Deploying AI
Ethically & Safely
Putting AI where people use it, watching if it goes wrong later, and treating people fairly — explained simply.
↓ scroll to begin
Putting AI where people use it, watching if it goes wrong later, and treating people fairly — explained simply.
Training a model on your laptop is only half the story. Real people need a safe, fast way to send a photo or question and get an answer — every day, not just once in class.
What changes when you go live:
Before launch checklist (simple):
Figure — Build, launch, then keep checking it still works.
Figure — If any step fails, user sees error — not a blank screen.
Ways to share AI with people
| Way | What it means | Used for | Example |
|---|---|---|---|
| Website / app | User sends text or photo online | Homework helper, support chat | School study website |
| On device | Model runs on phone or robot | Fast, more private, offline sometimes | Keyboard word suggestions |
| API | Other apps call your AI over internet | One brain, many apps | Translation used inside many apps |
| Batch job | Runs overnight on many items | Reports, not instant chat | Score 10,000 forms by morning |
| Embedded in machine | Inside factory PC or camera | Stop belt when defect seen | Vision PC next to conveyor |
What to watch after launch
| Measure | Plain English | Why care |
|---|---|---|
| Speed | Seconds until answer | Users leave if too slow |
| Uptime | Is the service online? | Homework due at midnight |
| Error rate | How often it crashes | Trust drops fast |
| Complaints | “Wrong answer” reports | Early sign of drift or bias |
| Cost | Cloud bills for AI calls | Popular app can get expensive |
Write two sentences a 12-year-old would understand. What should NOT be shown (internal error codes)?
Drift means the real world moved but the model stayed the same. New slang, new phone cameras, new spam tricks — the old brain is still answering yesterday’s world.
Think of a spam filter trained in 2020. It learned old scam phrases. In 2026, scammers use new emoji patterns and new links. The filter still runs, but more spam slips through until someone retrains it on fresh emails.
Three simple types of drift:
What teams do: Watch simple charts, read user complaints, retrain on new data, or roll back to the last model that worked — like undoing an app update.
Figure — Gap grows when training data no longer matches today.
Figure — Same cycle as deployment: never “launch and forget.”
Signs you may need to retrain or fix
| Sign | What users notice | Example |
|---|---|---|
| More wrong answers | Complaints go up | Spam filter misses new emoji spam |
| New inputs | Photos or words look different | New phone camera colour style |
| Rules changed | Law or school policy shifted | Stricter privacy for student photos |
| Old software | Security holes, crashes | IoT camera never updated |
| Sudden “always yes” | Model says OK to everything | Broken sensor feeding zeros |
Who does what (small team view)
| Role | Job in plain English |
|---|---|
| Developer | Builds app, connects model, fixes crashes |
| Data person | Collects new labelled photos or text |
| Domain expert | Teacher, nurse, engineer — says if answers make sense |
| Operator | Watches dashboards, restarts service if down |
If training photos mostly show one skin tone, one age group, or one language, the system may work worse for everyone else. That is unfair — and sometimes against the law.
Bias usually comes from data and design choices, not because the computer “hates” someone. Examples:
Privacy: Photos, voice, and location are personal. Collect only what you need. Say why. Let people say no when possible.
Explainability: People ask “why did the AI say no?” Simple reasons help (e.g. “flagged as spam because link pattern”). Deep neural networks are harder to explain — another reason to keep humans in the loop for big decisions.
Figure — AI assists; human decides for loans, medical, discipline, policing.
Fairness ideas — simple checklist
| Question | Why it matters | Example |
|---|---|---|
| Who is in the training data? | Missing groups → worse results for them | Face unlock fails on some skin tones |
| Who gets hurt if wrong? | Pick human review for big decisions | Loan denied by mistake |
| Did we ask consent? | Photos and voice are personal | School hallway camera |
| Can we explain the answer? | Appeals and trust | “Spam because suspicious link” |
| Can user opt out? | Respect choice | Optional “smart” grading assist |
Privacy — types of data and care level
| Data type | Sensitivity | Care needed |
|---|---|---|
| Public web text | Lower | Still check copyright and bias |
| Student homework | High | School rules, parent consent, secure storage |
| Face image | Very high | Consent, limit storage time, secure servers |
| Health records | Very high | Law, doctor oversight, encryption |
| Location from phone | High | Explain why; allow off switch |
List two reasons yes and two reasons no. What could go wrong if the camera lighting is bad?
Laws and school or company policies ask: What is this system for? What must it never do alone? Who is responsible if someone is harmed? How can people complain?
Governance is not only lawyers — it is clear rules so teams do not ship risky AI by accident. For student projects, a one-page “model card” is enough practice.
High-risk areas (need extra care): health, policing, hiring, loans, exams, critical infrastructure. Often require human oversight, logs, and testing before wide use.
Incident book: When something goes wrong, write it down — what happened, who was affected, what you changed. Same idea as a school accident log.
Figure — Many people share responsibility for high-risk AI.
Model card — what to write down (one page)
| Section | Plain English — what to write |
|---|---|
| Purpose | What job is this AI for? (e.g. “flag possibly bruised apples”) |
| Not for | What it must never decide alone (e.g. “fire an employee”) |
| Training data | Where photos or text came from; how many examples |
| Known weak spots | Dark rooms, wet apples, cracked camera lens |
| Metrics | How accurate on held-back test set (simple %) |
| Contact | Who fixes problems; how to complain |
| Last updated | Date and version number |
High-risk uses — extra rules people expect
| Use area | Why extra care | Typical safeguard |
|---|---|---|
| Medical diagnosis | Life and health | Doctor decides; regulated testing |
| Police surveillance | Freedom and privacy | Law, oversight, audit logs |
| Exam scoring alone | Fairness for students | Human grades; AI only assists |
| Child monitoring | Vulnerable users | Parent consent, minimal data |
Purpose, not for, known weak spot. Share with a partner — can they explain your project back to you?
10 easy questions on using AI safely in the real world. Instant feedback on every answer.
Module 7 in short: ship carefully, watch for drift, and think about who gets hurt if the AI is wrong.