title: Ojibwe Chat
aliases: []
tags: [example, tool]
summary: A public Ojibwe translation chatbot that confidently returns incorrect output.
genre: Tool
themes: [Data quality and review, Representation and misrepresentation]
technology: [MT, LLM]
stance: Cautionary

Ojibwe Chat

Ojibwe Chat is a publicly accessible chatbot that offers to, among other things, translate English into Ojibwe. The page carries no information about who built it, how it was made, what data or models it relies on, or who is accountable for what it says.

Context

A workshop participant came across Ojibwe Chat just a few weeks before the workshop. They had been watching Ojibwe video lessons online, went looking for more resources, and this turned up in the search. This very natural path to discover the tool is part of what makes it dangerous. The people most likely to seek out a translation tool are learners, and learners are, almost by definition, not yet able to judge whether the output is correct. One workshop participant, though, was able to judge its output. Here is the exchange they tested:

Notice what happens after the correction. The model simply apologizes and produces a second answer with exactly the same confidence. That second answer is also wrong.

After this, some participants suspected the tool was little more than a thin wrapper around a general-purpose LLM, so they asked it directly:

The tool appears to be a general-purpose OpenAI model dressed up as an Ojibwe chatbot. The Ojibwe-specific framing makes it more dangerous than the underlying model alone because it lends the tool an authority it hasn't earned. Intentionally or not, the framing alone confers a legitimacy the tool doesn't deserve. For learners who can't judge the output's correctness and do not have easy access to native speakers (as is the case for many extremely low-resource languages) this misinformation can be especially harmful.

Summary of issues

• Inaccuracies. The tool confidently returns incorrect output, which is especially dangerous for learners who can't judge the output's correctness.
• Lack of transparency. The tool provides no information about who built it, how it was made, what data or models it relies on, or who is accountable for what it says
• Misrepresentation. Giving a general-purpose model an "Ojibwe" face lends it an authority it hasn't earned.