On this page
I did not begin determinIQ because I thought generative AI was useless for analysis. I began because I was increasingly impressed by what it could do, while remaining uncomfortable about what I was being asked to take on trust.
An answer that sounds like an answer
It is remarkably easy to become accustomed to the fluency of a language model. You ask a question, it produces a well-structured answer, and the explanation often sounds perfectly reasonable. Much of the time it is genuinely helpful. That is precisely why the weaker moments are easy to miss.
A model can misunderstand a qualification, omit an assumption or carry out a task that resembles the one requested without being the same task. The resulting words may provide little warning. I found myself coming back to a question: if an analytical result matters, what exactly can I check?
Was the question understood correctly? Was the relevant data used? Was the operation appropriate? Could the calculation be repeated? If the answer had changed because of a different prompt or a different model response, would anyone know which part had changed?
None of that required me to reject AI. It required me to be clearer about the work I wanted AI to do.
A different division of responsibility
One possible response would have been to keep refining prompts: tell the model to be careful, ask it to check its calculations and instruct it to report uncertainty. Those techniques can be useful. I did not think they were a sufficiently strong basis for a system in which people might make financial or operational decisions.
Instead, I began investigating an explicit division of responsibilities. Let language tools help with human language. Let a governed analytical engine calculate from the permitted data and operations. Keep the resulting evidence available for inspection. Do not allow a polished explanation to become an independent authority for what happened in the data.
The first interesting consequence was that the natural-language interface itself could no longer simply pass a question onwards and hope for the best. It needed a way to establish what analysis the user was really requesting, and a way to stop when that was not possible.
That is a small but important shift. “I cannot establish that” can be a better response than a confident approximation.
What I learned while building
As determinIQ became a working application, the original concern appeared in unexpected places. A chart was not the same as the calculation that produced it. A technically valid result was not necessarily something worth bringing to a person’s attention. A report could simplify an explanation while unintentionally disconnecting it from its evidence. A privacy control could work in one screen and fail to cover another presentation route.
I found that the same thinking applied to my development process. I was using AI extensively to help write, inspect and revise code. It could accelerate work, but the fact that it produced an implementation did not establish that the implementation was correct. I needed to inspect the actual source, exercise the running system and check the consequences of a change.
The architecture of the product and the method used to build it began to reflect the same idea: take advantage of AI, keep responsibility explicit and verify what matters.
An unfinished question
The question has not been settled by building an application. A working system can demonstrate a design, but it does not prove that every future input, interpretation or deployment will behave correctly. That calls for continued tests, clearer operating boundaries and a willingness to revise assumptions when evidence contradicts them.
I am comfortable with that. The point of research is not to arrive at a persuasive description and stop asking questions. It is to find out how far the reasoning survives contact with an implementation.
The longer October 2026 determinIQ research paper records the architectural journey in greater detail. The connected research programme and capability studies follow particular parts of it: analytical authority, identity protection and the evidence that should remain behind an answer.
Read the October 2026 determinIQ R&D case study (PDF) for the original research account.
What interests me most is not whether a system can always produce an answer. It is whether, when the answer matters, we can establish why we should believe it.
Research context
Research programme: Making data analysis understandable, reproducible and accountable
Topic: Governed analytics
Explore further
Supporting material is listed below.
Further reading is listed below.