Forty minutes on what AI is actually doing in BIO5 labs in 2026 — then twenty minutes of hands-on practice on the U of A's own platform. You'll leave with one prompt you'll use in your lab this week.
Bioscience, biomedical engineering, biotechnology, statistics, and computational biology. No coding background required. Some high-school biology helps — the examples translate to whatever your lab does.
| Block | Duration | Activity |
|---|---|---|
| 1 | 5 min | Why this lesson exists |
| 2 | 10 min | What generative AI actually is |
| 3 | 15 min | What it's useful for in your lab |
| 4 | 10 min | What it's risky for — and the policies that protect you |
| 5 | 20 min | Hands-on at genai.arizona.edu — four exercises, five minutes each |
AI didn't replace researchers in 2024–2026. It changed the daily work of nearly every working scientist — including, almost certainly, your KEYS mentor.
A task where reaching for an AI assistant genuinely speeds you up — drafting, summarizing, debugging — without costing you accuracy.
A task where the AI will hand you a confident, plausible answer that is wrong — and you won't catch it. Citations, protocols, statistics.
A task where pasting the data would violate your lab's policies, your mentor's trust, or federal law. Unpublished data, patient info, IP.
That's the whole trick — and understanding it tells you exactly where these systems are brilliant and exactly where they fall apart.
A pattern-recognition system trained on enormous amounts of text. Given everything so far, it predicts the most likely next word. Then the next. Then the next.
Models like Claude, GPT, Gemini, Llama, and Gemma are not databases. They do not "look things up." They generate plausible continuations of your text.
They're so good at next-word prediction that the output feels like reasoning. Useful — as long as you remember what's actually happening underneath.
Want the deeper version — transformers, training, foundation vs. fine-tuned models? See the AI Landscape lesson. For today, "next-word predictor trained on text" is enough.
Four kinds of work you'll actually do this summer — and where an AI assistant genuinely pulls its weight in each.
A chat interface like genai.arizona.edu is the right starting point. For deeper AI-assisted coding, see the Vibe Coding lesson.
DeepMind. Protein structure prediction from sequence. The 2024 Chemistry Nobel.
Meta. Protein language models — structure and function straight from amino-acid sequence.
Recursion · Isomorphic · Insilico. AI generates and screens candidate molecules.
Hundreds of FDA-cleared AI/ML medical devices — mostly in medical imaging.
Evo · Borzoi. Systems that learn directly from DNA sequence at scale.
AI assists hypothesis & pattern recognition. Experiments validate.
AI assists hypothesis generation and pattern recognition. Experiments validate.
A model that predicts a protein structure or proposes a candidate molecule still needs wet-lab or clinical confirmation before anyone trusts it. Your KEYS project is the experiment side of that loop.
Four failure modes — and the U of A policies that exist to protect you and your mentor's research from the worst of them.
Specific paper titles, authors, journals, and DOIs that look completely real — and are not.
Concentrations, incubation times, temperatures, buffer recipes — invented out of thin air.
Test results, p-values, and sample sizes pulled from nowhere and stated as fact.
FERPA-aligned. Your conversations are private and not used to train models. When you need to work with sensitive material, this is where you go.
Working with human-subjects data? Your project has IRB rules. Ask your mentor first.
Many journals and fellowships now require disclosure of AI use. Your KEYS deliverable may too — ask before you lean on it heavily.
Research finds that disclosing AI use can make people trust you less. Disclose anyway — undisclosed use discovered later is far worse than the honesty hit.
AI is not your mentor, not the experimental method, not your scientific judgment. Treat it as a fast junior collaborator — occasionally wrong, with no stake in the outcome.
You still own the conclusions. Your name goes on the poster.
Four exercises, five minutes each, on genai.arizona.edu. Log in with your NetID. The point is to feel what we just covered — especially the hallucination test.
An OpenWebUI interface that gives you free access to multiple frontier models — Claude, OpenAI, Gemma, Meta Llama, Amazon Nova — through a single chat window. Conversations are private and are not used to train models.
Switch models from the dropdown at the top — central to Exercise 1.
FERPA-aligned; your chats are not training data.
No personal API key, no subscription. Just your NetID.
Pick a topic from your KEYS lab. Ask one model the prompt on the right. Then switch models (selector at the top) and ask the exact same prompt.
Ask the model for three papers supporting a claim from your field. Then paste each DOI into doi.org. Some resolve to the paper described. Some redirect elsewhere. Some will not exist at all.
Paste a short excerpt of a protocol you're learning this week — nothing your mentor flagged as confidential. Ask it to flag risks and silent-failure points.
Context · Role · Action · Format · Tone. Run it, then change one piece and re-run — notice how much output quality shifts from small changes.
AI is a fast junior collaborator — useful, occasionally wrong, no stake in your results. Let it draft, summarize, and explain. Then you read the abstract, check the DOI, run the experiment, and put your name on the poster.