebookLM is the quiet standout for fellows. You upload your own sources, a pivotal trial PDF, an NCCN guideline, a stack of lecture notes, and it builds summaries and study guides grounded in exactly those documents, then generates a podcast style audio overview you can listen to on a commute. Because it only works from what you upload, it hallucinates far less than an open web chatbot. The catch is the flip side: if your source is outdated, so is the summary, and very long documents can lose detail, so spot check the parts that matter. It has a free tier, which is where to start.
General assistants like ChatGPT and Claude are the flexible workhorses for explaining a mechanism simply, reformatting your notes, or drafting practice questions. They are freemium, with paid tiers around 20 dollars a month. But they are not grounded in medical evidence, so they carry the highest risk of confident wrong facts and fabricated citations. Use them to explain and organize, never as your only source of clinical truth.
Clinical documentation
Notes are not learning, but they eat your time, and time is the scarcest resource in fellowship.
AI scribes such as Abridge and Nuance DAX listen to a clinic visit and draft the note, which can give you back real time at the end of a long day. Adoption is growing fast across health systems. Two caveats: use only the scribe your institution has approved and integrated, since patient audio is protected health information, and always read and edit the draft before you sign it. The note is your responsibility, not the model's.
Who benefits most at each stage
In your first year, lean on tools that build a foundation: a heme/onc question bank with AI explanations for daily learning, OpenEvidence for fast answers on service, and NotebookLM to turn the trials you keep hearing about into something you actually retain.
In your senior years, the center of gravity shifts toward trial level depth and the ITE and boards. This is where consistent question practice plus a strong content resource pays off, supported by research tools for your scholarly work. In the final board crunch, narrow your stack. Heavy question practice with good explanations beats collecting more tools. The fellows who score highest are not the ones using the most apps. They are the ones doing the most active recall.
Where these tools still fall short
I build these tools, and I will still tell you to be careful.
The biggest risk is confident wrongness. AI systems regularly produce fluent, plausible, incorrect oncology information, including fabricated trial citations, and they do it with full confidence. Always verify dosing, sequencing, and trial outcomes against the primary literature or the guideline.
The second risk is dependence. If you let AI build the differential or pick the regimen for you, you do not learn to do it yourself. Use these tools to pressure test your reasoning after you have committed to it, not instead of it. The third is privacy and policy. Anything touching patient data has to run through your institution's approved tools and rules. When in doubt, ask before you paste.
Frequently asked questions
What is the best AI tool for ITE and board prep? A heme/onc focused question bank with AI explanations and spaced repetition, used consistently, will move your score more than anything else. Pair it with ASCO-SEP or ASH-SAP for breadth.
Is it safe to use AI on rounds? For your own learning and for forming questions, yes. For decisions that touch a patient, treat the output as a starting point, verify it against the guideline or trial, and follow your institution's policies.
Which of these are free? OpenEvidence is free for verified clinicians, NotebookLM and the basic tiers of ChatGPT and Claude are free, and Consensus, Elicit, and Perplexity have freemium tiers. The dedicated question banks and content programs are paid.
Can I use ChatGPT instead of a question bank? No. General chatbots are strong explainers but are not calibrated to board content and can invent facts and citations. Use them alongside a real question bank, not in place of one.
How do I avoid getting burned by AI hallucinations? Prefer tools that cite their sources, check those sources for anything high stakes, and never trust a regimen, dose, or trial result that you cannot trace back to the primary literature or a guideline.
