AI responsibility

Cities are complex. The decisions made about them — where a school goes, which neighbourhood receives investment, how a street is redesigned — affect millions of people, often for decades. When AI enters that decision space it has to earn trust the same way any other tool does: by being transparent about how it works, where it can fail, and what good practice looks like when using it.

GUS (the Geographic Understanding System) is the AI assistant that powers our analysis layer. It helps planners, policy-makers and operators interrogate large quantities of urban data in plain language. It is fast, broad and useful — and, like every AI system, it is fallible.

This page explains, in plain terms, what GUS is, where it can be wrong, the controls we have built to reduce that risk, and the role we expect a human expert to play. We treat this as a living document and update it as the system, the data behind it, and our understanding of responsible AI evolve.

Our position in one sentence: GUS is designed to make qualified humans faster and better informed — not to replace expert judgement, due diligence or democratic accountability.

1 · How GUS works

GUS combines three layers: a curated geospatial data foundation, a set of analytical models tuned for urban questions, and a large language model that translates between human questions and structured analysis.

The data foundation

GUS draws on a wide range of authoritative sources — government statistical offices, transit authorities, satellite-derived measurements, OpenStreetMap, mobility providers, environmental sensors and our own primary research. Every dataset is documented with its source, vintage, spatial resolution and known limitations.

The analytical layer

On top of the data sit deterministic models for things that should never be guessed — distance calculations, accessibility scoring, demographic intersections, density measures. These run as inspectable code, not as language-model output.

The language interface

A large language model interprets your question, decides which datasets and analyses are relevant, calls them, and summarises the result. The LLM is not the source of truth — it is the interpreter. Results are returned with the underlying datasets, queries and citations so the work can be traced back to its origin.

The split matters. Where the answer requires arithmetic, geometry or a defined methodology, GUS routes the work to deterministic code. The language model is used to navigate, explain and contextualise — never to fabricate numbers it should be looking up.

2 · Why AI can be wrong

Even with a careful architecture, AI systems can produce mistaken or misleading answers. Being honest about how that happens is the first step in defending against it.

Pattern matching is not understanding

Language models generate responses based on statistical patterns in the data they were trained on. They have no first-hand knowledge of your city. When a question falls outside familiar patterns, an LLM can produce a confident answer that is subtly — or completely — wrong.

Urban data is messy

Census boundaries change. Definitions of “household” vary between countries. Land-use codes are inconsistent. Sensors have outages. Two seemingly identical datasets can be measuring different things. AI cannot always detect these inconsistencies, especially when the underlying metadata is incomplete.

Statistics have limits

Small sample sizes, sampling bias, suppressed cells in public datasets and confidence intervals can all distort a tidy-looking number. Headline metrics often hide important variation between neighbourhoods or demographic groups.

Bias travels with data

If a city historically under-counted residents in informal settlements, models trained on that data will under-serve the same residents. We try to surface these gaps, but no system can repair history. AI inherits the choices made by the humans who collected, classified and published the underlying data.

Complex relationships are easy to misread

Cities are systems of systems. Causation, correlation and coincidence look very similar in a chart. AI can describe what is correlated but cannot, on its own, decide what caused what. Causal claims require domain knowledge, triangulation and — sometimes — experiments.

3 · Our safeguards

We build GUS as if every answer might be challenged in a planning committee — because eventually one will be. The following safeguards are baked into the platform, not bolted on.

• Real-time validation against authoritative sources. Where authoritative data exists, GUS is wired directly to it rather than relying on cached or modelled approximations. Validation rules flag responses that drift outside expected ranges.
• Cross-referencing across multiple data sources. Important findings are triangulated. If one source says population density is X but two others say Y, GUS surfaces the disagreement instead of silently picking a winner.
• Transparent methodology, citations and sources. Every answer can be expanded to show the underlying datasets, how they were combined, which spatial unit was used and what the confidence interval is. Users can follow citations through to the original publisher.
• Determinism where it matters. Calculations that should never be guessed — distances, counts, intersections, accessibility scores — are run by deterministic code, not language models. Two users asking the same question of the same dataset get the same answer.
• Regular model updates and quality monitoring. We continuously evaluate the system against known-good benchmarks, regression test suites and reviewer feedback. When models drift or new failure modes appear, they are investigated, fixed and documented.
• Uncertainty made visible. Where GUS is not confident — small sample, stale data, methodology that doesn’t fit the question — it says so. A calibrated “I don’t know” is more useful than a confident guess.

4 · Human in the loop

Some decisions are too important to be left to a machine. Our platform is designed around the idea that the human is always the decision-maker — GUS is the analyst at their elbow, not the authority at the head of the table.

Human review is required for:

• Material decisions — investment, planning consent, public spend, regulatory action.
• Equity-sensitive analysis — anything that touches vulnerable groups, protected characteristics or public services.
• Forecasts and projections — model outputs that depend on assumptions which should be challenged before they are quoted.
• Causal claims — “X caused Y” must be evaluated by a domain expert, not a language model.

For our delivery work, qualified analysts review GUS outputs before they leave the studio. For platform users, we provide the same evidence base — citations, methodology and uncertainty — so that your team can apply equivalent judgement.

5 · How to use GUS well

A few habits dramatically improve the quality of work that comes out of GUS — and out of any AI assistant.

• Treat the first answer as a draft. Use GUS to surface candidate insights, then probe — change the question, change the spatial unit, look at the raw data, ask “what would I expect this to be and why.”
• Verify the load-bearing numbers. If a number is going into a board paper, a planning report or a policy decision, click through the citation. Reading the source for thirty seconds prevents a great deal of embarrassment.
• Cross-reference with the people who know. A community worker, a transit operator or a long-tenured planner often knows things the data hasn’t recorded yet. Use GUS to frame the conversation, not to short-circuit it.
• Ask GUS what could be wrong. A useful prompt: “What are the strongest reasons this analysis might be misleading?” The system is reasonably good at red-teaming itself when invited to.
• Document the chain of reasoning. When GUS supports a recommendation, capture the question you asked, the datasets used and the assumptions involved. Decisions held to public scrutiny need a paper trail.

6 · Known limitations

We try to be candid about what GUS is not yet good at. The following are areas where we recommend particular caution.

• Very recent events. Some datasets refresh monthly, some annually. For fast-moving situations — a closure, an incident, a brand-new development — GUS may not yet reflect the on-the-ground reality.
• Informal, undocumented or contested data. Informal settlements, unregistered businesses and disputed boundaries are systematically under-represented in official datasets. We supplement these where we can, but gaps remain.
• Hyper-local nuance. GUS is strong at neighbourhood, district and city scales. For an individual building or a single street, on-site observation often beats remote analysis.
• Predictions and futures. Forecasts depend on assumptions about behaviour, policy and economics. GUS can model scenarios; it cannot tell you which scenario will happen.

7 · Reporting and feedback

If something looks wrong, please tell us. The fastest way for GUS to get better is for the people using it to flag what does not match their on-the-ground knowledge.

For a specific data point or output that looks incorrect, email ai@silacities.com. Please include a link or screenshot of the answer, the question you asked, and the source you believe is correct. We acknowledge reports within two working days.

For broader questions about our approach, partnership on responsible AI, or research collaboration, email hello@silacities.com.

For our wider commitments to data, security and editorial standards, see our transparency report, security page, data processing notice and editorial policy.