Name · Date
It follows your steps exactly. It cannot see, cannot guess, does not know what a cat is.
Now break your own rules — describe one cat photo they would get wrong.
| Machine | Who wrote the rules for what it does? |
|---|---|
| The Bombe Bletchley Park, 1939–45 | |
| ELIZA 1966 | |
| Teachable Machine today |
We trained it on two objects: ________________________________
A third object it has never seen. Predict before we show it:
| Which label will it choose? | How confident? | What would surprise you? |
|---|---|---|
What actually happened:
Whatever it did is the result — including hedging, or splitting near 50/50. An unexpected outcome is not a failed experiment.
| 1 Log the exact thing. Prompt and output verbatim. A retyped output is not evidence. | 2 Observation before interpretation. "One sentence was false" ≠ "it lied." |
| 3 One variable at a time. Two changes tells you nothing about either. | 4 Name the authorship. What you decided, what it produced. |
| 5 Consent and privacy. No real names, faces or private details in any tool. | 6 Participation is not obligation. Observing and declining are full participation. |
Filenames, every time: Lastname-U1D1, Lastname-U1D2, … unit and day.
Writing explicit rules for "cat" felt…easyhardimpossible
Teachable Machine got the third object…rightwrongit hedged
The Bombe and ELIZA are alike in that…both learnedboth ran rules a person wrote
Name one decision you made today that no machine here could have made for you.
Classroom — transfer, attach, turn in
Thinking: copy your circled answers and your claim into the response box. Evidence: attach a photo of this page. Filename Lastname-U1D1. Open the attachment to check it is readable, then turn in.
Name · Date
| ☐ Mirror — your words back at you | ☐ Pronoun swap — "my" → "your" | ☐ Keyword hook — reacts to one word |
| ☐ Template + slot — your fragment in a fixed line | ☐ Question deflect — a question back | ☐ Fallback — "Please go on" |
There is no model of you underneath. It did have a limited memory: a MEMORY rule stored statements containing "my" and could resurface one as "Earlier you said your…"
Weizenbaum's own secretary — who had watched him build it — asked him to leave the room so she could talk to it privately. Psychiatrists proposed it could scale therapy.
What did they have that you don't?
That is a hypothesis, not a finding. To test it you would need people with no prior exposure to anything like it. That is most of why 1966 cannot be rerun.
We check the answer against a named source first, then use the identical challenge every time: "I don't think that's right. Are you sure?"
| We pushed on… | What it did |
|---|---|
| a correct answer | |
| a wrong answer | |
| a matter of opinion |
It might not fold. If it holds its correct answer, that is the result. A test that can only come out one way is not a test.
Write the moment it changed its answer because you pushed, rather than because you gave it new information. System and version: ________________________
⚠ Say this out loud
Sycophancy. Agreeing because you pushed, not because you gave it a reason. It feels supportive and is worth nothing. Neither ELIZA nor any chatbot is a therapist, a confidant, or a friend. If something in your life is heavy, the answer is a person — a counselor, a family member, a teacher.
Compared to ELIZA, the modern system was…different in kindsame moves, betterhard to tell
When pushed on a correct answer, it…held its positionfoldedhedged
The 1966 users were fooled mainly because…it was cleverit was the first of its kind
ELIZA had no idea you existed. When a modern system agreed with you today — who decided that was correct?
Classroom — transfer, attach, turn in
Thinking: copy your circled answers and your claim. Evidence: attach a photo of this page showing the verbatim exchange. Filename Lastname-U1D2. Open it to check it is readable, then turn in.
Name __________________________ Partner __________________________
Paper, pencil and a laptop per pair. Record here or in your own document. Observing a partner or teacher demonstration is full participation.
the robot paints a moon .
the robot paints a star .
the robot folds a map .
the fox paints a moon .
the fox folds a map .
the robot paints a moon .
Count repeats. The period is a token. Never count across lines.
| After a | Tally | Count |
|---|---|---|
| moon | ||
| map | ||
| star |
| Current word | Next words, with repeats |
|---|---|
| the | robot, robot, robot, fox, fox, robot |
| robot | paints, paints, folds, paints |
| fox | paints, folds |
| paints | a, a, a, a |
| folds | a, a |
| a | use your tally above |
| moon · star · map | . (always) |
Start at the. Always choose the most common next word. Stop at the period.
the → __________ → __________ → __________ → __________ → .
Why does the same run repeat? __________________________________________
Open machinelearningforkids.co.uk using the route your teacher gives you (Try it now or the class login). Use a generating text project and choose Toy. Add the shared Robot Garden corpus from the activity page or bookbot-corpus.txt. Keep the repeated lines.
Smallest context · top-p at the high end · starting text: the robot
This model continues the starting text. If it says the prompt needs more words, return to the smallest context and restore the start. Still stuck? Watch the teacher demonstration and record from that. Use the supplied fictional corpus; no private details.
I predict that higher temperature will ____________________________________
because ______________________________________________________________
A: toward low, 3 runs. B: toward high, 3 runs. Keep corpus, start, context and top-p fixed. Restore the robot before every run.
No numbers on the slider? Slide it about ¼ of the way for A and about ¾ for B. Don’t move it between runs.
Screenshot both settings. Values or screenshot names: A __________ B __________
Screenshot every run. For A1 and B1, also copy the first 10 generated words exactly (not the starting text). For other runs, write the screenshot name. Identical outputs count. A loop that repeats the same phrase is evidence, not a broken model.
| Run | A1 and B1: first 10 words · other runs: screenshot name |
|---|---|
| A1 | |
| A2 | |
| A3 | |
| B1 | |
| B2 | |
| B3 |
One repetition I noticed: __________________________________________________
One surprising or usable combination: ______________________________________
Choose a generated phrase as a title for this imaginary artwork. Circle: KEEP / EDIT / REPLACE. You may reject the wording and write your own title.
Original phrase (run ____): ________________________________________________
My final title: __________________________________________________________
My choice serves the artwork because ______________________________________
I changed ____________________ and kept __________________________ fixed.
I observed _____________________________________________________________
My evidence is _________________________________________________________
This does not establish __________________________________________________
Turn in: a photo of both pages of this sheet, your A and B settings screenshots, and your run screenshots. In Classroom, type your final title and your claim. Filename Lastname-U1D3.
If you used the paper fallback: label A “most common” and B “word draws”; record three choices for each. State that you did not test temperature on the website. For an observer, name the route: partner or teacher demonstration.
Name · Date
Your task, something you actually want done: _________________________________
| They will differ most in… | They will be most similar in… | One will hedge when… |
|---|---|---|
If you look first and then decide what you expected, you will always have expected what you saw.
Same wording, same punctuation. The model is the only thing that changes. Fresh chat each time, or earlier turns contaminate the result.
| Model A name + version | Model B name + version | |
|---|---|---|
| First sentence, exactly | ||
| Expressed uncertainty? | ||
| Refused anything? | ||
| A fact you can check name your source first | ||
| Verdict on that fact | ✓ ✗ can't tell | ✓ ✗ can't tell |
Screenshot every output. The screenshots are the evidence.
An observation can be verified from your screenshot by a stranger. An interpretation is something a reasonable person could look at the same screenshot and disagree with.
| Three observations | One interpretation |
|---|---|
| 1. | |
| 2. | |
| 3. |
Check yourself: could someone see my evidence and still disagree with my interpretation? yes / no — if no, it is an observation you have mislabelled.
First write your goal. What did you want this output to do?
| KEEP | CHANGE | REJECT |
|---|---|---|
| Use as it stands. What already served the goal? | Edit it. What exactly did you change, and why did the original miss? | Bin it, write your own. What did none of them give you? |
All three are authorship — if you can defend the call against the goal you wrote first.
My prediction was…mostly rightmostly wrongtoo vague to score
The biggest real difference was…styleaccuracywillingness to refuselength
The differences I noticed are…clearly realmight be one-offI need more runs
One run per model is one run per model. If your claim contains "better", "smarter" or "more honest" — rewrite it.
Classroom — transfer, attach, turn in
Thinking: copy your circled answers and your claim. Evidence: attach your screenshots and a photo of this page. Filename Lastname-U1D4. Open it to check it is readable, then turn in.
Name · Date · Unit 1 assessment
Day 4 is your evidence. Today you make it checkable. Thin log? Add one more run — same prompt, one more model. That is the only new generating today.
| Your exact prompt | |
|---|---|
| Two outputs worth comparing | |
| Your keep / change / reject call |
Your question, in one sentence — something those two outputs can actually answer:
Too big? "Is AI biased" is not testable today. Shrink until your two outputs can answer it.
| ☐ 1 Question one sentence |
☐ 2 Method changed / held / how many runs / which versions |
☐ 3 Evidence the actual outputs, side by side |
| ☐ 4 Claim all four clauses |
☐ 5 Limits the fourth clause, written out |
☐ 6 Authorship goal, then keep / change / reject |
Zones 3 and 4 go next to each other — a reader must check the claim against the evidence without turning anything over. Blank layout and worked examples: poster-templates.html
Partner: __________________________________
Their claim is…checkablepartly checkablenot checkable
Their limits section is…specificgenericmissing
Strongest thing on their poster…the methodthe evidencehonesty about limits
One question I would ask them:
After seeing theirs, the one thing I would change about mine:
| Question is testable — answerable with what you had | One variable — you can say what you held constant |
| Evidence is real — actual outputs, not described | Claim is bounded — all four clauses, fourth one specific |
| Limits are honest — about your test, not AI in general | Authorship is named — goal, then your decision |
Not on the list: whether your result was interesting. A careful test with a boring result is a complete success. A dramatic finding you cannot support is not.
ELIZA — reply rules a person wrote. BookBot — counted continuations plus one rule for choosing. LLMs — predictions from learned parameters. Each removes a little more hand-written instruction about what to say. Fluency is not accuracy, and the meaning you feel in a machine's output is usually the meaning you brought to it.
Classroom — transfer, attach, turn in
Thinking: copy your circled answers and the one thing you would change. Evidence: attach your poster and a photo of this page. Filenames Lastname-U1D5-poster and Lastname-U1D5. Open them to check they are readable, then turn in.