A personal note: On Tuesday, 28th April, one month ago, I posted this on LinkedIn: "In between courses, I built Understories, a prototype at the intersection of communications and AI. Understories gives climate advocates research-backed feedback on their messaging."

Merging artificial intelligence with climate communication is a contrarian bet I’m willing to defend. I believe everything — not anything — should be tried and tested to overcome what I see as a story problem, with internal enemies: a tendency to over-engineer messages.

What follows is an account of where things stand. To everyone who engaged with the original post or reached out directly: thank you.

Caroline - 26 May 2026

—————————————

1 . What the numbers show

The multiplier effect: 1 post, dozens of clicks on the link, dozens of messages analysed, hundreds of unique visitors, thousands of events on the website.

• The geographical distribution: 1 post, 26 countries: Australia, Canada, Colombia, Czech Republic, Dominican Republic, France, Germany, India, Ireland, Italy, Morocco, Netherlands, New Zealand, Poland, Portugal, Romania, Slovakia, South Africa, Spain, Switzerland, Thailand, UK, Uruguay, US, Vietnam, Zimbabwe.

• Uneven traffic: Three countries account for more than 50% of visits: the US, the UK, and the Netherlands. But the post and website link have travelled further than I would have expected — all the more striking given the English-language barrier: Thailand, Vietnam, Zimbabwe, Morocco, Dominican Republic.

The number that stayed with me was 26 countries in one month with a unique organic post, the friction of being desktop-only, no paid promotion, no cross-posting. The problem Understories is solving isn’t local — it’s structural.

2. The tech and its surprises

The tool's limits are a map of AI's limits in this space. Three orchestration tools, three architectures, three LLMs to find what’s now a stable structure. This is what I learned in the process.

• The free choice: Understories is built with free tools only — I haven’t subscribed to any AI system. The implication is deliberate: anyone can do this. My training on AI systems and expertise in communications helped, but autodidacty is the real foundation here. Ethically and environmentally, lighter models, lower compute, no data sold to fund the infrastructure: I think that’s a feature, not a bug.

• The weight of the research: My own synthesis of 50 reports and academic papers was too heavy for a free model to carry on iteration — it came down to 7 key syntheses. After a submission cap, the tool downgraded the LLM for one of the agents without warning. I didn’t notice a difference in evaluation quality, probably because the initial training was sufficient to cover the gap. The initial goal was to test whether the foundation and architecture were working. They do.

• The hallucination: One occurred, from a known source, in a closed context. It was contained. For a RAG designed to work in a governed, context-aware loop, that’s within an acceptable range — and worth being transparent about.

You “rent” the tool you use; you don’t own it. The dependency variable is one of many considerations to have in mind. To integrate the 50 research analyses, I would have to build an external vector base; this is the possible next step. The LLMs used were Claude Sonnet 4.5 and Claude Haiku 4.5. Lovable was considered but not used for this build.

3 . Perspective from the outside

An observation: almost no one pushed back. The comments were warm: “This is so cool,” “This is impressive,” and privately: “This could be a game changer.” I had prepared for backlash. It didn’t come.

That’s worth interrogating. Did people self-censor out of kindness? Was the audience already convinced? Is the tool too niche to provoke disagreement — or simply not yet visible enough to attract it? In a space as contested as AI-assisted communication, the absence of public critique is itself a data point.

One Sociology researcher offered the harder questions I was looking for. I’m grateful. Three points, and my replies:

• The authenticity paradox: Using AI to write conviction-led content risks killing the very thing that makes it convincing. The greatest and most engaging speeches are personal and carry a human voice [Understories doesn’t write — it edits and balances an existing piece. The human voice is the star ingredient. No AI tool replaces that.]

• The psychosocial toll: Tools like this could deepen what many workers already feel — that their job is meaningless. AI degrading self-worth is a real point of attention. [True. But also: pending weeks on back-and-forth only to publish a version you know is weak or performing tasks that don’t change anything: that’s the bullshit job. Understories is a reminder to be audience-centric first — not a replacement for judgment.]

• The replicability point: This is a RAG anyone can build. The entry barrier is low, and a series of prompts can do the job. [Exactly. It can be replicable. The moat of Understories sits at the intersection of three governance layers, each holding its level of complexity: a research base, an 8-criteria focus, and an ethical framework. Will communicators have the time or resources to build it? Climate communication should be part of the commons. This is an attempt to move the conversation in that direction.]

Another thing that surprised me: not a single question about carbon footprint, AI model choices, English-language bias, or open-source alternatives. Those silences are as telling as the critiques that did arrive.

—————————————

On Ethics and Environment

Artificial intelligence has been described as "extracting Earth's geological history" by USC researcher Kate Crawford. The environmental and ethical costs of AI are real and demand accountability.

Organisations can develop an AI policy addressing: which tools are used and why, their energy footprint, data sovereignty, and how AI augments rather than replaces the communicators and frontline voices it is meant to serve.

—————————————

4 . What’s next

I received questions about what comes next. They came in various forms. Three considerations worth sharing:

• The business model: Understories began as an exercise, after working on a similar project during an AI program at MIT, where financial viability was part of the thinking. The honest tension is between two things: a tool like this should be free and governed by the communities that use it, and free doesn’t sustain itself. The financial scenarios are familiar, but I haven’t resolved the tension. It can also simply live as a research pilot: a proof of concept for others to build something similar, better suited to their needs.

• The scope question: Understories is built for last-mile messaging — the moment when you need to reach a real audience in vivid, actionable terms. That makes it less relevant for technical or internal communications, and that’s fine. The harder question it raises is for organisations that want their work to resonate beyond their immediate field: do you actually want to build a movement? If yes, communications can’t be an afterthought. Understories is an attempt to make that question harder to avoid.

• The research base: Building Understories took about four weeks part-time, spread between February and March, and 75% of that time went into my own synthesis of 50 reports and academic papers, most of which I had already read and studied before. The stack is still not finished. I am adding methodology notes and licensing status (Creative Commons, etc.) as I go. I plan to share the full research stack this summer with whoever is interested. 

Climate is a story problem. I’m glad I’ve been able to work on it in my own limited capacity.

——

The monthly update is shared by email first to users, researchers, AI and comms practitioners I happen to meet while building Understories.

I publish every other Thursday, a "newsletter" you can subscribe to: The Reading List. It's made of short explainers, key findings, and data points from selected reports as they go live, plus exclusive content.