When AI was science fiction, the fear was that machines would grow too powerful and slip beyond human control. Now that AI is real, and the prospect of systems that can outthink us no longer sounds far-fetched, the danger in front of us is a different one: we still cannot trust what these systems do. In a new essay, our founder and CEO William Tunstall-Pedoe argues that trustworthy AI calls for a different architecture, and sets out the one UnlikelyAI has built, which he calls STRIDE (Semantic Translation Into Deterministic Execution).

At the heart of STRIDE is the semantic artifact: a structured representation of the meaning of complex data, in a form that can be reasoned over with non-statistical methods. A policy becomes a tree of clauses and exclusions; an approved answer becomes a validated, self-contained response. The neural network does the reading, and the artifact is what the reasoning runs on, deterministically.

This inverts how AI is usually built. Neural networks are statistical approximations, so on their own they carry a permanent error rate, they hallucinate, and they cannot reliably explain themselves or give the same answer twice. A bigger model makes those problems smaller without making them disappear. STRIDE keeps neural networks at the edges, doing what they are unmatched at, and moves the reasoning, the decisions and the explanations onto symbolic machinery that can be checked.

The biggest change is one the field has chased for years: answers that can actually be explained. Because the reasoning runs on the artifact rather than inside the model, every conclusion can be traced to the rule it applied and the facts it used. As William writes, ‘a symbolic trace is not a narration of the reasoning. It is the reasoning.’ Ask a language model whether an insurance claim is covered and it may answer differently each time. Rendered as a symbolic Policy Tree, the same question is answered the same way every time, with an explanation that cites the actual clauses.

None of this is theory. STRIDE is delivering today in finance, insurance and accountancy, and it is the first step towards something larger: a shared symbolic substrate of meaning that UnlikelyAI calls Universal Language. That substrate, William writes, ‘is where trustworthy AI, built piece by piece, is heading.’

Read the full essay, Breaking Through AI’s Trust Ceiling, and follow us on LinkedIn to hear more from the team.