Trust is the currency of AI systems. A user who trusts an answer acts on it; a user who doesn't either ignores the system or double-checks everything it says — either way, the value evaporates. And trust in AI comes from explainability: the ability to understand and verify why the system said what it said. Here's the argument that most discussions of AI explainability miss: explainability isn't a presentation layer you bolt on at the end. It starts with retrieval — because you can only explain and cite an answer if you retrieved the sources it's based on in the first place. Trust is built from the ground up, and the ground is retrieval.
Trust requires verifiability, and verifiability requires sources
Why do users distrust AI? Because it can be confidently wrong, and they can't tell when. A fluent answer with no way to check it demands blind faith — and people, sensibly, don't extend blind faith to a system they've seen hallucinate. The antidote is verifiability: give the user a way to check the answer against its source, and trust becomes possible, because now they can confirm rather than believe.
But here's the crucial dependency: you can only show a source if you retrieved one. Verifiability requires citations; citations require provenance; provenance requires that the answer was grounded in retrieved content you can point back to. If the system generated its answer from the model's opaque internal knowledge, there's no source to show and nothing to verify. So the ability to build trust is determined at the retrieval stage — long before the answer reaches the UI. No retrieval, no provenance, no citation, no verifiability, no trust. The chain runs all the way down to retrieval.
Explainability is built bottom-up
Think of trust as a stack, built from the bottom:
- Grounded retrieval (the foundation). The system retrieves real passages from real sources to answer. This is what creates something to explain with. An ungrounded system has no foundation for trust — there's nothing to point to.
- Provenance carried through. Each retrieved passage's source (document, section, position) must travel with it through the pipeline to generation. If provenance is dropped anywhere, you can't reconstruct it — the model doesn't remember where a passage came from unless you tracked it.
- Citations shown with the answer. The answer surfaces its sources, ideally per-claim, so the user can trace each assertion to its origin.
- User trust (the top). Because answers are verifiable, users extend trust — they can check, so they don't have to blindly believe.
The order matters: the top of the stack can't exist without the bottom. You cannot add trustworthy citations to a system that didn't retrieve sources, any more than you can add a roof to a building with no foundation. This is why "add explainability" is the wrong frame — it implies a late addition, when in truth it's a property of the whole pipeline starting at retrieval.
Why retrieval quality is trust quality
It goes deeper than "you need retrieval to have citations." The quality of retrieval shapes the quality of trust:
- Good retrieval → citations that hold up. When retrieval surfaces genuinely relevant, correct sources, the citations the user checks confirm the answer, reinforcing trust. Every verification that pans out deepens confidence.
- Bad retrieval → citations that expose failures. When retrieval surfaces the wrong sources, the citation reveals the mismatch — the user checks and finds the source doesn't support the claim. This breaks trust, and rightly so.
- The honest case: no good source → decline. When retrieval finds nothing relevant, the trustworthy behavior is to say so rather than fabricate. A system that admits "I don't have a source for that" preserves trust; one that invents an uncited answer to fill the gap destroys it.
So retrieval isn't just the enabler of explainability — its quality directly determines whether the explanations build trust or erode it. Investing in retrieval quality (hybrid search, reranking, freshness) is investing in trust, because it's what makes the citations users check actually check out.
What this means for building trustworthy AI
If trust starts at retrieval, then building trustworthy AI means getting the foundation right:
- Ground answers in retrieved sources, always — never let the system answer high-stakes questions from opaque internal knowledge with no provenance.
- Preserve provenance end-to-end, so every answer can carry its citations. Design the pipeline so sources ride along with content from retrieval to generation. (Kognita returns the source with every retrieved passage for exactly this reason — provenance is built in, not reconstructed.)
- Invest in retrieval quality, because it determines whether citations confirm or expose. Better retrieval means citations that build trust.
- Show sources with answers, so users can verify — turning "trust me" into "here's my evidence."
- Decline honestly when there's no good source, preserving the trust that fabrication would destroy.
The takeaway
Trust in AI systems is built on verifiability, verifiability is built on citations, and citations are built on grounded retrieval — so explainability, and the trust it produces, starts at the retrieval layer, not at the UI. You cannot bolt trust onto a system that didn't retrieve sources to point to; it's a property of the whole pipeline, from the ground up. And retrieval quality is trust quality: good retrieval produces citations that confirm answers and deepen confidence; poor retrieval produces citations that expose failures; and the honest "I don't have a source for that" preserves trust where fabrication would burn it. The teams building AI people actually trust aren't the ones adding an explainability feature at the end — they're the ones who understood that trust is engineered from the foundation, and the foundation is retrieval. Build the bottom of the stack well, and the trust at the top follows.