PlanetWise — Trustworthy AI for Climate Policymaking

Context

AI locked out

civil servants are often prohibited from using public GenAI platforms over data privacy, security, and accuracy concerns

Data overload

"we receive too many reports weekly to analyze deeply in available time" — Paul Vetter, Dutch Ministry official

Beyond-human

inspired by James Bridle's Ways of Being: ecosystems and biodiversity as stakeholders in the policy process

Climate policymakers work under heavy information loads and tight deadlines, yet the AI tools that could help are either off-limits or untrusted. PlanetWise asks what a GenAI tool would look like if it were designed for this reality from day one.

Challenge

How can we design a generative AI tool for climate policymakers that leverages AI's strengths and addresses its limitations through transparency, expert validation, and multi-perspective analysis?

Research

Our process combined speculative design with Agile methods. Early on, limited access to policymakers meant we leaned on desk research and expert literature, iterating on concepts weekly and validating with subject-matter experts later in the process.

Desk research

policy workflows, trust in AI, and AI ethics — from the EU AI Act to OECD and UNESCO frameworks

Stakeholders

interviews with Dutch Ministry officials, the Australian Climate Council, and AI experts

Weekly iteration

paper sketches → Figma prototypes → expert feedback, every week

I was involved in every stage from initial research to final prototype, led concept development and ideation, and coordinated the stakeholder interviews and expert consultations.

Key insights

Transparency is a precondition, not a feature.

Three priorities emerged from the evidence base and directly shaped the final concept:

Transparency in AI reasoning

Policymakers mistrust "black box" outputs; sector-specific models are preferred over general-purpose LLMs.

Mandatory expert validation

"AI findings always must be checked by our experts before informing policy decisions." — Sara Saberi, Australian Climate Council

Multi-perspective analysis

Seeing a range of outcomes — ecological, social, economic — supports better evaluation of trade-offs.

Exploration

Each ideation round paired a research insight with a concept, then tested it against expert feedback — and the feedback redirected us more than once:

Concept PEET prompting wizard + real-time guidance for ethical, inclusive prompts

→

Expert feedback Ministry officials found prompt coaching less critical than clear, structured outputs and expert review

→

Outcome Dropped — focus shifted to transparency and workflow integration

Concept Scenario generator — input a policy, view ecological, social and economic implications

→

Expert feedback An interesting "nice-to-have", but not the main focus

→

Outcome Kept as a supporting feature behind transparency and validation

Concept Color-coded validation — experts flag bias, misinformation, and ethical issues directly in AI outputs

→

Expert feedback Valued — with requests for clearer flag explanations and closer collaboration-tool integration

→

Outcome Became the Human Validation module at the core of the concept

Dropped concept: real-time guidance flags biased framing ("Why is renewable energy bad…") and suggests a balanced prompt — set aside after ministry interviews prioritized structured outputs.

Solution

PlanetWise answers a policy question in multiple perspectives — general, environmental scenarios, human expertise, and reasoning — built on climate-specific models (ClimateGPT) with curated datasets, designed as a desktop application for secure governmental networks.

Multi-perspective answers Ecosystem scenarios Visible reasoning & sources Expert validation workflow Presentation-ready proposals Reference library

1 · Ask a policy question — the general perspective. Policymakers comment for colleagues, give feedback to the AI, and add validated text directly to their proposal, with sources visible.

Ecosystem scenarios — short-, mid- and long-term "what if" futures with environmental data, giving voice to beyond-human stakeholders.

The transparency dashboard — every step of the AI's reasoning with clickable references, in line with the EU AI Act.

2 · Human validation — AI suggests which expert roles should review the result; proposal generation stays locked until every review is done.

3 · Create proposals — once validation is complete, PlanetWise generates a concise two-page report with verified content and visualizations, matching how ministers actually consume information.

Validation

In the later stages I conducted a user test with a Deloitte senior manager specializing in AI and data analytics, which reshaped the final experience:

Communicate value on the homepage Step-by-step linear flow Reasoning over confidence scores Scenario imagery only in final outputs Collaboration in a dedicated panel

"Transparency — through clear references and reasoning steps — will build more trust than an AI-generated confidence score." — Alireza Azarfar, Senior Manager AI & Data Analytics, Deloitte

Reflection

Honest limitations matter in AI work: the confidence-score indicator risked being misleading (it's generated by the AI itself), the tool remains more human-centered than its beyond-human ambition, and limited access to policymakers constrained primary testing.

This was my first time designing for the public sector with AI as both the design tool and the subject — navigating legal, ethical, and ecological dimensions at once, and learning to advocate for design decisions inside a team under tight deadlines.