AI & Data Career Foundations · Intermediate · ~7 hours
AI Safety, Security & Governance
Recognise AI privacy, reliability and security risks and design workflows with appropriate human oversight.
Status: Catalogue ready. This lesson version is available in review mode while subject-matter, accessibility and assessment checks are completed.
Govern AI, data and human oversight
Map AI data and risks
By the end of this lesson
- Map inputs, processing, outputs, users and storage in a fictional AI workflow.
- Distinguish privacy, security, reliability and fairness risks.
- Prioritise a small set of controls using impact, likelihood and reversibility.
Risk starts with the data flow
An AI risk review should begin with what information moves where, not with a list of fashionable controls. Map the source of inputs, the people who can submit them, the model or service that processes them, the output recipients, storage locations, retention and deletion paths. Include prompts, retrieved documents, logs, evaluation data, feedback and any connected tools. A workflow that appears to be a simple assistant may copy information into several systems. Classify the information by sensitivity and purpose. A public product description is different from a customer contract, a candidate record or an internal security note. Ask whether each field is necessary, whether the person authorised its use, who can access it and how long it should remain. Data minimisation is a control, not merely a privacy slogan. If a task can be completed with a fictional identifier or redacted excerpt, do not send the full record. Name different risk types. Privacy risk concerns unwanted use or disclosure. Security risk includes unauthorised access, manipulation or leakage. Reliability risk covers incorrect, unavailable or inconsistent output. Fairness risk concerns unequal error or treatment. Accountability risk appears when nobody owns a decision or correction route. A single workflow can carry all of these, so a control that addresses one does not automatically address the others.
Prioritise controls by possible harm
Risk prioritisation should help a team act. Describe the unwanted event, affected people, potential impact, likelihood evidence, existing controls, remaining uncertainty and owner. Use qualitative terms when precise probabilities are not supported. A low-likelihood privacy disclosure with severe consequences may deserve attention before a frequent minor formatting error. Consider scale and reversibility: an output reviewed by one person is different from an automated message sent to thousands. Controls can reduce the chance of an event, limit the data exposed, detect a problem, correct it or provide a safe alternative. Examples include approved-tool lists, field redaction, access roles, retention limits, source citations, test sets, review queues, audit logs and an immediate stop process. Link each control to a risk and define how it will be checked. A policy that is impossible to follow under normal workload is not an effective control. Record residual risk and the decision owner. The aim is not to claim that risk has disappeared. It is to make a proportionate decision that can be revisited when the model, data, users or purpose changes. Governance becomes stronger when affected people can understand the process and challenge a harmful result.
Evidence, judgement and a safe next step
The professional habit in this lesson is to make the decision process visible. Start by naming the AI data-flow risk in plain language, then identify the evidence that supports the current view and the evidence that is still missing. Evidence can be a buyer statement, a dated observation, a documented calculation, a permitted source or a clearly labelled hypothesis. It should be possible for another colleague to understand where a conclusion came from without relying on confidence or personal reputation. When the evidence is weak, reduce the strength of the claim rather than decorating it with persuasive language. The next step should be proportionate to what is known. Map where information goes, classify sensitivity and link each material risk to an owner and a checkable control. A small, reversible action is often more useful than a grand promise: ask one clarifying question, test one calculation, invite the right stakeholder, or record one assumption for review. Give the other person a genuine opportunity to agree, decline or change the sequence. This matters in commercial work because a fast process is not automatically a good process, and it matters in career development because a polished artefact is not automatically truthful evidence. Keep the boundary explicit. A risk register does not guarantee compliance, security, fairness, employment, placement or a model's accuracy. Do not put private information, speculative outcomes or unsupported performance claims into a learner portfolio, CRM, dashboard or application. If a decision affects another person, explain who owns it and how they can challenge or correct it. A good submission therefore shows both capability and restraint: it demonstrates what was done, states what was not tested, and identifies a sensible follow-up. Reviewers can reward that clarity even when the proposed answer is incomplete, because responsible professional judgement is part of the skill being developed.
Workplace example
A fictional assistant receives customer notes, generates a summary and stores a log. The risk map identifies unnecessary personal data, cross-team access and unsupported claims. The team redacts fields, limits roles, requires source checking and assigns an owner for corrections before testing further.
Try this walkthrough
- 1.Draw the workflow from input through model, output, storage, access and deletion.
- 2.Classify data sensitivity and list people affected by an error or disclosure.
- 3.Write five risk statements with impact, evidence, current controls and owners.
- 4.Choose proportionate preventive, detective, corrective and stop controls.
Watch for these traps
- Reviewing the model but not the surrounding data flow and connected tools.
- Calling all risks security risks and missing privacy or fairness concerns.
- Using precise likelihood scores without evidence.
- Treating a policy as a control without checking whether people can follow it.
Knowledge check
Choose the response you think is strongest. These checks are practice until the course version has passed subject-matter review; graded attempts will be handled server-side.
For the reviewed pilot, answers are checked on the server. This draft result is not yet added to your official learning record.
Practice studio
Make a decision. Test a response. Leave evidence.
Use the case from this lesson to practise judgement without creating a score or changing your course record.
Scenario decision
A fictional assistant receives customer notes, generates a summary and stores a log. The risk map identifies unnecessary personal data, cross-team access and unsupported claims.…
Choose a response
Branching roleplay
A colleague asks you to move this situation forward. What do you say next?
Evidence workspace
Draft a short known / assumed / next note from the fictional case. Use fictional details only.
This is a session-only practice draft. It is not saved, graded or added to your learning record. Never enter real personal or confidential information.
Practical assignment
Create a risk register for the fictional AI summary assistant. Include data flow, six risks across at least four categories, owners, controls, residual uncertainty and a stop condition.
Pass this lesson's knowledge check before recording it as complete.
Portfolio evidence
Practical portfolio project
Assess a fictional AI assistant, produce a risk register and define controls for privacy, evaluation, human review and incident response.
Turn the practical project into a bounded evidence record. Save a draft while you work, then submit it for human review when it is ready.
Sign in to save practical evidence. Sign in