Not all AI is equal: Model choice and purpose-built tools

Not All AI Is Equal: Why Model Choice and Purpose-Built Tools Matter

Most people assume that large language models (LLMs) like ChatGPT or Claude reflect “the internet” or “the data.” But that's only part of the story—and often the least concerning part.

As LLMs become the new gateways to knowledge, the question isn’t just what they know, but how they’ve been taught to respond. A recent TechOasis article

meco.app/share/fe2fc5e6-1d6c-49e5-b5db-32e… laid out a stark reality: AI models don’t simply reflect the web—they reflect their makers.

Bias can enter a model in at least three ways: • During training, through the data selected (or excluded) • During alignment, when developers shape what models are allowed to say • Through hidden system prompts, invisible instructions that heavily steer every answer

Many users are unaware of the system prompts running in the background of general-purpose chatbots. These prompts are often extensive and can be used to filter, shape, or even suppress certain types of content. And while alignment is necessary to avoid harm, it also gives the companies behind these models significant power to define what is “acceptable” knowledge or opinion.

This is why model choice matters—especially for researchers.

This is why it makes a difference whether you use a general-purpose chatbot or a purpose-built tool. Yes, designated tools also make use of large language models—but that doesn’t mean they simply have ChatGPT "built in." 😉 (And yes, I’ve seen this claim more than once—even in academic articles.)

At QInsights, for instance, we do use LLMs. But we’ve built our platform very differently. Our system is based on a Retrieval-Augmented Generation (RAG) approach, which means that responses are grounded in the data you upload—not in the model’s pre-existing beliefs or pre-tuned worldview.

When you ask questions or run a thematic analysis in QInsights, you’re not pulling answers from the entire internet or from hidden system prompts. You're engaging with your own dataset, and our system shows you the source material that supports each response. That transparency matters.

But here’s the key point: no AI model is neutral. Whether it's OpenAI, Anthropic, or Mistral, each embeds its own design choices and values. That doesn’t mean you shouldn’t use them—it means you should understand how to work with them.

AI can be a brilliant assistant in qualitative analysis, but only if we remain in the driver’s seat. That means: • Knowing where the answers come from (in Qinsights case, based on your data). • Understanding that interpretation is a human task • And staying critical—even in a co-intelligent dialogue with AI

So next time you use an AI assistant—whether it's in a chatbot or an analysis tool—ask yourself:

Whose voice is speaking? And how will I engage with what it says?

Because while you may not control the model’s responses, you are responsible for how you interpret and use them. This is especially true when it comes to meaning-making. In qualitative analysis, we need to remain reflective and critical—not just about our data, but about the tools we use to explore it.