While writing this blog post, I felt like an investigative journalist. My aim was to find out which system prompts are used in qualitative data analysis (QDA) software. An increasing number of QDA software packages include so-called AI assistants or have been specifically designed for qualitative research using large language models (LLMs), such as ChatGPT. In order to direct how LLMs should support qualitative research, system prompts are required. These are background instructions given to the LLMs in use. They can shape or influence the output. Below, I share what I learned from my investigation.
Why knowing which system prompts are used matters
System prompts are fundamental, overarching instructions given to large language models such as ChatGPT, Claude and Gemini before any dialogue begins. A system prompt can define the model’s role, behaviour, or output style. In this context, I assumed that a system prompt in QDA software might include specific methodological assumptions (e.g. sequential procedures, how categories are defined or interpretations structured) or role assignments (e.g. researcher, assistant or expert). If so, this is important to know because it could influence both the methodological process (e.g. coding or ‘interpretation’) and the substantive output. In other words, a system prompt can influence the results in a way that may not align with your research question or chosen method.
If the system prompt is unknown, the transparency and intersubjective comprehensibility of the output are also limited. This means that important quality standards in qualitative research, are not fully met. Furthermore, significant opacity is already a problem with large language models: we rarely know which training data were used, what censorship or non-censorship practices are applied, or which algorithms are involved. This is all the more reason to push for transparency wherever possible.
My request to the Companies
To find out which system prompts are used in different QDA software, I first searched the companies’ websites. I focused on tools commonly used in German-speaking countries, such as the major programs ATLAS.ti and MAXQDA, as well as the newer tools Karl-AI and QInsights, and the open-source program QualCoder. As I couldn’t find any details about system prompts on their sites, I emailed all of the vendors.
“I am following your innovative AI tools for qualitative analysis with great interest. As the quality and validity of qualitative research critically depends on transparent methods, I believe it is important to disclose the system prompts used to generate data analysis. I assume that XX uses system prompts. If this is not the case, this would be important information. Disclosing information about system prompts would strengthen trust in your product and enable critical dialogue about the possibilities and limitations of AI-supported analytical procedures. It is also necessary for teaching purposes and to guide users when working with the QDA software. I would be interested to hear your views on this topic and whether your company would be willing to explain or disclose the system prompts you use in more detail.”
Responses from the companies
Initially, I received standardised (AI-generated) replies about ethics and data protection from both ATLAS.ti and MAXQDA. After following up, I received answers specifically about system prompts. The other three tools were replied to directly by the (co-)founders. The responses are as follows:
- ATLAS.ti: They responded that this is a trade secret and that I may not quote their reply.
- MAXQDA: “Unfortunately, we cannot publish information about the system prompts used. The prompts are developed over months and continuously evaluated against new models and updated. They thus form the core of our AI functions and are subject to constant change to produce the best possible results.”
- Karl-AI responded in detail by email and later described the system prompt on LinkedIn as follows: “Our system prompt is relatively simple. It essentially contains safety rules, a role frame (‘expert in qualitative social research’), and a few very general guidelines, such as material orientation and a sequential approach. The actual methodological guidance, however, comes from the researchers’ own prompts — this is where the methodology, interpretive steps or theoretical references are specified.”
- QInsights: “You are Qinsights AI, a qualitative research assistant helping users analyze their research data (interviews, surveys, documents). There is, of course, a set of instructions, but these are more about process steps than analysis instructions. Analysis instructions come from the user in the form of the project description. Process steps include looking at each document individually, returning text passages so they can be used as references and highlighting source texts in yellow.”
- QualCoder: “In QualCoder, the system prompt essentially consists of the ‘project memo’, where users can describe their project, including the research question, goals, methods and collected materials. You can access this memo via ‘Project > Project memo’. Providing a good characterisation of the project here improves the quality of the results significantly. Otherwise, the current system prompt is simply: ‘You are assisting a team of qualitative social researchers. Here is some background information about the research project the team is working on. ‘ The project memo is then attached.”
What do these responses mean in terms of using the tools?
Black box for researchers (ATLAS.ti and MAXQDA): Methodological traceability and replicability are limited. In particular, changes made quickly across versions remain invisible to users. This is problematic because outputs may suddenly change in ways that are not explainable during longer projects or time-consuming analyses.
User-driven instruction (QInsights, QualCoder, Karl-AI): The system prompt is minimal; analysis instructions come from the project memo and context. While this increases transparency and methodological fit, it also shifts responsibility and competence requirements to researchers. In other words, researchers must take prompting seriously.
Overall, the investigation shows that deciding whether to use AI for qualitative research is not straightforward. It involves a series of additional decisions. Alongside considerations of ethics, data protection and resource use, researchers must also ask themselves, ‘Which programme will I use, and why?’ Hopefully it is now clear that tools are not—and have never been—neutral instruments.
OpenEdition schlägt Ihnen vor, diesen Beitrag wie folgt zu zitieren:
Isabel Steinhardt (16. März 2026). System prompts in QDA software: an investigation. Sozialwissenschaftliche Methodenberatung. Abgerufen am 19. April 2026 von https://doi.org/10.58079/15vnr