AI literacy and task selection for public servants

Learn in your browser · Chapter 01 of 08

The useful question is not whether a public servant is enthusiastic about AI. It is whether they can explain what a proposed system does, recognise a boundary and protect the citizen when an answer is uncertain.

Larchmere Council and all people, documents and numbers here are fictional. Its housing-enquiries induction helps staff explain services. Eligibility, vulnerability ranking, complaint outcomes and adverse actions remain with the authorised human process. The narrow training boundary prevents practice from becoming an experiment on residents.

The enquiries team wants residents to find correct hours and a usable contact route without a misleading case promise. Manager Mara compares a draft assistant with a maintained template. Choosing the template may be a valid improvement. Mickai's Trust Agent mission centres trustworthy, human-controlled AI; practical excellence includes declining an unsuitable use.

AI includes different computational techniques. A classifier assigns categories; a generative model produces content from learned patterns and context; retrieval finds material. An application can connect these components to actions. Ask what input, transformation, output and external effect are enabled rather than assuming the word assistant describes them.

Finding the current hours sheet is retrieval; shortening it is generation; sending it is an action. The wrong version can be found, an exception omitted or a message misdirected. Prompting addresses only part of this chain. Source selection, permissions and release controls need separate checks.

Official presentation can make fluent text appear authoritative. An assistant may promise an appointment without a booking check. That changes what a resident expects, even if spelling is perfect. Look for implied promises, omitted qualifications and wording that turns a process explanation into a personal decision.

Artificial general intelligence, or AGI, has contested definitions and predictions. This course claims no AGI capability and bases no service choice on its arrival. Evaluate demonstrated functions. Greater capability would still require a lawful purpose, controlled access, public authority and a way to stop.

Give staff permission to ask questions and report errors without blame. Existing skills in spotting ambiguity and assisting people belong in the control design. Learning should make that judgement visible, not become a prompt-writing contest or imply that automation alone solves workload.

Learn in your browser · Chapter 02 of 08

Choose the smallest useful task and describe where its output goes. Low-sensitivity input does not make a consequential output harmless, and a drafting label does not remove its influence on a decision.

List tasks before products. Record purpose, input category, output recipient, error consequence, ordinary-tool alternative and accountable owner. Use task descriptions rather than citizen records. Frontline staff and governance colleagues can then challenge the service map without unnecessarily collecting personal information.

Information sensitivity and output consequence are separate. Public rules can support a consequential eligibility recommendation; a spelling task can expose a private file. Consider combinations of details as well as individual fields. Unknown information or effect requires investigation, not an assumed low score.

For induction, choose an ordinary tool, a bounded training candidate or a stop/referral. These are teaching outcomes, not legal classifications. Apply exclusions before benefit scores: expected time savings cannot cancel a citizen-rights boundary. An eligibility recommendation stays excluded even in a vendor's invented demonstration.

The twelve proposals below produce three training candidates: public-note rewriting, fictional acknowledgement practice and invented feedback grouping. Candidate status covers this exercise only. Changes to data, recipients, integrations or purpose require reassessment before live use is considered.

Scroll sideways to see every column.

Grouping invented feedback differs from ranking real residents. A queue can affect waiting time, attention and access, so intermediate recommendations matter too. The Canadian federal scope guide includes some decision-support uses even with a human final decision; actual local applicability still needs checking.

Record alternatives and rejected proposals. A checked formula suits T04's exact arithmetic. For T01, compare the draft with the maintained template: does improved clarity justify its checking effort? The register should explain that choice rather than reward AI use as an objective.

When someone proposes changing T02 into personalised replies, open a new task record. Do not inherit the synthetic exercise's approval. Name the new input class, recipient, effect and review requirements. This simple change-control habit prevents a successful training session from silently becoming a customer-facing service.

Learn in your browser · Chapter 03 of 08

An approved task does not automatically approve its data or its supplier. Before entering anything, establish what information is needed, where it can travel and who can change the arrangement.

Replacing a resident's name with Resident A does not make a real letter synthetic. Address fragments, unusual circumstances and dated interactions can still identify someone. Stop the upload, keep the original in its authorised case system and create a fresh invented scenario with only the details needed for practice.

The induction permits current public sources or wholly synthetic material only. No real case records, staff files, credentials or private correspondence enter the tool. This is a conservative teaching boundary, not a worldwide legal prohibition. Real processing requires separate purpose, necessity, lawful basis, security and organisational review.

The ICO's data-minimisation guidance asks whether information is relevant and necessary for its purpose. For acknowledgement practice, a fictional category and invented reference suffice. Income, medical details and an address teach nothing additional. Spare fields are not a reason to collect them; a model cannot grant disclosure permission.

Map the data journey: hosted processing, retained logs, support access, subprocessors and connectors. A browser extension can introduce another recipient. Information-governance and commercial leads must check the actual contract and settings. Enterprise-ready is a sales description, not proof that this workspace has the approved arrangement.

Evidence of sovereign control covers data access, key custody, processing and support locations, model changes, decision rights and exit. A local server may still involve remote telemetry, backups or maintenance. Learners should understand the boundary; qualified technical colleagues must verify how it works.

For each supplier question, Mara records an evidence owner, required document or test, and unresolved issue. Discovery moves no real information. An unanswered retention or exit question remains a hold; a longer user disclaimer cannot create the missing control or right.

Scroll sideways to see every column.

An instruction inside a comment, such as export the register, is content rather than authority. This exercise has no network connector or live register. Real containment also requires permissions, separation and testing, not prompt wording alone. Stop scope changes and use the established security route when appropriate.

Learn in your browser · Chapter 04 of 08

A useful public-service draft preserves the source meaning and makes uncertainty visible. The reviewer needs to see which fact supports each claim, not just a polished paragraph.

Larchmere supplies two fictional source cards. Card LC-H1, version 1 dated 1 September 2026, says reception opens Monday to Thursday from 09:00 to 16:00 and Friday from 09:00 to 13:00. Card LC-C1, version 1 on the same date, says an enquiry can request contact from the team, appointment availability must be checked separately, and reception staff do not decide eligibility. These are invented facts for this workbook, not information about a real council. The exercise never uses them to advise a real resident.

The first draft says: Visit any weekday before 16:00 and we will arrange your appointment immediately. It changes Friday hours and converts a contact request into a promise. Record both discrepancies and hold the draft. Better punctuation or asking the model to confirm itself does not repair meaning.

Record each source's identifier, version, date, owner, permitted purpose and relevant passage. Open and check citations. If approved sources conflict, hold the affected claim and ask the owner; do not select the convenient answer. A search result is not automatically current operational policy.

A strong prompt defines role, constraints, method, verification and output structure. Ask for an internal draft from the two source cards, preserve exceptions and separate unsupported questions. Exclude personal advice, eligibility and sending. Clear instructions make checking easier but cannot replace comparison and release controls.

Distinguish clearer presentation from changed substance. Short sentences and headings may help; replacing may with will or adding a deadline changes meaning. Friday closes at 13:00 in the fictional source. Extending those hours would be a service-policy proposal, not an editorial correction.

Role: internal drafting assistant for a fictional staff exercise.
Inputs: approved source cards LC-H1 v1 and LC-C1 v1 only.
Task: propose a short reception information note.
Constraints: preserve hours and qualifications; no personal records,
eligibility decisions, appointment promises or external sending.
Method: list facts, draft, compare every claim with its source.
Return: draft | source support | unknowns | reviewer questions.
If sources conflict or a requested claim is unsupported, stop that claim.

Scroll sideways to see every column.

Check omissions as well as additions. Individually true sentences can mislead if the Friday exception disappears. Ask a colleague what they would do using only the draft, and compare that interpretation with the authorised service. This practice requires no real resident history.

Keep draft and source versions together. A new source needs a fresh dependency check. A hash can compare bytes with a reference, but cannot establish truth or authority. The source owner and governance record supply separate evidence. The later checker inspects declarations, not the truth of prose.

Translation can change meaning without an obvious new fact. Identify the need and route official notices through appropriate language and service review. Do not claim an unqualified staff check provides language assurance. Retain reviewed templates and accessible human explanations where available.

Learn in your browser · Chapter 05 of 08

A human name beside an output is not enough. The person needs the exact proposal, relevant evidence, time to consider it and authority to refuse or escalate before the service acts.

The induction tool has no booking, eligibility or messaging access. A learner proposes, a reviewer examines, and the real authorised process controls any later communication. Training success does not delegate decision rights. A login does not prove deliberate approval of a particular message.

Assign review by competence: service information, data permission or specialist language knowledge. Name who resolves each uncertainty and how it reaches them. Seniority alone does not confer every competence, and assigning a reviewer title cannot supply missing expertise.

Bind review to the draft version, source pack, audience and intended use. Record concerns, required changes and limits. Changed drafts or sources need another check before release. Opening a screen is evidence of interaction at most, not understanding, current intent or authority.

Twenty drafts arriving five minutes before closing may make review impossible. Reduce the batch, defer the new process or use the maintained template while preserving service access. A diagram without realistic workload provision can conceal rubber-stamping rather than prevent it.

A usable stop records the precise missing fact or authority, assigns its owner and provides a fallback. Hold new promises, missing sources and case decisions. Keep the ordinary human service available: residents should not wait for model repair when an established route can help.

Explaining the enquiries process is different from generating reasons for a refusal. Consequential decisions and their reasoning belong to the authorised procedure and applicable duties. The induction does not teach automated housing determinations or replace professional or statutory judgement.

Scroll sideways to see every column.

An attestation record may describe a declaration, but its value depends on identity, authority, evidence and procedure. It does not alone prove presence, intent or truth. The exercise's reviewerAssigned boolean is merely supplied data; the program neither authenticates a person nor provides professional attestation.

Role-play a changed Friday time while the source owner is unavailable. The learner should state a specific hold, use the fallback and identify the evidence needed to reconsider. Guessing an answer to finish quickly is not success. Observe the explanation and decision, not just the completed form.

Learn in your browser · Chapter 06 of 08

A useful service must work for people who cannot or do not wish to rely on the AI interface. Clarity, accessibility and a human route are part of the service design, not extras added after a pilot.

In a fictional role-play, the accurate note appears only as an image and directs all further help through a chatbot. Staff acting as residents with reading or internet barriers cannot use it easily. Restore structured text and a human route before judging the draft's usefulness.

Plain language preserves qualifications, choices and help routes. Explain enquiry, appointment and decision rather than assuming residents know internal structures. Shorter sentences should make the next step clearer without implying that an application succeeded or a complaint was settled.

Check document structure, headings, contrast, keyboard operation, labels, reading order and visual alternatives. Combine this with appropriate user and assistive-technology testing. Record methods and limits. One automated score does not establish that everyone can use the complete service.

Assisted digital support helps people who cannot use a digital route independently. The GOV.UK Service Manual treats it as user support and service design. Retain a staffed route; verify actual contacts, availability and language arrangements with its owner. Never invent details to complete a draft.

Criticism of AI is not automatically hostile input. A complaint may concern accuracy, exclusion, privacy or treatment. Preserve an appropriate record and use the authorised human complaints process. The assistant cannot judge validity, rank the complainant or decide an adverse outcome.

Withdraw a wrong note and use the maintained version while its owner checks the correction. This restores service but does not settle the complaint or assess harm. The proper investigation and further review route depend on the service and jurisdiction; do not invent universal deadlines.

Scroll sideways to see every column.

Do not infer disability, ethnicity or vulnerability from names, accents or writing style for training metrics. Use inclusive synthetic cases and identify evidence gaps. Real equality evaluation requires suitable expertise, lawful information handling and participation methods, not a reassuring overall satisfaction score.

Explain relevant AI use and checking without transferring responsibility to residents. AI may be wrong is not a substitute for staff accountability or accessible help. Keep public explanations useful while protecting personal and security-sensitive information.

Learn in your browser · Chapter 07 of 08

A useful governance record says which rule applies, to whom and why. A policy title, a training certificate or a selected region cannot establish that an actual service is compliant.

Begin with the real public body, service function, people affected, information flows and proposed decision role. Then select a jurisdiction pathway and identify the qualified owner who will check applicability. Larchmere is an invented teaching setting, not a device for importing one country's rules into another. Its task-selection rules are intentionally conservative training boundaries. They do not state that every excluded induction task is prohibited by law in every jurisdiction.

Separate four kinds of material: legislation, institutional policy, professional or regulator guidance, and an educational method. Their authority and scope differ. The UK Government AI Playbook supports practical AI understanding and meaningful human control. It is guidance, not a certificate or a replacement for current law. The Algorithmic Transparency Recording Standard has a defined mandatory scope and exemptions policy. Do not state that all councils or every administrative tool automatically fall within that mandatory scope.

Australian Government AI policy v2.0 took effect on 15 December 2025 and applies to non-corporate Commonwealth entities with exceptions. It addresses accountability, staff training and use-case governance. A local council or state service must check its own arrangements rather than treating that federal policy as automatically binding. In Canada, the federal scope guide distinguishes covered administrative decision-support uses from other work. Some partial automation remains in scope even where an officer makes the final decision; institution, function and production conditions still matter.

For an EU setting, use current AI Act materials, the Commission's AI literacy guidance and the competent national authorities. The Commission page emphasises tailoring literacy to the people, systems and context and does not present copying a training practice as automatic compliance. This workbook does not resolve every changing commencement, amendment or sector obligation. A course attendance record and a knowledge score are useful learning evidence, not proof that the organisation has fulfilled its legal duties.

The NIST AI Risk Management Framework is intended for voluntary use. NIST states that version 1.0 is being revised. Its risk-management vocabulary can structure questions, but it is not law or accreditation and this course offers no NIST certification. The ICO AI guidance also carries a notice that changes following the Data (Use and Access) Act require review. Avoid repeating an old blanket automated-decision rule as if the legal position never changes. Keep the course's no-eligibility-automation boundary as a training rule while qualified leads determine actual obligations.

Scroll sideways to see every column.

For T01, record public-information training only, no personal data or live sending, and no deployment approval. A real case-summary proposal needs a separate decision-influence and information review. Keep sensitive advice and case information in authorised internal systems, not the public workbook.

Update the record when the purpose, tool connection, affected population or source changes. Do not turn legal uncertainty into an instruction to do nothing indefinitely: name the question, evidence owner and fallback. Equally, a deadline or an enthusiastic sponsor does not answer the question. The team can continue synthetic learning while the actual live proposal remains held.

Learn in your browser · Chapter 08 of 08

Turn knowledge into observable behaviour. A useful induction shows that staff can detect an error, refuse an unsuitable task, explain the reason and continue the service through an appropriate route.

Larchmere's four-week induction permits T01 public-note rewriting, T02 fictional acknowledgement practice and T03 invented feedback grouping. Mandatory stops are an eligibility recommendation, a real case-file input and an unsupported appointment promise. No activity sends to residents, changes queues or determines outcomes. Training documentation is not deployment approval.

Week one establishes vocabulary and the twelve-task register. Week two develops source checking and minimal-input practice. Week three rehearses human review, accessibility and complaints handling. Week four repeats challenging cases with changed details, measures the complete workflow and makes a reasoned learning decision. Learners who need more support receive another practice opportunity. Attendance and confidence surveys are useful context but are not substitutes for observing whether someone spots the actual problem.

Use the optional downloadable task-selection checker to inspect a rule rather than trust a model. It uses Python's standard library, fixed fictional task definitions and synthetic JSON. Run python -m unittest -v to exercise the tests, and python task_selector.py fixtures.json to view the twelve dispositions. It makes no network request and does not call an AI system. Read its source and README first, use an approved local environment, and never substitute a real citizen file. Staff without Python can trace the identical fields on the supplied paper worksheet.

The checker validates exact fields, then applies exclusions and holds before selecting an ordinary tool or training candidate. Unknown fields, types and task IDs are rejected. Its declarations are not authenticated: a favourable result establishes neither truthful input, actual review nor legal permission. This limitation is central to the lesson.

Scroll sideways to see every column.

For a fictional measurement example, the current template process takes 8 minutes per note. Across ten comparable synthetic notes that is 80 minutes. A draft assistant takes 2 minutes to generate each note, 5 minutes to review it and 20 further minutes across the batch to repair mistakes. Its total is 10 multiplied by 7, plus 20, or 90 minutes. Generation alone appears fast; the complete observed process is slower in this small invented comparison. The numbers teach a method and make no claim about a real product or public service.

Review quality and access alongside time. Count unsupported promises, missed exceptions, successful stops, and whether the designed support route remained usable. Do not sum downloads, registrations and page visits as if they were unique competent staff. Avoid collecting sensitive personal characteristics for convenience. A learning lead can keep proportionate training records under the organisation's actual rules and identify where a qualified assessment is required. Small synthetic results do not establish performance with real residents.

Agree the next decision before running the session. GO may mean another bounded training session, not live deployment. HOLD means resolve a named evidence or capability gap while retaining the ordinary service. STOP means end an unsuitable proposal, preserve relevant authorised evidence and explain the reason. Reconsider when source versions, connections, permissions or the service purpose change. Reward a well-justified stop as evidence of competence rather than a failure to adopt technology.

The accompanying assessment is knowledge-only: a twenty-question sitting drawn from a forty-question bank. Passing requires strictly above 80 percent, so 16 of 20 is insufficient and 17 is sufficient. The written exercises and capstone are ungraded practice. A named Mickai certificate, when issued through the checked official exam route, records the assessment result and issuance; it is not accreditation, proof of identity, professional authorisation or a legal compliance finding.