Zum Inhalt springen

English:Explainable predictions and user-centred interfaces - AI Year 8

Aus MOOCsWiki Staging
aiMOOC-Siegel aiMOOC

Explainable predictions and user-centred interfaces - AI Year 8

QR-Code



Introduction

Hands preparing a paper interface prototype on a wooden table
Sage Ross photographed a paper prototype for a Wiki Education interface. Our fictional museum cards below are a separate original exercise.

A useful prediction display helps people understand, question and act. Compare two paper designs for a fictional museum activity guide, check explanations against the actual rule and communicate missing evidence honestly. Build an inclusive review plan without collecting real visitor information. The supplied program produces text records, not a live museum service or a tested accessible website.


Explain what the system actually did

Our original museum-guide prototype suggests either Workshop or Gallery from two stated preferences. One input asks whether the fictional visitor would like a hands-on activity; the other asks whether they prefer a quieter setting. Each answer is zero for no, one for yes, or unanswered. These are activity preferences, not ability, age, health or a judgement about a person's interests. No real visitor fills in this course fixture.

The supplied fixed rule gives three points for hands-on=yes and one for quieter=yes. A total of at least three suggests Workshop; a smaller total suggests Gallery. The weights are authored examples, not a model fitted to visitor data. The score is not a probability, confidence measurement or quality rating. Calling a display AI-powered would not change that fact.

Hands-on preference Quieter-setting preference Rule score Suggestion
No No 0 Gallery
No Yes 1 Gallery
Yes No 3 Workshop
Yes Yes 4 Workshop

Notice a design weakness: changing the quieter-setting preference changes the score but never the suggested activity on these four valid combinations. A visitor selecting both preferences still receives Workshop. That does not prove that a workshop will be quiet. The example deliberately makes this limitation visible so that a team can question the rule instead of hiding it behind a polished explanation.

For hands-on=yes and quieter=no, an accurate explanation is: the hands-on answer contributes three points, the quieter answer zero, and the total meets the threshold of three. Saying the system observed you enjoying craft would be invented evidence. Saying a quieter preference caused this suggestion would misrepresent the calculation. A faithful explanation of a poor rule is still possible; explanation quality and recommendation suitability are different questions.

NIST's explainability report distinguishes giving reasons, making them meaningful to intended users, matching the actual process and respecting the system's limits. Apply these ideas to our card: show the real calculation, use understandable wording, and state what the rule cannot determine. A clear story that invents how a result was reached is not a faithful explanation. Different audiences may need different levels of detail.

CodeAI introduces machine learning. Compare fitting from examples with our supplied scoring rule. No trained classifier or calibrated confidence was created here. A real learned model may need a different explanation method; the three-point arithmetic is not a universal explanation of neural networks.

Attempt before checking: change only the quieter-setting answer from zero to one for the same hands-on visitor. Hint: the total rises from three to four. After attempting, check that the label remains Workshop and explain why neither result verifies actual noise, availability or accessibility. Then change hands-on to zero and quieter to zero: the result becomes Gallery. This is a change in the authored rule's output, not proof that a person's real preferences or opportunities have changed.


Compare two cards and communicate uncertainty

Prototype A is intentionally poor: it shows Workshop, a bright badge reading 97% certain and one button labelled Go. It gives no basis for the percentage, no visible route to change inputs and no explanation of what the suggestion excludes. These are authored design defects, not claims about an existing museum or MOOCwiki feature. An appealing badge can encourage unwarranted reliance when its meaning is unsupported.

Prototype B uses a short layered explanation. First comes the suggestion and the fact that it comes from a fixed rule. Next come the actual input contributions and threshold. The limitation remains visible: current availability, step-free access and actual noise are unknown, and the score is not a probability. Then come clearly named options to change choices, choose the other activity or ask museum staff. Extra technical details such as version and full calculation can be placed in a secondary layer without hiding the essential limitation.

Card section Example content for hands-on yes and quieter no
Suggestion Workshop — fixed-rule suggestion
Why this appeared Hands-on contributes three and quieter contributes zero; total three meets threshold three
What remains unknown Current availability step-free access and actual noise; score is not a probability
Your next step Change choices; choose the other activity; ask museum staff

The distinction between uncertainty and a missing fact matters. We have no current availability record, so say that availability is unknown rather than inventing an 80 percent chance of a place. We have not measured how often this rule gives useful suggestions, so a calibrated probability cannot be inferred from the points. A visitor may also disagree with the objective itself. Explain the decision boundary and the limits without treating disagreement as a faulty user.

If either preference is unanswered, the program returns needs_input with no score or suggestion. It does not silently turn missing into no. The message says that a choice is unanswered and that no suggestion was calculated. In a paper design, offer the user an understandable way to answer or obtain staff help. Requiring a private explanation of why a choice is unanswered is unnecessary for this fixture.

The following complete program was actually executed. It generates structured text records for the paper cards, enumerates all four answered combinations and records a before-and-after change. It does not create buttons, send requests to staff, run an accessibility test or implement the hypothetical 97% badge. The output's action labels are proposed controls, not working service links.

"""Text records for a fictional paper prototype, not a deployed user interface."""
import json

VERSION = "museum-rule-v1"
LIMIT = ("This fixed rule does not know current availability, step-free access "
         "or actual noise. Its score is not a probability.")
ACTIONS = ["Change choices", "Choose the other activity", "Ask museum staff"]


def checked_choice(value):
    if value is None:
        return
    if type(value) is not int or value not in (0, 1):
        raise ValueError("Use 0, 1 or None for an unanswered choice")


def build_card(hands_on, quieter):
    checked_choice(hands_on)
    checked_choice(quieter)
    if hands_on is None or quieter is None:
        return dict(version=VERSION, status="needs_input", suggestion=None,
                    score=None, components=None,
                    reason="A choice is unanswered. No suggestion was calculated.",
                    limitation=LIMIT, actions=list(ACTIONS))
    hand_points = 3 * hands_on
    quiet_points = quieter
    score = hand_points + quiet_points
    suggestion = "Workshop" if score >= 3 else "Gallery"
    comparison = "at least" if score >= 3 else "below"
    reason = (f"Hands-on choice contributes {hand_points}; quieter-setting choice "
              f"contributes {quiet_points}. Total {score} is {comparison} "
              "the threshold 3.")
    return dict(version=VERSION, status="suggestion", suggestion=suggestion,
                score=score, components=dict(hands_on=hand_points,
                                             quieter=quiet_points),
                reason=reason, limitation=LIMIT, actions=list(ACTIONS))


def demonstrate():
    cards = []
    for hands_on in (0, 1):
        for quieter in (0, 1):
            cards.append(dict(inputs=dict(hands_on=hands_on, quieter=quieter),
                              card=build_card(hands_on, quieter)))
    return dict(cards=cards, missing=build_card(None, 1),
                before_change=build_card(1, 0), after_change=build_card(0, 0),
                classifier_trained=False, probability_calibrated=False,
                user_study_conducted=False, live_interface_implemented=False)


if __name__ == "__main__":
    print(json.dumps(demonstrate(), indent=2))

Model the trust boundaries: preference input → type and missing-value checks → fixed scoring rule → faithful explanation record → proposed display → user choice or human review. A designer or generative model must not insert an invented reason between the calculation and the displayed card. Keep the rule version and inputs available for checking, with no unnecessary identifying information.

Try a failure case: enter unanswered for hands-on and yes for quieter. First hint: missing is not zero. Partial step: the missing-value branch runs before arithmetic. Check after attempting that the record contains no score and no suggestion. Explain what message and human-help route the paper interface should show. For a changed rule version, recheck both the label and explanation instead of keeping an old explanation merely because it sounds plausible.


Evaluate usability without inventing a user study

A paper walkthrough is useful for finding questions and obvious gaps. It is not a completed user study, a screen-reader test or proof that everyone understands a message. Our six fictional cases below are authored review scenarios. No participant responses, comprehension percentages or actual accessibility results have been collected. Record a proposed test separately from evidence that it has been performed.

Fictional situation Question for the reviewer Proposed evidence
A visitor asks why Workshop appeared Can the visitor connect the input contributions with the result Ask for an explanation in the visitor's own words during an approved future test
Both preferences are yes Could the card be mistaken for a quietness guarantee Check whether the unknown-noise limit is understood
A screen-reader user changes a choice Would a changed status be announced without stealing focus Test a real implementation with appropriate assistive technology
A keyboard user wants staff help Can the focused control be found and operated Test keyboard order visible focus and meaningful control names
A preference is unanswered Does the interface explain the missing state without guessing Inspect the needs_input message and alternatives
A visitor changes their mind Do the result and explanation update consistently Compare the old and new rule records and then test the implemented interaction

W3C guidance supports giving input labels and useful instructions, showing which keyboard control has focus, and making qualifying status changes available to assistive technology without requiring focus to move. In the proposed card these lead to specific tests: meaningful labels for both preferences, visible keyboard focus for actions, and a way to announce a new result. A table of intentions does not implement those behaviours. Passing one criterion or producing readable text would not establish full WCAG conformance.

For each case use four separate review questions: does the person understand the purpose, identify the actual reason, recognise a limit and find an appropriate next step? Do not compress these into one fabricated success rate. If an authorised future test occurs, obtain suitable permission, collect only necessary evidence and report the task conditions and limits. This course's invented personas must not become a reason to collect real pupils' private profiles.

CodeAI discusses training data and bias. Relate the wider question of whose needs are represented to our fixture, while remembering that this rule was authored rather than fitted. A larger explanation box cannot fix an unsuitable objective, excluded users or missing service information. Review both the method and how the result is presented.

Write a risk record with responsibility and evidence. A false confidence badge risks overreliance: the designer removes unsupported percentages and checks each claim against the rule. An inaccessible changed-result message risks missed information: the implementation lead tests real keyboard and assistive-technology behaviour. A misleading quieter preference risks unmet needs: the museum-content lead reviews actual service information before any pilot. An ineffective review route risks leaving someone stuck: the service lead checks staff authority, response arrangements and an alternative route. Each entry needs a residual concern and a monitoring action; these controls are proposals, not completed operational safeguards.

Transfer studio: redesign the card for a fictional school-club information guide. Compare a minimal explanation with a layered one using clarity, fidelity, cognitive load, privacy, accessibility and the ability to choose otherwise. Identify what the rule knows, what it does not know and who can review an error. Do not use the card to grade pupils or assign them automatically. Recommend revise at paper stage until the objective and proposed interface have suitable evidence. This is an editorial Year 8 application of computational abstraction, not a claim of statutory AI coverage or a deployed guidance service.


Interactive Tasks


Quiz: Test Your Knowledge

What kind of model produces the museum suggestion? (An authored fixed scoring rule) (!A classifier trained on real visitors) (!A calibrated probability model) (!A sensor measuring personal interests)



What happens for hands-on yes and quieter no? (Score three suggests Workshop) (!Score one suggests Gallery) (!A measured probability of ninety seven percent) (!No result can ever be calculated)



What does changing only quieter from no to yes do here? (It changes the score but not the suggested activity) (!It guarantees a quiet workshop) (!It proves a causal change in visitor behaviour) (!It automatically trains new weights)



Which explanation faithfully describes the source of the result? (The stated answers contributed points to a fixed threshold rule) (!The system observed the visitor enjoying craft) (!An invented confidence badge proved the choice) (!The algorithm knows every activity is available)



Why should prototype A lose its confidence badge? (There is no evidence supporting the displayed percentage) (!Every percentage is forbidden in every interface) (!The number is smaller than one hundred) (!A coloured badge is always accessible)



What does the program do with an unanswered preference? (Return needs_input with no score and no suggestion) (!Silently convert it to no) (!Invent a private reason for the visitor) (!Assign the most popular activity)



What does a label such as Ask museum staff represent in this code? (A proposed paper control rather than a working service link) (!A message already sent to a real staff member) (!A completed complaints investigation) (!A guarantee of immediate support)



What evidence do the six fictional walkthrough cases provide? (Authored review scenarios and questions for future tests) (!Measured comprehension from six participants) (!A full screen-reader audit) (!Proof that every user prefers prototype B)



Why should a changed result be considered in accessibility testing? (Users need to become aware of important status changes) (!Focus must always jump away from the active control) (!A visible percentage replaces accessible status information) (!Keyboard users never change preferences)



What remains necessary after producing a clear explanation? (Review the objective evidence limits and real user needs) (!Assume the underlying rule is suitable everywhere) (!Treat explanations as proof of fairness) (!Remove every way to choose otherwise)





Memory Game

Fidelity An explanation matches the actual process
Uncertainty A limit on what the evidence supports
Prototype An early representation of a proposed design
Persona A fictional scenario for considering a user need
Focus The current keyboard interaction location
Review An accountable opportunity to question or correct an outcome





Drag and Drop

Show how the actual points produced a suggestion Faithful explanation
State that activity availability is unknown Knowledge limit
Keep no score when a preference is unanswered Missing-input handling
Offer a clearly named way to change choices User control
Inspect changed-result announcements in a real implementation Accessibility test





Crossword Puzzle

REASON What explains how a result was reached
LIMIT A boundary on the supported claim
FOCUS The keyboard interaction location
LABEL Text identifying the purpose of a control
CHOICE An option the user can select
REVIEW An opportunity to challenge an outcome





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

Complete the explanation.
The museum uses a fixed

. Hands-on yes contributes

points. Quieter yes contributes

point. A score of at least three suggests

. A missing answer is not silently changed to

. A faithful explanation matches the actual

. Unknown availability is a

. A keyboard user needs visible

. The six personas are

. A paper prototype is not a completed user

.



Open Tasks


Easy

  1. Read the four combinations: Calculate every score and suggestion from the two binary preferences. Hint: hands-on contributes three and quieter one. Explain why score two is not reachable with these valid answers.
  2. Explain the example honestly: Write a short explanation for hands-on yes and quieter no. Hint: state the actual contributions threshold and a limit. Do not invent observation training or a confidence percentage.
  3. Identify missing evidence: List three things the rule does not know about a real visit. Hint: activity names do not verify availability noise or step-free access. Explain how a visitor could seek appropriate help.
  4. Handle an unanswered preference: Describe the needs_input record and an understandable paper message. Hint: unknown is different from no. Offer answering changing or staff help without asking for private reasons.


Standard

  1. Compare the two cards: Evaluate prototype A and B using reason fidelity uncertainty and user control. Hint: visual polish does not support a confidence claim. State a possible burden of too much explanation as well as missing explanation.
  2. Map the trust boundary: Draw inputs checks fixed rule explanation proposed display and human choice. Hint: a generated explanation must not invent the model process. Add the rule version and responsibility for checking claims.
  3. Walk through three personas: Choose three fictional cases and specify the task review question evidence and unknown. Hint: do not report a test as performed. Distinguish missing data misunderstanding and inaccessible interaction.
  4. Design accessibility checks: Write proposed tests for input labels keyboard focus and changed-result status. Hint: the text generator does not implement a web interface. Name the implementation and assistive-technology checks still required.


Advanced

  1. Expose a rule limitation: Explain why the quieter preference cannot change the output label in the current rule. Hint: compare both values for each hands-on answer. Recommend a review of objective and service evidence rather than a more confident explanation.
  2. Plan a responsible future study: Design a small approved comprehension study with four review questions and minimal data. Hint: fictional personas are not actual participants. State permission safeguarding accessibility and reporting limits without collecting data now.
  3. Create a risk register: Write four risks with controls owners evidence residual concerns and monitoring. Hint: include overreliance an inaccessible update unsuitable quietness inference and an ineffective staff-help route. Mark every unimplemented safeguard.
  4. Transfer to school-club guidance: Compare minimal and layered explanations for a fictional information guide. Hint: revisit the rule user need excluded uses and ability to choose otherwise. Do not use the exercise for automated pupil assignment or grading.



Learning Check

  1. Calculate and interpret the rule: Give all four valid outputs and the meaning of their scores.
  2. Audit explanation fidelity: Why is a plausible story insufficient?
  3. Handle limits and missing values: How should the display handle unavailable knowledge?
  4. Compare interface designs: What improves the second card and what remains unproved?
  5. Separate scenarios from evidence: What do the six fictional personas establish?
  6. Recommend the next stage: What makes a responsible paper-stage recommendation?


Worked checks and assessment guidance

  1. Inputs no no score zero Gallery; no yes score one Gallery; yes no score three Workshop; yes yes score four Workshop. Scores are authored point totals not fitted confidence or probabilities. The quieter answer changes the score but cannot change the label in this rule.
  2. An explanation should reflect the actual process. For hands-on yes and quieter no three plus zero meets threshold three. Claiming the system observed craft enjoyment or that quietness caused the label invents evidence. Even a faithful explanation does not establish that the rule meets actual needs.
  3. State that availability step-free access and actual noise are unknown rather than inventing probabilities. An unanswered preference produces needs_input with no score or suggestion. Explain what is missing and offer suitable choices or staff help without inferring a private reason.
  4. The layered card provides a reason an essential visible limitation and meaningful action options while avoiding an unsupported 97 percent badge. Extra detail should be available without hiding crucial information. The program outputs text records only and does not implement actions or prove comprehension accessibility or suitability.
  5. They structure review questions about purpose reason limits and next action. They are not six observed people and give no measured success rate. Real implementation tests must inspect labels focus and status messages including assistive technology. A planned check or one passed criterion does not establish full WCAG conformance.
  6. Review the objective the quietness limitation provenance versions error consequences and user control. Maintain risks controls owners evidence residual concerns and monitoring. Obtain appropriate permission and minimal evidence for any later study. A changed school-club context requires fresh review and must not become automatic pupil grading or assignment.


Learning evidence

Retain the four score records, one faithful explanation with explicit unknowns, a missing-input example, a before-and-after choice change, a comparison of two cards and a labelled trust-boundary diagram. Add six fictional walkthrough cases, proposed accessibility and comprehension tests, a four-entry risk register and a school-club transfer. Separate executed text-generation checks from an unimplemented interface, future user research and any claim of accessible deployment.


OER and Wikipedia

NISTIR 8312: Four Principles of Explainable Artificial Intelligence supplies the distinction between reasons, audience meaning, fidelity to the process and operating limits. Relevant abstract and executive summary inspected 15 November 2026. The original museum calculation and card comparison are not NIST experiments or endorsed products.

W3C: Labels or Instructions, W3C: Focus Visible and W3C: Status Messages inform the proposed input, keyboard and changed-result checks. These Understanding pages explain selected criteria; the course does not claim a full conformance audit or implementation. Relevant sections inspected 15 November 2026.

DfE: Computing programmes of study supplies the KS3 computational-abstraction anchor. Year 8 AI studio objectives are editorial applications, not additional statutory Year-specific requirements.

Paper prototype of website user interface, 2015-04-16.jpg, Sage Ross, CC BY-SA 4.0. Photograph embedded unchanged. It depicts a Wiki Education interface prototype at WINTR, not our museum cards or a user test conducted for this course.

https://www.youtube.com/watch?v=OeU5m6vRyCk — CodeAI: AI: What is Machine Learning? https://www.youtube.com/watch?v=x2mRoFNm22g — CodeAI: AI: Training Data & Bias. Actual metadata playability embedding and preview bytes verified; previews visually inspected. Original videos embedded unchanged without transcript copying endorsement redistribution or relicensing implied. All required work is possible from the text without playback.

The museum rule weights card wording poor confidence-badge example fictional personas risk register and school-club scenario are original teaching constructions. The complete text-record program was actually executed. No real visitor study learned model calibrated probability accessibility certification working staff-help service local generative LLM or paid API was used.


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-Hauptseite

Mediathek

Suche verfeinern

Dein Lernplan bestimmt, was oben steht. Alle anderen Inhalte bleiben erreichbar. Seitenaufrufe beeinflussen die Auswahl nicht.

Inhalte werden geladen ...

Mediathek wird aus dem Wiki geladen ...