AI Foundations for Healthcare (HAIC Track 1) — FN HAIC

FN HAIC · Certification Course

AI Foundations for Healthcare (HAIC Track 1)

All staff — from front desk to physicians

1.2h learner time 4 modules 16 lessons assessed · pass ≥ 80%

Curriculum

Module 1 of 4 — What AI is — and isn't
Why this module · AI is a pattern machine, not a knowledge base · Generative vs. predictive — know which one you're facing · Hallucinations: fluent, confident, and wrong
Module 2 of 4 — PHI & privacy: what never goes into an AI tool
Why this module · PHI is more than a name · The bright line: sanctioned vs. unsanctioned · Minimum necessary — the prompt-writing habit
Module 3 of 4 — AI you already use: recognizing it in your daily tools
Why this module · The AI that doesn't announce itself · Inside the EHR specifically · Same rules, quieter packaging
Module 4 of 4 — Safe-use habits: verification, escalation, and when to ask
Why this module · The 20-second verification pass · The three stop-and-ask situations · Report it like a safety culture

Module 01

Module 1 of 4 — What AI is — and isn't

In the next few minutes you'll be able to: - Explain, in plain language, what today's AI tools actually do - Tell a generative AI tool from a predictive one — and why the difference changes how much you trust each - Recognize a hallucination and know your responsibility when AI drafts something for you This module is the foundation for everything else in the program — including the rules about patient data that come next.

Lesson 1.1 · 3 min · reading

Why this module

In the next few minutes you'll be able to:

Explain, in plain language, what today's AI tools actually do
Tell a generative AI tool from a predictive one — and why the difference changes how much you trust each
Recognize a hallucination and know your responsibility when AI drafts something for you

This module is the foundation for everything else in the program — including the rules about patient data that come next.

Lesson 1.2 · 5 min · reading

AI is a pattern machine, not a knowledge base

The AI tools arriving in healthcare — ambient documentation, drafting assistants, chat tools — are prediction systems. They generate the most plausible next words based on patterns learned from enormous amounts of text. They do not look facts up in a verified database, and they do not know your patient.

> The one-sentence version: AI produces what is likely, not what is verified. Plausible and true are different properties — your judgment supplies the second one. That's not a reason to avoid these tools. Used well, they remove real drudgery. It's the reason every AI-assisted task keeps a human accountable at the end.

Lesson 1.3 · 5 min · reading

Generative vs. predictive — know which one you're facing

Generative AI creates new content: a draft note from an ambient recording, a suggested reply, a summary. Its failure mode is confident fabrication — output that reads perfectly and is wrong.

Predictive AI scores or flags: a sepsis risk score, a deterioration alert, a scheduling forecast. Its failure mode is silent bias or drift — it can be systematically off for certain patients without looking wrong at all.

> Why it matters: you proofread generative output before it enters the record; you interpret predictive output as one signal among several — neither is a decision made for you.

Lesson 1.4 · 5 min · reading

Hallucinations: fluent, confident, and wrong

Because generative AI optimizes for plausibility, it sometimes produces details that were never in the source: a medication the patient never mentioned, a finding nobody documented, a citation that doesn't exist. This is a hallucination — and it arrives in the same fluent, confident tone as everything else.

Your protocol when AI drafts anything you'll sign or send:

Verify against the source — the recording, the chart, the conversation. Never against your memory of them.
Assume the polish proves nothing. Fluency is what the tool is best at — including when it's wrong.
Your signature means you checked. An AI-drafted note you sign is your note, fully.

Module 1 Assessment — pass ≥ 80%

1. An ambient documentation tool produces a beautifully written note that includes a medication allergy the patient never mentioned. What happened?

The tool accessed the patient's old records
A hallucination — the model generated a plausible detail that wasn't in the source
A software bug that IT needs to patch

Reveal answer

b) A hallucination — the model generated a plausible detail that wasn't in the source — The module's standard: A hallucination — the model generated a plausible detail that wasn't in the source

2. A deterioration-risk score seems consistently low for one group of patients on your unit. Which AI failure mode is this?

Hallucination
Prompt error
Silent bias or drift in a predictive model

Reveal answer

c) Silent bias or drift in a predictive model — The module's standard: Silent bias or drift in a predictive model

3. You sign an AI-drafted note after a quick skim because it reads flawlessly. What does this module say about that?

Fluency proves nothing — signing means you verified it against the source
Acceptable — the tool is validated by the organization
Acceptable if you flag it as AI-assisted

Reveal answer

a) Fluency proves nothing — signing means you verified it against the source — The module's standard: Fluency proves nothing — signing means you verified it against the source

Module 02

Module 2 of 4 — PHI & privacy: what never goes into an AI tool

This is the most important module in the program.

Lesson 2.1 · 3 min · reading

Why this module

This is the most important module in the program. By the end you'll be able to:

Recognize protected health information (PHI) in all its forms — not just names
Apply one bright-line rule before typing anything into any AI tool
Tell a sanctioned tool from an unsanctioned one, and why the same sentence can be fine in one and a violation in the other

Lesson 2.2 · 5 min · reading

PHI is more than a name

PHI is any health information that can be tied to a person — and the tying takes less than you'd think. Names, yes, but also dates of service, room numbers, ages over 89, rare-diagnosis-plus-town combinations, a photo of a wristband in the background. De-identifying is harder than deleting a name.

> The test: could a determined person, holding this text and public information, figure out who the patient is? If maybe — it's PHI. That vivid case you want advice about is often identifiable to anyone in your unit precisely because it's vivid.

Lesson 2.3 · 5 min · reading

The bright line: sanctioned vs. unsanctioned

The question is never just what you type — it's where. A sanctioned tool is one your organization has vetted and contracted: a business associate agreement in place, data handling audited, retention understood. An unsanctioned tool is everything else — including the excellent free chatbot on your phone.

> The bright line: PHI may only ever enter tools your organization has explicitly approved for PHI. Not tools that are popular, paid-for, or 'probably fine' — approved for PHI, on the published list. The same prompt — 'summarize this discharge note' — can be routine in the sanctioned tool and a reportable breach in the free one. If you don't know whether a tool is on the list, that is your answer until you check.

Lesson 2.4 · 5 min · reading

Minimum necessary — the prompt-writing habit

Even inside sanctioned tools, good practice is to include only what the task needs. Asking for help wording patient education about insulin? The tool needs the topic and reading level — not the patient's name, MRN, or history. This is the old 'minimum necessary' principle wearing new clothes.

Strip identifiers by default — most drafting and explaining tasks don't need them.
Never paste whole documents when a paragraph does the job.
When you slip — and people do — report it the same day. Every organization has a path for this, and early reporting is the difference between a correction and a crisis.

Module 2 Assessment — pass ≥ 80%

1. You want advice on a complex case and type it into a free chatbot, carefully removing the patient's name first. The case involves a rare condition, the patient's approximate age, and your hospital's name. Is this PHI?

No — the name was removed
Yes — the combination is likely identifiable, and the tool isn't approved for PHI anyway
Only if the patient complains

Reveal answer

b) Yes — the combination is likely identifiable, and the tool isn't approved for PHI anyway — The module's standard: Yes — the combination is likely identifiable, and the tool isn't approved for PHI anyway

2. Your organization's approved AI tool and a popular free chatbot both draft equally good letters. What makes the approved tool different for PHI?

It writes more accurately
Vetting, a business associate agreement, and audited data handling
It's faster

Reveal answer

b) Vetting, a business associate agreement, and audited data handling — The module's standard: Vetting, a business associate agreement, and audited data handling

3. You accidentally pasted a full patient note into an unapproved tool. What does this module say to do?

Delete the chat and move on
Report it the same day through your organization's process
Only report it if someone noticed

Reveal answer

b) Report it the same day through your organization's process — The module's standard: Report it the same day through your organization's process

Module 03

Module 3 of 4 — AI you already use: recognizing it in your daily tools

AI didn't arrive at your organization the day someone bought a chatbot license — it's been seeping into your existing tools for years.

Lesson 3.1 · 3 min · reading

Why this module

AI didn't arrive at your organization the day someone bought a chatbot license — it's been seeping into your existing tools for years. By the end of this module you'll be able to:

Spot the AI already embedded in your EHR, inbox, and dictation tools
Explain why 'invisible' AI still needs the same verification habits
Name the features your organization has actually turned on — and where to find out

Lesson 3.2 · 5 min · reading

The AI that doesn't announce itself

Autocomplete in your email. Smart phrase suggestions in the EHR. Dictation that cleans up your grammar. Priority sorting in your inbox. Chart search that 'understands' what you meant. None of these shipped with an AI badge — all of them are prediction systems making choices about what you see and what gets written.

> The shift to notice: the question isn't 'do we use AI?' — you do. It's 'which of the hundred small suggestions I accepted today were machine guesses, and did any of them matter?'

Lesson 3.3 · 5 min · reading

Inside the EHR specifically

Modern EHRs ship with AI features your organization turns on selectively: draft replies to patient messages, visit summaries, coding suggestions, ambient documentation integrations, risk flags on your patient lists. Two things follow:

Your build is unique. What's live at your organization is a configuration decision — the feature you saw at a conference may be off here, and vice versa.
Suggestions inherit your authority. A machine-drafted patient reply sent under your name is your reply. The drafting moved; the accountability didn't.

> Find the list: your informatics or training team maintains what's enabled. Knowing your build is part of being AI-literate here.

Lesson 3.4 · 5 min · reading

Same rules, quieter packaging

Embedded AI deserves the same habits as the obvious kind — it's just easier to forget because it looks like the software you've always used.

Autocompleted text is generated text. Read the whole sentence before it enters a record; the last half may not say what you meant.
Sorting and flagging shape attention. If the inbox ranks a message low, that's a prediction, not a triage decision — spot-check what the machine buries.
Convenience is the camouflage. The more seamless the feature, the more deliberate your verification habit needs to be.

Module 3 Assessment — pass ≥ 80%

1. Your EHR suggests a complete sentence while you chart, and you accept it with one keystroke. What is true?

Nothing changed — it's just autocomplete
That was generated text, and it carries your authority once accepted — read it whole
The EHR vendor is accountable for it

Reveal answer

b) That was generated text, and it carries your authority once accepted — read it whole — The module's standard: That was generated text, and it carries your authority once accepted — read it whole

2. A colleague says your organization 'hasn't rolled out AI yet.' Based on this module, what's the most accurate response?

Correct — no chatbot has been purchased
Embedded AI is already in the EHR, dictation, and inboxes — the rollout happened quietly
AI only counts if it's called AI

Reveal answer

b) Embedded AI is already in the EHR, dictation, and inboxes — the rollout happened quietly — The module's standard: Embedded AI is already in the EHR, dictation, and inboxes — the rollout happened quietly

3. Where do you find out which AI features are actually enabled in your organization's EHR build?

The vendor's public website
Your informatics or training team's enabled-features list
Trial and error

Reveal answer

b) Your informatics or training team's enabled-features list — The module's standard: Your informatics or training team's enabled-features list

Module 04

Module 4 of 4 — Safe-use habits: verification, escalation, and when to ask

The final module turns the first three into habits.

Lesson 4.1 · 3 min · reading

Why this module

The final module turns the first three into habits. By the end you'll be able to:

Run a 20-second verification pass on any AI output before you use it
Recognize the three situations where the right move is to stop and ask
Report an AI problem so it actually gets fixed — instead of becoming everyone's private workaround

Lesson 4.2 · 5 min · reading

The 20-second verification pass

Verification doesn't mean re-doing the work — it means checking the parts that carry risk. Before using any AI output:

Names, numbers, negations. Doses, dates, lefts-and-rights, and every 'no/not' — these are where fluent text hides consequential errors.
Anything you didn't say. Scan for content that wasn't in your source. Extra 'facts' are the signature of hallucination.
The claim you'd have to defend. If a sentence would matter in front of a colleague, a patient, or a lawyer, confirm it against the source.

> Habit, not heroics: twenty focused seconds on the risky parts beats five distracted minutes re-reading everything.

Lesson 4.3 · 5 min · reading

The three stop-and-ask situations

AI literacy includes knowing when not to proceed. Stop and ask a human — your supervisor, informatics, or compliance — when:

You're about to enter information you're not sure is allowed. The check takes a minute; a breach report takes a season.
The output disagrees with your judgment. Don't average the two. Disagreement between you and a machine is a signal worth thirty seconds of thought, not a coin flip.
You're being asked to use a tool nobody sanctioned. Including by a well-meaning colleague who found something great. Enthusiasm is not approval.

Lesson 4.4 · 5 min · reading

Report it like a safety culture

When an AI tool produces something wrong, biased, or unsafe, the organization can only fix what it hears about. Healthcare already knows how to do this — it's the same muscle as medication-error reporting.

Capture the specifics: which tool, what input (no PHI in the report itself), what it produced, what should have happened.
Use the channel your organization designates — and if none exists yet, that's worth raising too.
No workarounds culture: silently distrusting a tool and privately routing around it leaves your colleagues exposed to the failure you already found.

> The mindset: reporting an AI failure isn't complaining about software — it's the same professionalism as flagging a mislabeled sample.

Module 4 Assessment — pass ≥ 80%

1. You have 20 seconds to verify an AI-drafted message before sending. Where do they go?

Read the greeting and closing
Names, numbers, negations, and anything the AI added that you didn't say
Check spelling and tone

Reveal answer

b) Names, numbers, negations, and anything the AI added that you didn't say — The module's standard: Names, numbers, negations, and anything the AI added that you didn't say

2. An AI suggestion contradicts your professional judgment. The module says:

Defer to the AI — it has seen more data
Split the difference
Treat the disagreement as a signal: stop, think, and ask if needed — don't average

Reveal answer

c) Treat the disagreement as a signal: stop, think, and ask if needed — don't average — The module's standard: Treat the disagreement as a signal: stop, think, and ask if needed — don't average

3. You discover an AI tool consistently mis-summarizes a type of document, so you quietly stop using it for those. What's missing?

Nothing — you solved your problem
A report through the designated channel, so the failure you found doesn't keep reaching colleagues
A social media post warning others

Reveal answer

b) A report through the designated channel, so the failure you found doesn't keep reaching colleagues — The module's standard: A report through the designated channel, so the failure you found doesn't keep reaching colleagues

Try a full course, right now.