Librarian Joss sets a thick folder on the library table. "The Tech Review Panel needs three student reviewers," she says. "Comet, Wren and you."
Comet flips open the cover and reads aloud. "SH-1, an AI study helper. Works for every class. Never makes things up." She grins. "Let's approve it today!"
Wren taps his clipboard. "What does the evidence say? A cover is a promise, not a test."
Nova hovers over the table and projects the packet's five sections onto the wall. "Would you like a hint?" she asks. "Start with what SH-1 is supposed to do."
Comet turns to you. "Reviewer, is SH-1 even an AI system? Where do we start?"
Harbor Point High School is the crew's school. An outside builder has offered the school SH-1, short for Study Helper, version 1.
The school's Tech Review Panel must decide whether to try SH-1 in a pilot. Over twelve weeks, the crew will review it for the panel.
Harbor Point, SH-1, the builder and every number in this review are a made-up example. The review skills you practice are real.
This course is unplugged. You need paper, a pencil and sometimes index cards or a die. You never use a real AI tool, app or account.
| Packet section | What the packet says (made up for the story) |
|---|---|
| P1 Purpose | "SH-1 answers study questions, writes practice quizzes from class notes, and checks your answers." |
| P2 How it works | "SH-1 is built on a large pre-trained language model from another company, tuned on study material." |
| P3 Training data | "Public text and study material." (No list of sources.) |
| P4 Test results | "92 out of 100 right on our benchmark." |
| P5 Privacy | "SH-1 keeps a log of every question with the student's name, class and time, to improve SH-1." |
| Cover claim | "Works for every class. Never makes things up." |
SH-1 has three features. Ask answers a question typed into a chat box. Quiz writes a practice quiz from notes a teacher uploads.
Check, short for "Check My Answer," marks a student's answer right or wrong.
NIST, a United States standards agency, wrote a careful definition of an AI system. Here it is in plain words.
An AI system is an engineered or machine-based system. For a set of objectives, it produces outputs such as predictions, recommendations or decisions.
Those outputs affect real or virtual places. AI systems also work with different levels of autonomy, meaning how much they act on their own.
| Part of the definition | Evidence about SH-1 from the packet |
|---|---|
| Engineered or machine-based | The builder made it on a large pre-trained language model (P2). |
| A set of objectives | Answer study questions, write practice quizzes and check answers (P1). |
| Outputs | Written answers, practice quizzes and right-or-wrong marks. |
| Some autonomy | Ask replies to a student in the chat box with no teacher reading each answer first. |
The definition names three kinds of output: predictions, recommendations and decisions.
Generative AI is the kind of AI that makes new "synthetic" content, such as images, video, audio and text. That content imitates the data it learned from.
So the crew sorts outputs into four kinds: prediction, recommendation, decision and content.
| Kind of output | Made-up example |
|---|---|
| Prediction | "The bus will arrive in about six minutes." |
| Recommendation | "You might like this mystery novel next." |
| Decision | "This answer is marked wrong." |
| Content | A brand new practice quiz written from today's notes. |
| Statement | True or false? |
|---|---|
| An AI system produces outputs for a set of objectives. | ? |
| Generative AI makes new content, such as text and images. | ? |
| Every AI system acts with exactly the same amount of autonomy. | ? |
| SH-1, Harbor Point and the builder are real. | ? |
| The cover claim "Never makes things up" has already been tested by the crew. | ? |
Strong start, reviewer. Tomorrow the crew looks at how AI drives software and things in the world.