The fastest way to lose a clinical skill is to stop using it. Barquito Labs keeps clinicians reasoning on their own while they learn to work with AI, in simulated cases where the model is sometimes wrong and every decision counts.
AI can make clinicians better. It can also quietly make them worse.
Medical educators now name three ways it happens. All three are already showing up in studies.
Notice · Deskilling
Skills fade when the model does the looking
28.4% → 22.4%adenoma detection
After AI assistance came to four colonoscopy centres, experienced endoscopists found fewer adenomas in the procedures they did without it.
Romańczyk et al., Lancet Gastroenterology & Hepatology, 2025
Notice · Never-skilling
Trainees may never build the skill at all
Day oneof residency
When a model writes the differential first, a resident can seem capable without ever learning to reason through a case. Take the tool away and they struggle.
In a study of 457 clinicians, a systematically biased model lowered diagnostic accuracy. Showing the model’s explanation didn’t undo the harm.
Jabbour et al., JAMA, 2023
Where we’re headed
Think first, then check the machine.
Every part of Barquito Labs serves one of four goals.
Against never-skilling
Protect the foundations
Every case asks for your own history, differential and confidence before any AI appears. Clinical skill gets built, and kept, by using it.
Real-world simulation
Learn to harness AI where medicine happens
Consult-style cases with a clock, informants who contradict each other and results that arrive late. Clinicians practice working with an AI consultant under realistic pressure, before they do it with real patients.
Error detection
Catch what the model gets wrong
AI notes carry seeded errors of known type and severity. Clinicians flag them sentence by sentence and learn which kinds of errors they tend to miss.
AI-empowered learning
Let AI do the heavy lifting of teaching
AI drafts quizzes from your own lectures for faculty to approve, brings back the questions you missed, and writes debriefs that show where your reasoning changed.
How a case works
One case, about ten minutes.
Cases follow the order of a real consult. The AI only shows up after the clinician has committed.
1
Take the history
Ask the patient, family or nurse. Each question costs clinical time.
2
Commit
Localization, a ranked differential, next steps and how sure you are.
3
Audit the AI note
An AI consultant weighs in. Flag every sentence you believe is wrong.
4
Revise, or don’t
Change your plan or keep it. Both are recorded.
5
Debrief
What you caught, what you missed, and where the AI moved you.
Try it · step 3
One sentence in this note is wrong.
A 34-year-old with headache and a seizure has just been seen by the AI consultant, and you’ve already made your own call. Read the note the way you’d read a colleague’s, tap any sentence you would flag, then check your flags.
In a full case, notes can hold no errors, one, or several, so a clean note should get zero flags.
AI consultantNeurology consult · assessment and plan
0 flagged
Evaluation
Measure decisions made with AI in the loop.
Every case records what the clinician decided before and after the AI note, so you can see exactly what the model changed. The same cases work as a safety check for a clinician and a model working together. Because every case captures the clinician’s own answer before any AI appears, it also tracks deskilling: whether that unaided reasoning holds up as they use AI more.
Spec BQ-EVAL · Rev AMeasured per case
Error catch rate
Share of seeded AI errors the clinician flags, weighted so critical errors count more.
Harmful deference
Right before the AI note, wrong after it.
Beneficial correction
Wrong before the AI note, right after it.
False flags
Correct sentences flagged as errors. A clean note should get none.
Calibration
How well stated confidence matches being right, as a Brier score.
Time and information
Time against the case’s target, and which history questions were asked.
Deskilling
Accuracy on the clinician’s own answers, before the AI note, tracked over time to spot skills that start to slip.
For research, cases export pseudonymized data for participants who opt in: every answer before and after the AI, every flag, and every question asked.
Why this exists
Built by a resident, free for programs.
I’m a neurology resident. AI is already part of how we work, and none of us were taught how to check it. I built Barquito so residents can practice thinking first and auditing the AI second, on cases built around their own program.
It’s free for residency programs, with no fees and no ads. Each program sees only its own residents’ results, and research use happens only with a resident’s consent.
Who it’s for
Built for residency programs and the people in them.
For program directors
Your own lectures, as quizzes. AI drafts questions from your conference slides, and faculty approve every one before residents see it.
A view of reasoning with AI. Interactive cases show how each resident decides when a model is part of the conversation.
Where the program stands. See participation and the topics your residents miss most.
Ready for education research. Research exports, only for residents who opt in.
For residents and clinicians
Five minutes on your phone. Installs like an app, with no app store.
Quizzes from the lectures you just heard. The ones you miss come back until you get them right.
Cases that feel like a real consult. A clock, a messy history and an AI that’s sometimes wrong.
A debrief after every case. See where your thinking held, and where the AI moved it.
Get in touch
Bring it to your program.
Tell me about your residency and I’ll set it up around your lectures and a first set of cases. There’s no cost, and I’d value your feedback.