AI for Scientists — the Two Courses Worth Taking, and What Actually Changed in the Lab

AI for scientists comes down to two courses, not a degree. For the foundation, take fast.ai’s free Practical Deep Learning for Coders if you already write code, or the Stanford Online / DeepLearning.AI Machine Learning Specialization (free to audit, $49/month for the certificate) if the maths is the real gap. Then spend two free hours on DeepLearning.AI’s Claude Code course: agentic coding is the skill that changes your actual week. Skip the $3,600 three-day executive courses.

What AI for scientific research actually changes

Four things, in order of how much time they give back: screening literature, writing analysis code, running domain models you could never have trained yourself, and drafting prose you then have to declare. Nothing on that list is “the AI does the science.”

The gap worth exploiting is narrower than the hype suggests. Anthropic surveyed 1,260 active quantitative social scientists in late February and March 2026: 81% had already used generative AI somewhere in their research process, but only 20% used coding agents more than once a week. Economists, 39%. Political scientists, 25%. Public health, education and communications, 4–6%. Just over a quarter of doctoral students and postdocs were weekly users; among tenured professors the rate fell by more than half.

Read those two numbers together. Generative AI is saturated and buys you nothing your competitors don’t have: everyone in your field already has a chatbot open. Agentic analysis is still rare enough that competence puts you in the top fifth. The survey only covers social science, so the exact percentages won’t be your field’s. The gap will be.

AI for scientists: the courses worth shortlisting

Every price and duration below was checked on the provider’s own page, most of them re-checked on 6 August 2026, then weighted by one test: does it change what you can do on Monday?

1. fast.ai — Practical Deep Learning for Coders (free)

Free · 9 lessons, each around 90 minutes · Part 2 adds 30+ hours · no certificate

Still the best course for a working scientist, and it’s free. The stated prerequisite is a year of coding, preferably Python, plus high-school maths — fast.ai teaches the calculus and linear algebra it needs as it goes. You build working models in lesson one instead of grinding through six weeks of theory. It covers vision, NLP, tabular data and deployment in PyTorch, fastai and Hugging Face Transformers.

Honest negative: no certificate, and the top-down teaching order irritates some people. If you want to understand gradient descent before using it, you’ll fight this course the whole way; anyone who thinks in derivations usually prefers the Stanford route below.

2. Machine Learning Specialization — Stanford Online + DeepLearning.AI

Free to audit · $49/month for the certificate · 3 courses, 33 + 34 + 28 hours · shareable certificate · 4.9/5 from 39,156 course reviews · 820,717 enrolled

Andrew Ng’s rebuilt classic, and the right answer if machine learning for scientists means “I need the foundations properly.” Supervised and unsupervised learning, neural networks in TensorFlow, decision trees, plus the judgment calls about bias, variance and data splits that cause most bad results in published papers. Cost tracks your pace, so finishing inside one billing cycle turns the certificate into a single $49 month instead of three.

Two things the page buries, both worth knowing before you pay. The “Enroll for free” button is audit access; the actual price sits inside a collapsed FAQ, where Coursera states “A Coursera subscription costs $49 / month”. And the 4.9/5 works two shifts — the header attaches it to this program’s 39,156 reviews, while the description attaches “rated 4.9 out of 5” to Ng’s original 2012 course. Coursera Plus, at $59/month or $399/year, is the wrong purchase here unless you want the other 10,000 courses with it.

Honest negative: it’s pitched at beginners with high-school maths, so a quantitative PhD will find the first course slow and should go straight to fast.ai. Machine learning for physicists is under-served by every general course here — none of them touch physics-informed networks or symbolic regression.

3. Claude Code: A Highly Agentic Coding Assistant — DeepLearning.AI

Free enrollment, for now · 2 hours, 10 video lessons · intermediate · taught by Elie Schoppik, Head of Technical Education at Anthropic

Two hours, addressing the exact gap the Anthropic survey found. You learn to point a coding agent at a real codebase, extend it with MCP servers, manage context with a CLAUDE.md file, and run parallel work with git worktrees. One of the three projects is refactoring Jupyter notebooks into a dashboard, a task most scientists recognise painfully well.

Honest negative: it’s a two-hour tool tutorial, not a course, the free tier omits the completion accomplishment (that needs paid PRO), and it teaches one vendor’s tool. Read the small print on the price, too: DeepLearning.AI states access is “free for a limited time during the DeepLearning.AI learning platform beta”, so the zero is a promotion, not a policy. The habits transfer; the commands don’t.

The two holes in that shortlist

No cheap certificate, and nothing about science. One closes for $90, the other for nothing. If your department wants a credential on file but won’t fund a subscription, NVIDIA’s Deep Learning Institute sells Getting Started with Deep Learning as a one-off: $90, eight hours, TensorFlow 2 with Keras and Pandas, certificate included. It’s the cheapest route from nothing to a named certificate here, and the GPU framing is closer to how computational work actually runs. The catch is that it’s an eight-hour on-ramp built around NVIDIA’s own stack, and it will not teach you the statistics.

The second hole is science itself. EMBL-EBI’s free AlphaFold: a practical guide is the only thing here that fills it, and it’s the model of what domain training should look like: three hours, built with Google DeepMind, running from the folding problem through pLDDT and PAE confidence scores to AlphaFold 3, AlphaFold Server and AlphaMissense. It’s free, it’s CC BY, and EMBL-EBI states plainly that its on-demand courses carry no certificate. Structural biologists should do this before any general ML course; everyone else should read it as the template for finding the equivalent tutorial in their own field.

Every option compared

All figures come from the provider’s own page. The Machine Learning Specialization, Coursera Plus, NVIDIA DLI, the EMBL-EBI tutorial, the Claude Code course and the MIT course were re-checked on 6 August 2026; fast.ai, Google’s crash course and AI for Medicine carry figures verified 30 July 2026. One number moved between the 5 and 6 August checks: the specialization’s review count, from 38,721 to 39,156. Assume that rate of drift on every enrolment and review figure here.

The research workflows that actually changed

Literature screening and systematic review

This is where the hours are. Elicit automates systematic-review screening and extraction, and advertises “over 125 million papers”, “99.4% Data extraction accuracy”, and “up to 80% time savings” that researchers self-report. Its one real benchmark is a run across 994 Cochrane reviews: 95% search recall, 97% abstract screening, 99% full-text screening, 96% extraction. Every one of those numbers is published by the vendor and independently audited by nobody. Consensus and Semantic Scholar handle discovery, Litmaps citation mapping.

The trap is citations. Models fabricate them, and NIH’s own notice puts “plagiarism, fabricated citations, or other kinds of research misconduct” in a single sentence. Every reference an AI hands you gets opened and read.

Analysis code

A coding agent that reads your repository, writes the pipeline, runs it and shows you the failing test is a different tool from a chatbot you paste snippets into. It is also what four in five of the quantitative social scientists Anthropic surveyed still weren’t doing weekly in March 2026.

What makes this safe is not trust, it’s the artefact. An agent hands you code and logs, and code and logs can be read, diffed, re-run and version-controlled by someone who was not in the room. A chatbot answer pasted into a manuscript cannot. Keep the agent on the side of the work that leaves a reviewable trail.

Domain models you didn’t train

Structural biology moved first. AlphaFold 3’s inference code now sits on GitHub under Apache 2.0, with the model parameters downloadable separately under their own terms of use. Those terms are the part to read: the Apache licence covers the code, not the weights. There’s also a hosted version at alphafoldserver.com for non-commercial use, with a narrower set of ligands and covalent modifications.

For microscopy, Cellpose-SAM is the default generalist segmenter, producing usable cell and nucleus masks across modalities with minimal tuning; a June 2026 update added DINOv3-based models and a revised Cellpose-SAM that throws fewer spurious masks in low-contrast regions. The licence is the catch nobody reads: every Cellpose model is trained on CC-BY-NC data, which is fine for a paper and a problem the moment your work is commercial.

Neither tool needs an ML course to run. Both reward knowing where they fail, which is the actual argument for taking one — and for AlphaFold specifically, EMBL-EBI’s free three-hour tutorial teaches exactly that: reading pLDDT and PAE, the stated strengths and limitations of AlphaFold 2, and what AlphaFold 3 and the AlphaFold Server can and cannot model.

Writing, submission and the two rules that matter

Elsevier asks for a separate AI declaration statement in the manuscript at submission, and that statement “will appear in the published article”. Basic checks of grammar, spelling and punctuation are exempt, as is assistive technology used for accessibility. AI tools may not be listed or cited as an author, because “authorship implies responsibilities and tasks that can only be attributed to and performed by humans.” The hard line is images: “AI tools must not be used to create or alter images that represent primary observed or experimental data that were not directly obtained in the research.”

The second rule cuts the other way, and it’s the one careful scientists trip over. Reviewers and editors “should not upload a submitted manuscript or any part of it into an AI tool”, because doing so may violate the authors’ confidentiality and proprietary rights. Pasting a manuscript you’re refereeing into a chatbot to help draft your report is a policy breach, not a productivity hack.

On grants, NIH’s NOT-OD-25-132 (17 July 2025) says NIH “will not consider applications that are either substantially developed by AI, or contain sections substantially developed by AI, to be” original ideas of the applicant — and the same notice caps each PI at six covered applications per calendar year from the 25 September 2025 receipt date. NIH says it uses AI-detection technology to screen applications. If AI is detected after an award, NIH may refer the matter to the Office of Research Integrity as possible research misconduct, and the grant can be terminated (verify).

Note what that risk profile actually is: not a rejection letter, a misconduct file. Sharpen your own prose with AI; don’t let it generate your Specific Aims.

Do not bother with the $3,600 three-day course

MIT Professional Education’s AI for Scientific Discovery is a real course taught by real MIT faculty (Wojciech Matusik and Hanspeter Pfister). It also costs $3,600 for three days, targets “entry- to mid-level professionals or researchers”, and asks only for a bachelor’s degree plus “a basic understanding of programming and data analysis, including python or similar tools”. That’s roughly $1,200 a day, for an entry-level prerequisite bar, on a page that doesn’t state what credential you walk away with. We have now checked the page on 30 July, 5 August and 6 August 2026. All three times the only cohort listed (8–10 June 2026) showed as closed, with no next date published.

With an institutional training budget and a room full of R&D peers, three days of MIT faculty is defensible. Out of pocket, at that price, against a free fast.ai course covering much of the same ground, it isn’t. Our read on the general “AI certification” badges is harsher still: we have found no evidence that journals, funders or academic search committees weigh them at all, and a preprint with your analysis code on GitHub does the job every one of them claims to. Our AI certifications guide sorts the credentials that carry weight from those that don’t.

The genuinely free path

Free, in this order, and you’ll be better equipped than most of your department. Google’s Machine Learning Crash Course for a conceptual base, then fast.ai Part 1 for real models, then the two-hour Claude Code course, then whichever domain tutorial your field publishes — EMBL-EBI’s AlphaFold guide if you touch protein structure. Add our free AI automation course if the wiring is what intimidates you. Total cost, zero; total certificates, also zero, which matters less here than in almost any other profession.

If that path runs out, two places to go next. Our free machine learning courses hub re-verifies every no-cost option monthly, so it catches the ones that quietly move behind a paywall; the best machine learning courses online ranks the paid ones by depth rather than by brand. Neither is scientist-specific. For the adjacent jobs, the rest of AI courses by profession and the wider free courses hub cover them.

Hit the ceiling of the free material and want structured training past the introductory tier? Towards AI runs practical AI programs our readers get a graduate discount on. Exhaust fast.ai and the free courses first; pay only once you know what you’re missing.

Moving out of academia, or hiring someone who can do this? Open AI and research roles are on our talent board.

FAQ

What is the best AI course for scientists in 2026?

fast.ai’s Practical Deep Learning for Coders, and it’s free: nine lessons of about 90 minutes, needing roughly a year of Python and high-school maths. For the foundations plus a certificate, audit the Stanford Online / DeepLearning.AI Machine Learning Specialization and pay $49 for the month you finish in. No subscription at all? NVIDIA DLI’s eight-hour course is $90 once.

Do scientists need to learn machine learning to use AI in research?

No. Running AlphaFold 3, Cellpose-SAM or a literature-screening tool needs no ML training. You need it to know where those models fail, to build something custom, and to referee papers that use them. Start with the tools; take a course when they stop being enough.

Can I use AI to write a grant application?

Carefully, and never to generate the science. NIH’s NOT-OD-25-132 says applications substantially developed by AI, in whole or in sections, won’t count as the applicant’s original ideas. NIH screens with AI-detection technology, and a post-award detection can be referred to the Office of Research Integrity as possible misconduct. Editing your own prose is fine; generating Specific Aims is not.

Do I have to disclose AI use in a paper?

Usually yes. Elsevier wants a separate AI declaration statement at submission, published with the article, exempting basic grammar and spelling checks and accessibility tools. AI cannot be an author. AI must not create or alter images representing primary data not directly obtained in the research. Check your target journal, because rules still differ.

Which AI tools for researchers are actually worth using?

The common 2026 stack: Consensus or Semantic Scholar for discovery, Elicit for structured extraction and systematic-review screening, Litmaps for citation mapping, a coding agent for analysis pipelines, and domain models such as AlphaFold 3 or Cellpose-SAM. Verify every citation any of them produce.