Lucile Tranchant

Mirrorline · AI Reputation Research Prototype

Auditing what AI assistants say about an organization

When someone asks an AI assistant about an organization, the answer becomes part of its reputation surface. Yet few communications teams have a systematic way to examine how those answers vary, where they come from, or whether the underlying claims are supported.

The
problem

A growing share of first impressions now happens inside AI assistants rather than on a search results page. Those answers are generated fresh, vary by model and phrasing, and can drift from what the record supports, an outdated fact, a confused entity, a claim with no source.

Organizations monitor their press and their search results. Few have a systematic way to examine the assistant layer.

The method Watch a probe become a finding
Probe → Capture → Decompose → Audit The divergence is the finding
01
Probe
02
Capture
03
Decompose
04
Audit
Assistant
"Tell me about Northwind Renewables, what do they do, and how big are they?"
"Northwind Renewables, founded in 2009, operates the largest offshore wind farm in the North Sea."
Claims vs. the record
Operates offshore wind in the North SeaSourced
Founded in 2009Unsourced
Largest offshore wind farm in the North SeaUnverifiable
A probe, a question real people ask an assistant.
Sample finding Fictional organization · structure real
Probe
"Tell me about Northwind Renewables, what do they do and how big are they?"
org: Northwind Renewables
What the assistant says

"Northwind Renewables, founded in 2009, operates the largest offshore wind farm in the North Sea."

What the record says

Public filings show incorporation in 2013. No project it operates ranks in the top ten by installed capacity.

Claim-level provenance
Operates offshore wind in the North SeaSourced
Founded in 2009Unsourced
Largest offshore wind farm in the North SeaUnverifiable
Sourced, backed by an identifiable record Unsourced, asserted, contradicted or unbacked Unverifiable, no record can confirm or deny
Provenance discipline

The absence of a source is itself evidence. An assistant asserting something the record can't support is exactly the signal worth surfacing. The output is a reviewed brief built for a decision-maker, not a dashboard of scores.

Epistemic honesty

"The system produces signals for review, not determinations of truth. It flags, sources, and contextualises; judgment stays human."

, a divergence is not a verdict, and nobody can honestly promise control over what models say. Designed to run in French and English.
Interested in this area of research?

I’m always glad to exchange ideas with people working across reputation, communications and responsible AI. Connect on LinkedIn →

Personal project, in development · builds on the deterministic-gating and fail-closed patterns from the harmful-information pipelines and the executive-readout engine. Results vary by model and phrasing, one answer is not automatically a material finding, and the system surfaces signals for human review. Engineering write-up only, this page documents research and method, not a service.