English:Learning systems and human-set objectives - AI Year 6
Introduction

The same examples can teach different patterns when people change the goal or labels. Compare two review objectives, deliberately test two inconsistent labels and trace what actually changes. Keep the whole system and its users in view; no real messages or pupil accounts are used.
People define the task and the whole system
A useful purpose comes before a model
A learning system does not independently decide what a school or community ought to value. People define a task, select examples, choose labels and decide how outputs will be evaluated and used. A fitted model may discover a pattern people did not explicitly write as a rule, but its learning setup and surrounding product still reflect human decisions.
Consider a spam filter trained from examples labelled spam or not spam. Before counting correct predictions, someone must define what those labels mean for the intended users. Advertising, an unfamiliar sender and a dangerous message are not interchangeable categories. A system that predicts one of them cannot automatically claim to detect all the others.
Our fictional message-card project uses no real messages, email addresses, links or pupil data. Its output is a queue suggestion for an adult: REVIEW or KEEP. KEEP means do not route under this particular worksheet rule; it does not mean safe, true or approved for children. REVIEW does not mean the author has done something wrong. No message is deleted or sent anywhere.
CodeAI: AI: What is Machine Learning? With teacher support, identify who supplies labels and what a model can learn from them. The publisher's broader unit is for older grades; our Year 6 activity uses concrete original records, a finite program and paper tracing. No live pupil AI service is required.
Map the whole workflow
Start with the users' need, then define labels, collect permitted examples, check data, fit a candidate, test it and decide what action is proportionate. Put a human review and correction route after the output. A correctly running model is only one part of that workflow: unclear labels or an interface that hides uncertainty can still lead to a poor outcome.
| Stage | Human choice | What could go wrong |
|---|---|---|
| Purpose | Choose what should enter the review queue | Confuse advertising with danger |
| Data | Select fictional or properly authorised examples | Use private messages without approval |
| Labels | Write the exact target definition | Mix incompatible definitions |
| Fitting | Choose candidate rules and scoring method | Optimise the wrong recorded target |
| Evaluation | Use the right reference and new cases | Treat a training score as unseen accuracy |
| Action | Give an adult evidence and a way to disagree | Delete a message automatically |
Define two different objectives on the same cards
Our feature promo is one when the teacher's invented card includes a promotional cue and zero otherwise. Unknown_sender is one when its fictional sender is outside the supplied known-sender list. Both are simplified supplied annotations, not a real text or identity detector. The target is one for REVIEW and zero for KEEP.
Objective A routes promotional cards for review. Objective B routes cards from an unknown sender for review. Neither definition is a complete spam, truth or safety detector. These are deliberately different review tasks. Changing the objective changes the meaning of a correct label even when the visible cards stay the same.
Diagnostic: a promotional card from a known sender has promo one and unknown_sender zero. Should A and B give the same queue suggestion? Explain the two purposes before answering. If necessary, trace the two feature flags on paper and write the target definitions beside them. Do not choose a target merely because the word AI sounds authoritative.
Compare objectives with a controlled label experiment
Inspect six original training records
The following table is a complete invented training worksheet, version V1. No commercial mail corpus or real account was used. A and B are separate human-defined target columns, not two independent measurements of one universal truth. Keep the definitions visible while tracing the examples.
| Card | promo | unknown_sender | Target A | Target B |
|---|---|---|---|---|
| M1 | 0 | 0 | 0 | 0 |
| M2 | 1 | 0 | 1 | 0 |
| M3 | 0 | 1 | 0 | 1 |
| M4 | 1 | 1 | 1 | 1 |
| M5 | 1 | 0 | 1 | 0 |
| M6 | 0 | 1 | 0 | 1 |
M2 and M5 have equal features but are different invented cards. Equal feature values alone are not proof that one record was accidentally copied. M3 and M6 likewise show the same feature combination twice. All four possible binary feature combinations appear, although two appear more often. The six-row fitting scores therefore weight those combinations differently from a four-case exhaustive feature check.
Fit candidates instead of pretending the model chose the purpose
We offer three candidate rules: promo predicts the promo flag; sender predicts the unknown_sender flag; keep always predicts zero. Fitting counts matches with the selected training target and chooses the highest-scoring candidate. A tie is resolved by the fixed candidate order promo, sender, keep. Choosing that rule family, score and tie order is itself a human design choice.
For target A, promo matches six records, sender two and keep three, so fitting chooses promo. For target B, the scores are two, six and three, so it chooses sender. The data features did not change. The target definition changed which pattern scored best. This is a tiny data-fitted classifier, not a large language model or an operational message filter.
Trace M1: both objectives KEEP. M2: A REVIEW, B KEEP. M3: A KEEP, B REVIEW. M4: both REVIEW. M5 repeats M2's feature-based decisions and M6 repeats M3's. A difference is not automatically a software bug: first check whether the two systems are answering the same question.
Introduce two inconsistent labels and inspect the consequence
Keep objective A's written definition unchanged, but deliberately change its M2 and M5 recorded targets from one to zero. The altered target column is zero, zero, zero, one, zero, zero. This is an original label-quality experiment, not a claim about real annotators. The two changed labels contradict the declared promotional-review policy.
On these inconsistent labels, promo matches four rows, sender four and keep five. Fitting now chooses keep. It runs correctly and gets five of six matches against the corrupted training labels, while missing all three review cases under the original clean A definition. A good score against a bad reference does not establish the intended task was achieved.
Keep the clean and changed versions separately. Investigate the definitions and source records rather than fixing the table by a blind majority vote. A genuine policy change would require a new documented target and evaluation; an accidental label error requires a documented correction. Those are different explanations for a changed result.
Execute or trace the complete finite experiment
The following Python program was actually executed during course preparation. It selects the three candidates on the two clean targets and the deliberately inconsistent one. It uses no online service, reads no emails and cannot take a real message action. Learners can trace it on paper or use a teacher-approved Python environment.
def predict(model, promo, unknown_sender):
if model not in ("promo", "sender", "keep"):
raise ValueError("unknown model")
if any(v is not None and (type(v) is not int or v not in (0, 1))
for v in (promo, unknown_sender)):
raise ValueError("features must be integer zero or one")
if promo is None or unknown_sender is None:
return None
if model == "promo":
return promo
if model == "sender":
return unknown_sender
return 0
def fit(features, labels):
if not features or len(features) != len(labels):
raise ValueError("nonempty matched training records required")
for (p, u), label in zip(features, labels):
if p is None or u is None:
raise ValueError("training needs known features")
predict("promo", p, u)
if type(label) is not int or label not in (0, 1):
raise ValueError("labels must be integer zero or one")
candidates = ("promo", "sender", "keep")
scores = {
model: sum(predict(model, p, u) == y
for (p, u), y in zip(features, labels))
for model in candidates
}
best = max(candidates, key=lambda model: scores[model])
return best, scores
features = [(0, 0), (1, 0), (0, 1), (1, 1), (1, 0), (0, 1)]
labels_a = [0, 1, 0, 1, 1, 0]
labels_b = [0, 0, 1, 1, 0, 1]
inconsistent_a = [0, 0, 0, 1, 0, 0]
for name, labels in [("A", labels_a), ("B", labels_b),
("inconsistent_A", inconsistent_a)]:
model, scores = fit(features, labels)
print(name, model, scores)Expected selections: A chooses promo with scores six, two, three; B chooses sender with scores two, six, three; inconsistent_A chooses keep with scores four, four, five. Dictionary keys print in the declared candidate order. Unknown prediction features produce None, which this worksheet displays as ASK; invalid known flags are rejected rather than silently coerced. Training requires known features and valid target flags.
A boundary and an exhaustive feature check
Compare known-sender promotional input one/zero with unknown-sender promotional input one/one. A stays REVIEW; B changes from KEEP to REVIEW. Now compare unknown_sender zero with unknown_sender missing. The missing value needs ASK, not an assumption that the sender is known. That is an input-quality boundary, separate from changing the target.
Enumerating the four possible feature pairs is a complete test of this tiny binary input space. It is not an unseen real-message evaluation: every feature combination already appeared in training, and real messages contain much more information. A new fictional trial with uncertain promotional wording remains proposed and unperformed here.
Evaluate the intended outcome and retain oversight
Define positive before reading an error table
The credited confusion-matrix diagram puts actual/reference classes in columns and predicted classes in rows. For this worksheet only, take positive to mean REVIEW under a named target definition. A false positive routes a card whose target is KEEP; a false negative misses a card whose target is REVIEW. Changing the reference from A to B changes which outputs count as errors.
The diagram supplies four categories, not our six-card measurements. In the clean A experiment, the selected promo rule has three true positives and three true negatives on its fitting records, with no mismatches. These are training results. The always-KEEP baseline has three true negatives and three false negatives against clean A. It does not detect danger, truth or real spam.
Evaluate the changed model against the intended reference
The model chosen from inconsistent A labels always keeps cards. Against the altered labels it matches five of six; against the original clean A targets it matches only three of six. Both calculations can be arithmetically correct while answering different questions. Record which target version and denominator you used.
For the four feature pairs zero/zero, zero/one, one/zero and one/one, clean A outputs KEEP, KEEP, REVIEW, REVIEW. Clean B outputs KEEP, REVIEW, KEEP, REVIEW. The inconsistent-A fitted keep rule outputs KEEP for all four. This controlled comparison shows a causal effect of changing the supplied labels within our finite program; it does not establish how any real product was trained or how it behaves.
Before reading the worked checks, mark the target column you are evaluating. Partial step: M2 has target A one and target B zero. Finish the error comparison for M3 and explain why a supposedly high score without a target definition is incomplete evidence. Then state what evidence you would need to settle a real disputed label.
Do not confuse a proxy with the real aim
A measurable target is often only a proxy for a broader purpose. For example, routing an invented promotional cue is not the same as protecting every learner from harm. Increasing the count of held messages could make a queue unusable, while keeping everything could hide cases that need attention. Assess the complete process and people affected, not only a convenient numerical score.
Human review needs the target definition, source record, uncertainty and reason for routing, plus time and authority to disagree. A reviewer should be able to correct a mistake or choose a simpler method. An interface that presents KEEP as guaranteed safe would misrepresent our experiment even if the underlying prediction code were unchanged.
CodeAI: AI: Training Data & Bias. With teacher support, ask how examples and labels shape learned patterns and whose conditions might be absent. This video does not certify our fictional policy as fair or a real service as suitable. A high aggregate match count does not settle the consequences of errors for different users.
Transfer with a new purpose and an evidence log
Suppose a fictional library wants to recommend books. Predicting whether someone clicks a cover is different from helping them find a suitable book they understand and enjoy. Define the intended user benefit, features, target and evidence again. A click is observable behaviour, not an automatic measurement of learning or wellbeing. Do not reuse message-filter labels for a book task.
Keep this investigation paper-first and fictional. No pupil inboxes, sender addresses, private reading histories or real moderation decisions are required. A separate live school AI activity needs staff approval and appropriate age, supervision, filtering and monitoring checks. The present program supplies no authority to connect to a real account or delete anything.
Retain target definitions A/B, the six original records, both changed labels, all candidate scores, chosen models, boundary results, baseline and human fallback. State two limits: tiny simplified features and a reference defined by people. Propose appropriate new data/evaluation and source/label review as responses. Keep an uncertain-wording trial explicitly untested. This AI Year 6 route is editorial enrichment, not an independent statutory AI requirement or real message-filter certification.
Interactive Tasks
Quiz: Test Your Knowledge
Who defines what the learning task should achieve? (People choosing the purpose targets and evaluation) (!The model independently inventing what good means) (!The image filename) (!The highest training score without a target)
What does REVIEW mean in this fictional project? (Route a card for adult review under the stated definition) (!Prove its author did something wrong) (!Delete a real email) (!Certify the content is dangerous)
Which candidate wins on clean target A? (promo) (!sender) (!keep) (!No candidate)
Which candidate wins on clean target B? (sender) (!promo) (!keep) (!A real spam service)
Why do M2 and M3 differ under A and B? (The target definitions ask different questions) (!The feature table secretly changed) (!Every unfamiliar sender is dangerous) (!The program changed the learner age)
Which A labels are deliberately changed? (M2 and M5) (!M1 and M4) (!M3 and M6) (!All six labels)
Which candidate wins on the inconsistent A labels? (keep with five matches) (!promo with six matches) (!sender with six matches) (!None with no matches)
How many clean-A targets does that keep model match on the six rows? (Three) (!Five) (!Six) (!Zero)
What should happen when the sender feature is unknown? (ASK rather than silently assume a known sender) (!Automatically use zero) (!Declare the message safe) (!Read a real pupil inbox)
Why is the four-pair check not real unseen-message accuracy? (Those feature combinations already appeared and the task is synthetic) (!Four cases can never be checked) (!All target labels are universal truths) (!A confusion matrix supplies new emails)
Memory Game
| Objective | The task or outcome the system is set up to pursue |
| Feature | An input available for the stated prediction |
| Target | The reference outcome supplied for learning or evaluation |
| Proxy | A measurable stand-in that may miss a broader aim |
| Baseline | A simpler comparison method |
| Oversight | People can inspect evidence and change or reject an action |
Drag and Drop
| The same cards with a new target definition | Re-evaluate the purpose labels and scores |
| Two labels contradicting an unchanged policy | Retain versions and investigate the labelling problem |
| A high match count on a corrupted reference | Compare again with the intended target |
| An unknown sender feature | Record ASK rather than invent a known sender |
| A model output displayed as guaranteed safety | Correct the interface and retain human authority |
Crossword Puzzle
| TARGET | What names the reference outcome used for learning? |
| PROXY | What is a measurable stand-in for a broader aim? |
| LABEL | What recorded category can be inconsistent with its definition? |
| MODEL | What fitted component is only part of the whole system? |
| HUMAN | What kind of reviewer can challenge the output? |
| PURPOSE | What should be defined before choosing the technology? |
LearningApps
Optional extra practice. Choose an appropriate activity with your teacher. If the external collection is unavailable, use the native activities and independent tasks in this course.
Cloze Text
Open Tasks
Easy
- Name the human choices: Identify purpose, labels and one evaluation choice in the workflow.
- Read two targets: Explain A and B for a known-sender promotional card.
- Trace one missing feature: Distinguish sender zero from an unknown sender value.
- Read the error diagram: Name the actual/reference axis, prediction axis and meaning of positive for target A.
Standard
- Calculate candidate scores: Trace all six records separately for A and B and select each model.
- Audit changed labels: Keep A fixed, change M2/M5 as stated and recompute the three scores.
- Check all feature pairs: List four pairs and three sets of model outputs without calling them real unseen tests.
- Keep a complete evidence log: Name target versions, counts, boundary, baseline, limits, untested condition and reviewer.
Advanced
- Challenge a high score: Compare five of six against inconsistent labels with three of six against clean A.
- Redesign a real-use interface: Explain why KEEP must not be displayed as guaranteed safe and how a reviewer can intervene.
- Transfer to book suggestions: Separate predicting clicks from the broader learning benefit and define new evidence needs.
- Plan a new evaluation: Add an uncertain-wording case as a proposed test, with source/label review and a human fallback.
Learning Check
- Purpose and workflow: Correct the claim that a model decides what good means.
- Clean training scores: Compute the three candidate scores for both targets.
- Inconsistent-label experiment: Change M2/M5 in A and trace the effect.
- Boundary and finite coverage: Compare four feature pairs and a missing field.
- Reference-dependent errors: Read the confusion diagram for clean A.
- Transfer and human control: Redesign for fictional book recommendations.
Worked checks and assessment guidance
- People choose the intended benefit, permitted inputs, labels, candidate family, score, tie order and action. A model can learn a pattern within that setup. The same six features select promo for target A and sender for target B, demonstrating the role of target definition.
- For A, promo matches six, sender two and keep three; choose promo. For B, the scores are two, six and three; choose sender. These are training matches on an invented six-row table, not real spam or safety accuracy.
- The recorded targets become zero, zero, zero, one, zero, zero. Candidate scores are four, four and five, so keep wins. Against original clean A it matches only three of six and misses the three REVIEW targets. Preserve both versions and investigate why labels changed.
- A outputs KEEP, KEEP, REVIEW, REVIEW for zero/zero, zero/one, one/zero, one/one. B outputs KEEP, REVIEW, KEEP, REVIEW; keep outputs KEEP throughout. Missing features produce ASK. All feature pairs were already represented in training; this is finite functional coverage rather than new real-message evidence.
- Positive means REVIEW under A. The clean promo rule has three true positives and three true negatives on the six fitting rows. Always-KEEP has three true negatives and three false negatives. A result must identify the target and dataset; changing A to B changes the meaning of correctness.
- Define a new audience benefit, features, target, baseline and evaluation. Clicks can be a proxy, not proof of learning. Tiny features and human-defined references limit conclusions; review data/labels and plan suitable new evaluation. Keep uncertain wording untested until actually tried and give a reviewer evidence and authority to reject.
Learning evidence
Retain objectives A/B, six fictional rows, the two deliberately changed labels, all three score sets, chosen candidates, four-pair check, missing-feature case and baseline. Identify the reference for every score, two limitations and responses, a genuinely untested case and a reviewer who can intervene. Training or finite functional coverage is not real message-filter accuracy or practical certification.
OER and Wikipedia
DfE Computing supplies the KS2 data-evaluation anchor; this annual AI sequence is editorial enrichment. Google: framing an ML problem and Google: understand the problem were consulted for task definitions, appropriate metrics and simpler comparisons. Our binary records, label experiment, program and message examples are original. Reviewed 15 November 2026.
ConfusionMatrix.png, Jackverr, CC BY-SA 3.0. Unchanged diagram; it supplies class-comparison structure, not the original six-row data.
https://www.youtube.com/watch?v=OeU5m6vRyCk — CodeAI: AI: What is Machine Learning? Original recording embedded unchanged.
https://www.youtube.com/watch?v=x2mRoFNm22g — CodeAI: AI: Training Data & Bias. Original recording embedded unchanged. Teacher-supported context; no live pupil AI service or real inbox required.
Linked learning areas
aiMOOC Projects
Schulfach+


aiMOOCs


aiMOOC Projekte


MOOCwiki · Deutsch
Nach dem Lernen ist vor dem Lernen
Entdecke direkt den nächsten Lernkurs. Weitere Inhalte erscheinen, wenn Du weiter nach unten scrollst.
Zur MOOCwiki-HauptseiteMediathek
Mediathek wird aus dem Wiki geladen ...
Keine passenden Inhalte gefunden. Bitte ändere Suche oder Filter.
NEWSLernweltNOAH fragen