Safeguarding the Arctic with AI: Why Messy Data and Context Matter

This reflection comes from the two-day Turing–Nordic workshop in Oslo, where researchers, industry, and policymakers gathered to explore how AI can help safeguard the Arctic—our seas, ecosystems, and security. It’s tempting to focus only on clean data, clear solutions, and success stories. But as the workshop reminded us, the real power of AI lies in handling messiness. Contradictions, failures, uncertainty—these are not bugs in the system, they are features of the real world.

When Context Is Everything

A heartfelt thank you to The Alan Turing Institute, NORA (The Norwegian Artificial Intelligence Research Consortium), and the British Embassy Oslo for organizing these two September days that felt different from most workshops I attend.

Waqar Ahmed from the Research Council of Norway walked through 2026–27 Horizon Europe calls with refreshing candor about what actually gets funded. Martina Ragosta from SINTEF broke down proposal writing with the kind of tactical clarity that's rare in public settings. When Nico Lang presented and included a slide on Climate AI Nordics, I was genuinely glad to see familiar faces: Aleksis Pirinen and Olof Mogren from RISE Intelligent Systems Unit. It's good to see these networks actually forming, not just being talked about.

But the moment that's stayed with me came from Malte Pedersen: 

"Useful data isn't just the good takes—it's also the ugly takes."

It landed because it's true, and because we don't talk about it enough. The messy data. The contradictory results. The experiments that didn't work. The edge cases that don't fit our hypotheses. These are the things that too often are condemned to live, at best, in the 'future work' or 'limitations' sections, the datasets we don't publish, the parts of our work we gloss over in updates.

Much of my work centers on modeling knowledge, and this insight connects directly to something fundamental when working with LLMs: everything is context engineering. LLMs are stateless functions that turn inputs into outputs. To get the best outputs, you need to give them the best inputs. As Andrej Karpathy puts it: 

"Context engineering is the delicate art and science of filling the context window with just the right information for the next step."

But here's the thing: if your context only includes the "good takes," you're not giving the model the full picture. A huge part of the value is in the context, and context is—you guessed it—often messy. It includes contradictions, failures, outliers. Sometimes knowing that three approaches failed is more valuable than knowing which one succeeded. This speaks to the relationship between implicit and explicit knowledge. If your context only includes data points themselves while excluding the rich context in the relationships, you're not giving the model the full picture and perhaps not "the right information for the next step."

This realization has implications that extend far beyond LLMs. It shapes how we should think about data infrastructure itself. A modern data stack needs to:

• Handle structured, unstructured, and semi-structured data Texture is information encoded in form itself. The way someone hesitates in a transcript, margin notes in documents, orphaned fields in JSON—these are signals precisely because they don't fit schemas. The modern stack preserves optionality: keep the texture because you don't know yet which aspects will prove meaningful.
• Extract structure from unstructured data Traditional extraction treats ambiguity as a symptom to fix. But ambiguity is often the signal. Or, more accurately, multiple signals collapsed into one.. When someone says "I think maybe possibly we should consider..." they're communicating uncertainty clearly. That uncertainty is data. Extraction should be lossy in a controlled way—you need to know what you're discarding and preserve confidence intervals alongside content.
• Combine memory and application data Application data is high-frequency, high-fidelity, low-interpretation. Memory is low-frequency, lossy, high-interpretation. These aren't different kinds of data—they're different points on a compression-interpretation spectrum. A modern stack makes this spectrum bidirectionally traversable. When memory diverges from fresh data, that's not an error—it's a signal that reality has shifted faster than your compression cycle.
• Know the difference between first-party and derived data First-party data can say "this happened." Derived data can only say "I inferred this from that, under these assumptions, at this time." The stack should make it grammatically impossible to confuse the two. Consensus is not truth, and outliers are not noise. Maintain a graph of epistemic dependency so you can trace insights back to their first-party foundation.

The Pattern

A modern data stack should be a system for managing uncertainty, not hiding from it. It preserves texture because texture encodes the shape of what we don't know. It treats ambiguity as data because ambiguity tells you where reality resists your categories. It maintains multiple temporal resolutions because truth looks different at different scales.

If you're working on problems where context is everything and messiness is a feature rather than a bug, let's stay in touch.

Let’s stay in touch

I am Stefan Wendin, and I recently became one of the unit managers at Computer Science at RISE. I wanted to share these reflections because this is not just relevant for the Arctic—it’s relevant for all of us navigating the messy realities of AI, data, and collaboration.

If you’re working on problems where context is everything, and where messiness is not a bug but a feature, let’s stay in touch.

Links

https://www.turing.ac.uk/

https://www.nora.ai/