Facilitator notes
Land this session early and repeat it often. The goal is to help teams use AI safely and responsibly before small design choices become product issues.
Keep the framing practical rather than procedural. Invite partners from Corporate Security, Privacy, Legal, HR, or equivalent teams to show up as collaborators who help teams learn.
Use AI safely and responsibly
Atlassian AI Builders Week curriculum
Take what's relevant and make it your own.
Why this matters
AI systems amplify choices. They automate decisions, act at scale, and can surface hidden biases from the data and instructions they rely on. Small design oversights can quickly become customer harms, internal inefficiencies, or trust issues.
Responsible AI is not a compliance checkbox. It is a way to make better product, design, engineering, legal, privacy, and security decisions together.
Start with the principles
A useful responsible AI practice starts with a small set of principles teams can apply in day-to-day work.
Open communication, no BS
Be clear about when and how AI is being used. Transparency helps users understand what the system is doing, what it is not doing, and where human judgment still matters.
Build for trust
Trust, privacy, and security are core design constraints, not follow-up tasks. Assume confidence is hard won and easy to lose.
Accountability is a team sport
AI outcomes are owned by people, not the model. Product, engineering, design, legal, privacy, security, and operations all help identify and manage risk.
Empower all humans
Design AI to support people, not replace their judgment. The goal is better work, better decisions, and better team outcomes.
Unleash potential, not inequity
AI should create value broadly and avoid reinforcing bias, exclusion, or preventable harm.
Use a responsible review before you scale
One practical habit is the responsible technology review: a structured conversation that helps teams think through users, harms, misuse, and safeguards before launch.
Do not wait until a feature is fully designed. The review is most useful when it happens early enough to influence the approach.
Put humans at the center
Start by asking who will use the system and who will be affected by it. Include direct users and people downstream of the workflow.
- What benefits should they experience if this works well?
- What harms could they experience if it goes wrong?
- What new burdens might the system create, even if it succeeds technically?
Explore best case, worst case, and misuse
Look at the system as a whole. Consider accidental failure, deliberate misuse, and cases where the system succeeds but still creates a poor experience or overwhelms internal teams.
- How could this leak or expose sensitive information?
- How could the model hallucinate, overstate confidence, or use the wrong context?
- How could a bad actor misuse this capability?
- How could this create poor experiences at scale?
Be explicit about limitations and transparency
Users and customers need to know enough to use AI appropriately. Be transparent about the role AI plays, especially when customers might otherwise assume they are interacting only with a human.
- What inputs are appropriate?
- What data should not be included?
- What outputs require review?
- When should people verify facts manually?
- How should people report issues or unexpected behavior?
Think in terms of human oversight
Human involvement is a spectrum, not a single pattern. Sometimes a human should review and approve AI-generated output before it is sent or published. In other cases, ownership, auditability, or monitoring may be the better control.
- How complex or creative is the task?
- How likely are outputs to vary or be wrong?
- Who will be affected by the result?
- How reversible is the outcome?
- How costly would a mistake be?
Handle data intentionally
Use only the data needed for the task. If identifiers or personally identifiable information are not necessary, remove them. In some cases, pseudonymous identifiers are enough. Make the decision deliberately rather than by default.
Treat prompts and model behavior as security concerns
Prompt injection and data leakage do not behave like ordinary software risks. Use layered controls: filters, access controls, review processes, data loss prevention, and product-specific guardrails.
Use prompts and instructions to reduce avoidable harm
System prompts and agent instructions can shape safer behavior, even if they are not a complete solution. Pair them with product design choices that reinforce the same expectations.
- Ask the model to use a factual, objective tone.
- Instruct it not to make assumptions about people without evidence.
- Require it to acknowledge uncertainty rather than invent an answer.
- Remind users that outputs may be incomplete or incorrect.
- Avoid anthropomorphic language that encourages people to treat AI like a person.
Know who can help
Responsible AI work is cross-functional. Know where to go for guidance before issues become blockers.
- Responsible technology or responsible AI programs for principles and review practices.
- Security teams for controls, filters, and threat scenarios.
- Privacy teams for data handling and minimization decisions.
- Legal and product counsel for go-to-market and regulatory considerations.
- Procurement and supplier legal for third-party AI tools.
- Access and enablement teams for internal AI tooling processes.
A practical way to use this
Who benefits?
Who could be harmed?
What is the best case?
What is the worst case?
How could this be misused?
What do users and customers need to know?
What level of oversight is appropriate?
What data is actually necessary?
Which team should we involve now?
In short
Using AI safely and responsibly is less about one perfect policy and more about building the habit of deliberate decision-making. Start with principles, review risks early, keep humans at the center, be transparent about limitations, use only the data you need, and treat responsible AI as shared work across teams.