How will people respond to a new rule, product or news story? MiroFish attempts to simulate these situations with AI characters. Its creators present it as a prediction system. I would first treat its output as material for asking better questions.
What the project describes
According to its repository, MiroFish creates a simulated environment with agents based on supplied information. The agents have characteristics and react to each other. Their interactions inform a report, and users can ask further questions within the simulation.
That is an interesting approach to exploring scenarios. The creators’ description is not independent evidence that the reports reliably predict reality.
A practical example
Suppose you want to change a subscription. A simulation may surface objections you have overlooked: unclear terms, awkward timing or a group of customers responding differently.
I would use those objections to prepare interviews or a small real-world test. You still need to ask actual customers whether the issues matter. Otherwise, the simulation may largely reflect its starting assumptions.
What would make it more convincing?
I would want to compare predictions recorded in advance with what subsequently happened. I would also want a simple baseline and an understanding of how changes in the input affect the result.
More agents do not automatically make a prediction more reliable. Agents sharing similar assumptions may reinforce them. A long report can appear more certain than its evidence warrants.
MiroFish is a project I would follow critically. Exploring possible reactions can be useful; relying entirely on its predictions is a different decision. This assessment is based on the project description, not my own benchmark.
Sources and further reading
Explanation and commentary based on the sources below. Examples are illustrative.