Chapter 36 — Field or community practice
Lesson map
Lesson 36 published. Map, example, EX106.
flowchart TD
N0["Learning outcomes"]
N1["36.1 Settings"]
N2["36.2 Useful outputs"]
N3["36.3 Working with non-specialists"]
N4["36.4 Authorization and confidentiality"]
N5["36.5 Reflection"]
N6["36.7b Leaving the setting better"]
N7["Synthesis"]
N0 --> N1 --> N2 --> N3 --> N4 --> N5 --> N6 --> N7
N7 --> EX["Worked example"]
EX --> Q["Exercise EX106"]
Logical drawing
flowchart LR
N0["Learning outcomes"]
N1["36.1 Settings"]
N2["36.2 Useful outputs"]
N3["36.3 Working with non-specialists"]
N4["36.4 Authorization and confidentiality"]
N5["36.5 Reflection"]
N6["36.7b Leaving the setting better"]
N7["Synthesis"]
N0 --> N1 --> N2 --> N3 --> N4 --> N5 --> N6 --> N7 --> Q["EX106"]
[Learning outcomes]
|
v
[36.1 Settings]
|
v
[36.2 Useful outputs]
|
v
[36.3 Working with non-specialists]
|
v
[36.4 Authorization and confidentiality]
|
v
[36.5 Reflection]
|
v
[36.7b Leaving the setting better]
|
v
[Synthesis]
|
v
[Worked example] --> [Exercise EX106]
Learning outcomes
You will (1) apply AI-security knowledge in an organisation or community setting under supervision; (2) communicate with non-specialists; (3) deliver a useful, proportionate output; and (4) reflect on the experience.
36.1 Settings
Placements might be with a company, a public body, a charity or a community group adopting AI tools. Small organisations often use hosted AI assistants without policies; a well-scoped contribution can make a real difference.
Rina's placement was with a small charity whose staff used a hosted assistant without any policy. The charity had no IT staff, so the contribution had to be something staff could keep doing after she left: a short policy, sensible settings and one training session. A larger organisation would need a different shape of work, but the same principle of fitting the setting applies.
36.2 Useful outputs
Typical outputs: an AI inventory; an acceptable-use policy for AI assistants; a short risk assessment of one AI tool with recommended settings; staff training; a checklist for buying AI products. Keep outputs proportionate to the organisation's size and capacity.
For the charity, an AI inventory would have had one entry, so Rina skipped it. The acceptable-use policy and the settings change addressed the real risk, which was staff pasting case notes. A buying checklist was left for later, since the charity had no plans to buy more tools. Choosing outputs this way kept the work achievable in six weeks.
36.3 Working with non-specialists
Avoid jargon; explain risks with examples relevant to their work; give a small number of clear actions; respect their constraints. Listen first: their knowledge of the work defines what harm looks like.
At the charity, Rina explained the risk by asking staff to imagine their case notes appearing in a stranger's draft letter, rather than describing training-data retention. Staff told her that some letters had deadlines tied to benefit appeals, which shaped her advice: the policy allowed using the assistant for structure and wording, but required names and case details to be added afterwards by hand.
36.4 Authorization and confidentiality
Agree in writing what you will do and what data you will see. Do not test systems beyond that agreement. Keep the organisation's information confidential.
Rina's agreement with the charity allowed her to see the assistant's settings and a sample of anonymised letters, but not case files. When a staff member offered to show her real notes to explain a problem, she declined and asked for an invented example instead. That kept her within the agreement and avoided any question about how she handled clients' information.
36.5 Reflection
Record what you did, what worked, what you would change and what you learned about how AI security works in practice.
Rina's reflection noted that her first draft policy was three pages long and staff did not read it; the one-page version worked. She also recorded that the settings change depended on the provider's options, which she could not verify fully, and that a follow-up after six months would show whether the policy was still being followed once she had left.
36.6 Worked example (synthetic LCH case)
Rina spent six weeks with a small charity (synthetic) that used a hosted AI assistant to draft letters to vulnerable clients. Staff pasted full client case notes into the assistant. Rina's outputs: a one-page AI use policy (what not to paste, how to check outputs), assistant settings that disabled conversation retention for training where the provider allowed it, and a 30-minute staff session. A follow-up survey found 80% of staff could correctly identify what data must not be pasted.
The survey was the follow-up, run four weeks after the session.
36.7 Errors and limits
- Recommending enterprise-scale controls to a small charity.
- Testing beyond the agreed scope.
- Outputs without follow-up.
A charity with no IT staff cannot run a policy service or monitoring dashboard, so recommending them wastes everyone's time. Testing the assistant beyond what the agreement allowed, such as probing the provider's systems, would have broken the charity's trust. Without the follow-up survey, nobody would know whether staff had actually changed what they paste into the assistant. Rina's report recommended repeating the survey yearly, since new staff join and habits drift.
36.7b Leaving the setting better
A field placement ends with something the host organisation can use without the student: for example, a one-page checklist for reviewing a chatbot vendor, written in plain language and tested with the host's manager. The student's report records what the organisation chose to adopt and what it set aside.
For the charity, Rina's report noted that the policy was adopted and the buying checklist was set aside. The charity's manager kept a copy of the policy and the settings notes, so a new staff member could be briefed without Rina.
Exercises
- EX106 — Proportionate output. Propose three outputs for a five-person firm using AI assistants. Rubric: 3 = proportionate, specific; 2 = two; 1 = generic; 0 = none.
- EX107 — Plain language. Explain prompt injection to charity staff in two sentences. Rubric: 3 = accurate, jargon-free, relevant example; 2 = partial; 1 = jargon; 0 = wrong.
- EX108 — Policy rule. Write two rules for the charity's AI use policy. Rubric: 3 = specific, actionable rules on data and checking outputs; 2 = one; 1 = vague; 0 = none.
Synthesis
Field practice tests whether AI-security knowledge helps real people. Proportionate outputs, plain language and authorised work make the difference.
Continue studying — chapter 36 of 40
← Chapter 35 (draft) · Course chapter index · Chapter 37 (draft) →