AI & evidence
Every causal model in Causality is generated by an AI system. The effects, the links between them, the wording on each card and the reasoning in the detail panel are all produced by a language model responding to the scenario you confirmed. None of it is a forecast, and none of it is a simulation.
What it will not do
- No invented numbers. The model reasons in five qualitative channels only: direction, magnitude, confidence, provenance and applicability. It is not permitted to write a figure into any sentence it produces, and a model that does is rejected rather than shown.
- A figure only where a source stated one. Where a retrieved source states a percentage, that figure is copied from the source’s own words and checked against the retrieved text before it is kept. You are shown the quoted words it came from. A range appears on an effect only where two or more sources overlap, and where they disagree the disagreement is shown instead of an average.
- No invented sources. The shape the model fills in for a source has no field for a URL, a title, a publisher or a year. There is no expression it could write that puts a web address into one. Addresses enter at exactly one place: the citation channel of the search tool.
- No borrowed credibility. Two of the provenance values below — research and observed — are absent from the list the model is allowed to emit. They are added afterwards, by the server, and only when a supporting source of a qualifying type actually applies directly.
- No cost control dressed as a finding. If a search is stopped by a time or spending limit, the model says so. “We stopped looking” and “we looked and nothing applies” are different statements and stay different.
How evidence is graded
Applicability is not confidence, and collapsing the two would be the easiest way to make this app dishonest. Confidence asks how certain a claim is, and the model is the authority on its own reasoning. Applicability asks how directly the available evidence fits this scenario — its geography, its period, its population — which is a property of an outside artefact and can be wrong in ways confidence cannot.
- direct
- the source is about this scenario's geography, period and population, so its finding lands here without adjustment
- extrapolated
- the source is about a different place, period, sector or population; the finding has to be carried across, and that carry may not hold
- insufficient
- the source is on topic but does not bear closely enough on this particular claim to move it either way
Where a claim came from
- research
- supported by published research or study findings
- observed
- grounded in observed data or a documented historical case
- assumption
- an explicit assumption this model makes
- inference
- reasoned by the model, not directly supported
The other three channels
- increase
- the quantity is expected to rise
- decrease
- the quantity is expected to fall
- mixed
- it moves in different directions for different groups, or the net sign is unclear
- low magnitude
- a noticeable but marginal change
- moderate magnitude
- a clearly material change
- strong magnitude
- a change large enough to reshape the area it touches
- low confidence
- plausible reasoning, easily wrong
- medium confidence
- well-argued, with real room for doubt
- high confidence
- follows almost definitionally, or from a very well-established mechanism
Judge it accordingly
A model with no sources is not a weaker version of one with sources — it is a different kind of statement, and the notice under the scenario title always says which you are looking at. Treat a generated model as a structured starting point for your own thinking, not as an answer.
Two kinds of model are not built for you when you open them. The worked example is a recording of one real run, replayed without new AI calls, and its notice gives the day it was recorded. A shared scenario was built for, and saved by, someone else; its notice says so, and that this app did not produce or re-check what it shows. Both are AI-generated.
What happens to anything you save is covered separately, under Privacy.