Common misconceptions

Most workplace AI incidents begin with a false belief about the tool, not with malice. These are the misconceptions worth correcting explicitly, because each one leads to a specific mistake.

“It is basically a better Google”

A search engine retrieves documents that exist and shows you their sources. A language model generates text and, unless it explicitly searched, cites nothing verifiable. Search failure looks like "no good results"; model failure looks like a perfect answer that is wrong. Some tools now blend both, showing sources for AI answers; when sources are shown, check them, because models can also misrepresent what a real source says.

“It remembers and learns from everything I type”

Half true, and the details matter. The model does not instantly absorb your words into itself. But the provider may retain conversations, and may use them to train future models, depending on the tool, the plan, and your settings. Enterprise agreements usually exclude training on customer data; consumer defaults often do not. The safe assumption for any tool your organisation has not vetted: whatever you type may be stored and read.

“AI is objective, it is just math”

The math is neutral; the data is not. As the Amazon case showed, a model distills the patterns of its training data, human prejudices included, and then applies them with mechanical consistency. "The algorithm decided" is never an explanation that removes responsibility.

“The answer is polished, so it must be right”

Fluency is the one thing a language model is directly optimized for. It is therefore the one signal that carries no information about correctness. Judge AI output the way you would judge a stranger's claim, not the way you would judge a trusted colleague's.