Why quiet drift is more dangerous than an obvious hallucination
A look at why a plausible AI answer with a small shift in category or function can be more dangerous than an obvious hallucination.
Gemini Pro
Read → Brand visibility in AI
The model remembers the trace of a brand, not the brand itself.
Atelier das Entidades studies how language models build a picture of a company from scattered textual traces: the website, product descriptions, competitor context, and recurring phrases. To a client, a brand may be clear and almost tangible, like a sign on a narrow Lisbon street. To the model, it can become a hazy category label. The lab studies where that entity gets lost — and why the error can sound so plausible.
«…a Portuguese B2B platform for client communication, [▲ a marketing agency] based in Lisbon, with around twenty-five staff…»
How the work is done
One answer at a time, before the conclusion.
For the lab, an observation is a specific model answer to a fixed prompt: how the company is named, which category it is tied to, and what neighbours appear around it. Samples are built around practical scenarios — choosing a vendor, comparing solutions, clarifying product functions.
Repeatability is sought in the semantic route that holds: if the brand is pulled toward the same wrong category again and again, there is already material for a finding.
▲ shifts the category↳ replaces the function~ pulls in a neighbour— leaves an empty space
In focus · current observation · jan 2026
«…a [↳ community-moderation] platform aimed at small Portuguese businesses, helping them manage customer feedback…»
The function replacement survives a reworded prompt. The model keeps the right category — B2B, Portugal — but loses the real work the product does.
From the observation corpus
Every analysis begins with a model answer that sounds almost right.
How a series of runs separates noise from stable drift
A piece on why one strange AI answer is not yet a conclusion, while a repeated semantic trajectory becomes research material.
ChatGPT-4o
Read →How One Prompt Sends a Brand Down Different Paths
A look at why the same AI prompt can preserve a brand in one run and blur its category in another.
ChatGPT-4o
Read →When a brand leaves a weak trace in language, the model fills in the picture on its own.
Atelier das Entidades analyses those fill-ins as observable model behaviour, in terms of repeated runs and methodological limits.
View the method →