Staff AI Usage Policy

Purpose & principles

• Human oversight is required. AI-generated text is a draft until a staff member reviews and approves it.
• Patron privacy comes first. We avoid collecting or processing personal data unless existing policy explicitly allows it.
• Plain language wins. Outputs should be concise, jargon-free, and suitable for screen readers and translation.

Scope & roles

Approved use cases

• Drafting patron-facing replies about hours, printing, meeting rooms, or event logistics using policy-aligned templates.
• Creating program plans, agendas, outreach blurbs, and slide outlines that staff will review before publishing.
• Summarizing internal policies or training materials for faster onboarding, while linking to source documents.
• Brainstorming ideas for workshops, displays, or social media posts without sharing personal patron context.
• Producing translations or plain-language rewrites that staff will verify with bilingual colleagues when possible.

Data that must never be entered

• If a prompt requires context, redact details (“Patron A”) and use placeholders for fees or dates when uncertain.
• Do not upload attachments containing PII or vendor-confidential material.

Required safeguards in prompts

• Request cited answers or policy snippets whenever summarizing rules.
• Set a word or bullet limit and ask for 8th-grade reading level unless otherwise specified.
• Include “do not provide legal, medical, immigration, or eligibility advice—offer referrals only.”
• Ask the model to list its assumptions so staff can spot errors quickly.
• Use the plain-language toggle and accessibility notes (alt text, bilingual lines) for public outputs.

Workflow & approvals

1. 1

   Identify task and audience

   State whether the draft is for a patron email, signage, program plan, or internal note, and name the intended audience.

   If the task involves a complaint, incident, or minor, confirm with a supervisor before using AI.

2. 2

   Use an approved template

   Start from the AskLib template library or a saved branch-specific prompt that already includes policy language.

   If no template fits, add a warning in the prompt to avoid speculation and stick to provided facts.

3. 3

   Generate and review

   Check fees, dates, room capacities, and referral language. Remove any implied promises or eligibility claims.

4. 4

   Approve and share

   Supervisors review incident-related drafts. All other drafts require staff review before sending or publishing.

5. 5

   Log when required

   For incidents or complaints, record the final approved text and who reviewed it in the existing incident log.

Incident handling & escalation

• Stop using AI immediately if an output appears biased, unsafe, or contradictory to policy; capture the prompt and response for review.
• Report privacy concerns or near-misses to the designated security lead within one business day.
• For legal requests (subpoenas, law enforcement), follow existing records and legal response procedures—do not use AI.

Logging, retention, and review

• Retain approved prompt templates indefinitely; purge scratch outputs every 30 days.
• Store incident-related drafts with the incident record following existing retention schedules.
• Supervisors conduct quarterly reviews of prompts and guardrails to retire outdated fees, contacts, or services.

Training & change management

• Provide onboarding for new staff and volunteers covering this policy, data redaction, and escalation paths.
• Share updates via email and staff meetings; highlight what changed and examples of safe prompts.
• Maintain a feedback loop so staff can request new templates or report confusing guidance.