Responsible Use of AI

CE Canvas includes AI features that help you plan engagement, draft documents, and analyse community feedback. This guide explains what each feature does, where human judgement is essential, and how your data is protected.

AI in CE Canvas is a support tool, not a substitute for professional judgement or community insight. Always review what it produces.

What AI does in CE Canvas

AI features only run when you choose to use them. Generated drafts are not published automatically. Your prompts, generated responses, and accepted drafts may be saved in CE Canvas as part of your project record.

How EVA stays grounded

Before EVA responds or drafts, it is designed to retrieve relevant information from your project — uploaded materials, context, canvas, plan, and results — and from CE Canvas guidance. This is called grounding. The underlying AI model also has general training, so grounding reduces unsupported answers but cannot eliminate mistakes.

• It is instructed to rely on your evidence. EVA is instructed not to fabricate specific dates, figures, budgets, or commitments, and to identify uncertainty when the available material does not support an answer. It can still make mistakes, so verify important claims against the source.
• Sources and confidence are shown where supported. Grounded drafts and answers can include source chips and confidence signals so you can inspect their basis. The absence of a warning or a high-confidence signal is not a guarantee of accuracy.
• It can identify missing context. When key information is missing, EVA is designed to flag the field as needing input instead of filling the gap with a plausible guess.
• Your data is used to answer, not to train. Grounding means your materials shape this response only — see “How your data is handled” below.
• Questions about your account are answered from your records. Which projects exist, how many, and their status come from the account’s own project records — scoped to what your permissions let you see — not from your uploaded documents. EVA cites those records as the source, and documents appear as sources only where they actually bear on the question.
• Organisation-level figures are computed, not generated. Every number in an Organisation insights answer or briefing is calculated from your records by CE Canvas — percentages included — and handed to EVA already worked out. EVA’s job there is to restate and frame them, so any figure it gives you should match one on the Organisation insights tab. It is instructed not to total, average or extrapolate beyond what it was given.

Three limits are worth knowing before you rely on an organisation-level briefing, because they are properties of the evidence and its scope rather than of the model:

• It doesn’t know what people said. The cross-project record holds what was run, who was reached, and which themes were coded — not the comments themselves. EVA is instructed never to characterise community opinion from it. To report what participants actually said, work from the project’s own results.
• Themes can’t be tracked over time. Themes are coded separately in each project, using each project’s own vocabulary, and prevalence isn’t stored historically. A theme appearing in more projects is not evidence that a concern is growing, and EVA is instructed not to describe one as rising, spreading or emerging.
• A filtered briefing says so. The briefing covers whatever the Organisation insights tab is scoped to, and is told to state that scope. Where something falls outside the filter — most often outcomes, when completed projects are excluded — it is instructed to report it as hidden by the filter, not as an absence or a failing. Even so, when a briefing will be read in public, generate it unfiltered or check the scope line before you rely on it.

EVA currently uses Anthropic Claude models accessed through third-party AI service providers. The specific model and provider may change as we improve the service.

Where human judgement is always required

AI can generate plausible-sounding content that is factually wrong, contextually inappropriate, or missing crucial local knowledge. There are specific areas in community engagement where human review is non-negotiable:

What is genuinely open to community influence EVA cannot know which decisions are fixed and which are open. The Negotiables / Non-Negotiables section of the Engagement Plan and the Decision Space section of the canvas must reflect your actual project constraints — not what sounds reasonable in a draft.

Stakeholder sensitivities and community history EVA has no knowledge of historical tensions, prior broken commitments, cultural protocols, or trust issues in your community. Sensitivities captured in Project Context and the People sections of the canvas must come from your own knowledge.

The engagement level you can honestly deliver Choosing Inform, Consult, Involve, Collaborate, or Empower is a commitment. EVA can suggest a level based on the project type, but only you can confirm whether your team, timeline, and organisational mandate can genuinely deliver it.

Closing the loop commitments The specific promises you make to communities about when and how they’ll hear what happened to their input must come from you. EVA can help structure the language — the commitment itself is human.

Any content shared publicly or with communities Review all AI-generated content before it goes outside your team. Check facts, verify context, and ensure the tone is appropriate for your community.

Practical review checklist

Before using any AI-generated content:

• Does this accurately reflect our actual project constraints and scope?
• Does it match what I know about this community and its history with our organisation?
• Are the stakeholder groups correct and complete for this specific project?
• Is the engagement level one we can genuinely deliver?
• Have I removed or corrected anything that sounds plausible but is actually wrong?

How your data is handled

• Your platform data is hosted in your customer region. Australian and New Zealand customer data is hosted in Australia, Canadian customer data in Canada, and United States customer data in the United States, as described in our Privacy Policy.
• AI requests may involve cross-border processing. When you use an AI feature, CE Canvas sends the prompt, relevant project context, retrieved source material, and instructions needed to generate the response to an AI service provider. It does not provide the provider with unrestricted access to your database. CE Canvas applies contractual, technical, and organisational safeguards where cross-border processing occurs.
• CE Canvas does not use your project data to train AI models. AI service providers process request content under their applicable commercial API data-handling terms.
• AI interactions may be retained in CE Canvas. Prompts, responses, sources, and related operational records may be stored to provide project history, traceability, security, and support. Provider-side processing and retention are governed by the applicable provider terms and our agreements with them.
• Data is encrypted in transit.
• AI features follow the same access controls as the rest of CE Canvas — team members only see data from projects they have access to.

Reporting a problem

If AI generates content that is inaccurate, biased, or inappropriate, contact us at support@cecanvas.com. Include the context or prompt you used, what was generated, and why it was a problem. Your feedback directly informs how we improve and safeguard AI features.

Our principles

CE Canvas’s approach to AI is guided by our global Responsible Use of AI Policy and applicable legal, ethical, and public-sector expectations:

• Human-centred — people remain in control; AI assists, it does not decide
• Fair and inclusive — we work to reduce bias and promote equitable engagement outcomes
• Private and secure — data is handled responsibly and kept safe
• Transparent — you always know when and how AI is being used
• Accountable — we take responsibility for our AI systems and welcome feedback