Drift comes in two familiar shapes. Sometimes the inputs change: customers start asking about a service you launched last quarter, or using vocabulary that did not exist when the system was set up. Sometimes the correct answers change: your prices went up, a location closed, a policy was rewritten. Either way the system keeps responding with the same confidence while becoming steadily less right.
The dangerous feature of drift is that it is silent. Nothing errors. No alert fires. The only symptoms are indirect, such as more customers asking follow-up questions, more corrections from staff, or complaints about information that turned out to be wrong. That is why the fix is a habit rather than a tool: review real conversations regularly and look for answers that no longer match reality.
For a website chat widget, the antidote is keeping the source current and having a fast way to correct mistakes. Clerkzo re-crawls your site so published changes flow through to answers, and its corrections loop lets you fix a specific wrong answer directly rather than waiting for a retrain. Reviewing your chat transcripts every so often is the cheapest drift check any small business can run.
Related terms
Browse all 142 terms- Document Object Model (DOM)The Document Object Model, or DOM, is the structured map a browser builds of your web page after loading it. Code can read and change that map, which is how anything on a page moves, updates or responds to a click.
- Model TemperatureModel temperature is a setting that controls how much randomness an AI model applies when choosing its next word. Lower values produce consistent, predictable output; higher values produce more varied and creative output.
- Transformer ModelA transformer model is a type of neural network architecture that processes text by weighing how much every part of the input relates to every other part. It is the design underpinning virtually all modern large language models.
- Large language model (LLM)A large language model (LLM) is an AI system trained on very large amounts of text to predict the next piece of language in a sequence. That single capability lets it answer questions, summarise documents, write code and hold a conversation.