Outlook
One AI-assisted practice task for social media training, with a manual baseline
Design one bounded AI-assisted practice task for England social media training with a manual baseline, human review, stop rules and rollback.
Start with one bounded formative task: generate alternative draft captions from a fictional brief, then require the learner to identify factual, claims, accessibility and brand risks. The human learner submits the final reasoning. The tool neither publishes nor grades.
This is an editorial scenario for an England employer, not evidence that any provider offers or has tested the application.
What to take away
- Use one bounded task where learners generate captions and identify risks, with humans submitting final reasoning.
- Keep retrieval, classification, generation and action separate; permission for one function does not permit the next.
- Test with synthetic cases before learner access, covering claims, accessibility, data and adversarial instructions.
- Record model, settings, prompt, output, reviewer, defects and final human decision for every case.
- Pause the task if prohibited data or unsafe claims pass, or if the service changes without approval.
Define the manual baseline
Without AI, the learner writes three alternatives, checks each against the brief and records revisions. The AI-assisted variant uses the same brief, time window, rubric and evidence fields. Compare only whether the exercise produces reviewable evidence, not campaign performance.
Record the model or service, edition, date, settings, prompt, output, reviewer, defects and final human decision. Prohibit production accounts, customer data, confidential campaigns and real employee performance records.
The AI Playbook for the UK Government discusses meaningful human control and life-cycle management for government organisations. Its scope does not clear a private training use, but it supports explicit decision points and ownership.
Keep four functions separate
Retrieval finds approved course material. Classification applies a stated rubric. Generation proposes text. Action publishes, scores or changes access. Permission for one function does not permit the next. In this exercise, only generation is enabled; retrieval is limited to approved fictional material, and action remains human.
Four Functions and Their Permissions
Retrieval
- What it does
- Finds approved material
- Enabled here
- Limited
- Who decides
- Approved sources
- Permission carries
- No
Classification
- What it does
- Applies stated rubric
- Enabled here
- No
- Who decides
- Rubric owner
- Permission carries
- No
Generation
- What it does
- Proposes text
- Enabled here
- Yes
- Who decides
- Learner
- Permission carries
- No
Action
- What it does
- Publishes or scores
- Enabled here
- No
- Who decides
- Human
- Permission carries
- No
Test before learner access
Build a synthetic set covering ordinary captions, regulated-product claims, image descriptions, ambiguous evidence, prohibited personal data and adversarial instructions. A learning reviewer checks alignment with the rubric. A claims reviewer examines whether the task teaches substantiation rather than fluent assertion. CAP's substantiation advice says objective claims need objective evidence; a generated sentence supplies neither.
Test keyboard operation, focus, screen-reader output, captions for any media and a valid no-AI route. GOV.UK's reasonable-adjustments guidance includes trainees and apprentices, while actual duties and configurations need specialist review.
Create an acceptance ledger before running the set. For each input, record expected learner decision, prohibited output, actual output, reviewer judgement, correction. Report cases in which the learner identified every pre-defined critical defect / all valid cases presented, with the set version, period, exclusions and reviewer.
This internal measure tests the exercise and learner response under one configuration, and it is not a benchmark, product-accuracy rate or forecast of workplace performance.
Compare with the manual baseline using the same cases. If reviewers change the rubric after seeing outputs, mark a new version and rerun both routes. Do not retain only successful examples.
Protect data, systems and rights
The ICO's AI guidance is under review after the Data (Use and Access) Act 2025. A named privacy and employment reviewer must address purpose, roles, minimisation, transparency, retention, worker monitoring and human decisions.
The NCSC's secure AI guidance covers security across design, development, deployment and operation. Test supplier changes, access, logging, leakage, incident response and restoration. Copyright review must cover course content, learner inputs and output use; no model licence is assumed.
The buyer also records who may change prompts, models and connected tools. An unapproved version change returns the exercise to HOLD even when the interface looks unchanged.
Stop and roll back
Pause the task if prohibited data enters a prompt, an unsafe claim passes review, the assessment rewards tool output rather than learner judgement, an access route fails or the service changes without approval. Disable the integration, preserve evidence, notify the named owners and return to the manual exercise. Restart only after the defect is reproduced, corrected and independently retested.
Before you act
- Define a manual baseline with the same brief, time window, rubric and evidence fields.
- Prohibit production accounts, customer data, confidential campaigns and real employee performance records.
- Build a synthetic test set covering ordinary captions, regulated claims, ambiguous evidence and prohibited data.
- Create an acceptance ledger recording expected decision, prohibited output, actual output and reviewer judgement.
- Name privacy, employment and security reviewers to address data protection and system changes.
- Pause and roll back if prohibited data enters, unsafe claims pass, or access routes fail.
Common questions
What is the purpose of the manual baseline in this exercise?
The manual baseline requires the learner to write three alternatives without AI, check each against the brief and record revisions. The AI-assisted variant uses the same brief, time window, rubric and evidence fields. Compare only whether the exercise produces reviewable evidence, not campaign performance.
How should the four functions be separated in this task?
Retrieval finds approved course material. Classification applies a stated rubric. Generation proposes text. Action publishes, scores or changes access. Permission for one function does not permit the next. In this exercise, only generation is enabled; retrieval is limited to approved fictional material, and action remains human.
When should the task be paused and rolled back?
Pause if prohibited data enters a prompt, an unsafe claim passes review, the assessment rewards tool output rather than learner judgement, an access route fails or the service changes without approval. Disable the integration, preserve evidence, notify named owners and return to the manual exercise. Restart only after the defect is reproduced, corrected and independently retested.
