Is AI a new form of Anthropological Intelligence?

Generative AI opens up new possibilities for qualitative research. Language models have therefore quickly become popular analytical sparring partners, capable of reading interviews, annotating images, creating overviews, identifying themes, extracting quotes, and drafting credible, well-argued report structures. Sounds great, right?

However, a number of questions arise: Can AI be used while maintaining professional control over the analytical work? And can we harness the potential of language models as text-generating collaborators without introducing unnecessary noise into the analysis phase?

At Epinion, we have prioritised dedicating time within our three innovation hubs to experiment with embedding AI as an analytical tool across the organisation.

In this blog post, I share four practice-oriented examples of our approach to AI in qualitative analysis, outlining the key considerations and principles that guide our work in this field.

1. Using Language Models to counter bias and blind spots

Analyses are often most valuable when they serve as eye-openers — challenging what we already know and raising meaningful questions about the future. This makes it essential for analysts to remain “challenged” and to maintain a broad perspective on their data. This is where AI can play an important role.

Language models excel at creating and maintaining overviews of large volumes of text. Beyond extending the cognitive capacity of analysts to process information during analysis, this overview can also be used to identify overlooked topics or potential imbalances in conclusions.

In doing so, we reduce the risk of “blind spots” or biases and help ensure that the analysis remains both relevant and sharp. This exploratory dialogue with the model’s overview can kickstart the process of uncovering additional nuances and identifying new themes across fieldwork, especially early in the analysis phase when sparring and input are most valuable.

In other words, language models can both expand our analytical capacity and challenge our own intuition and preconceptions.

2. Distinguishing between generative and verifying use

The use of language models presupposes professional expertise and close engagement with the data. Technology alone does not create a strong and effective analytical process — it emerges from the interaction between human and machine. This makes it essential that analysts themselves have been closely involved in the fieldwork and have a solid understanding of the project and research context.

The question, therefore, is not whether language models should be used, but when and how.

At Epinion, we consider early in project planning which functions language models can best contribute to. We ask ourselves questions such as:

• At what point in the analysis are we professionally equipped to engage with and challenge the model’s persuasive output?
• Should the model generate drafts, or contribute to validating existing hypotheses?

It is in the distinction between generative and verifying use that we find the most interesting synergies.

a. Generative functions

In a generative role, we use the language model as an idea generator. By attaching reference material, we can focus the model’s creative output within a defined framework. This allows us to shape the analytical expression while revisiting the data with fresh perspectives.

b. Verifying functions

We typically use language models for verification in two situations: when the analysis needs to produce insights within predefined categories, and when we want to test our own interpretations against the empirical data. In these cases, the model functions as a bias checker and helps surface overlooked perspectives.

3. Working with a closed, local and GDPR-compliant language model

Many qualitative analyses involve data that is confidential, sensitive and personally identifiable. For this reason, Epinion has prioritised access to a closed, local, GDPR-approved language model — a true game changer in working with interview data, notes and quotes. This offers several advantages:

a. Full control over data

With a closed model, we retain full control over data access. We can work directly with transcripts and notes without anonymisation or concerns about data security. Nothing is sent to external servers, and nothing is used for model training.

b. Tailored professional expertise and understanding

The model can be fine-tuned with concepts, terminology and methodological frameworks of our choosing. As a result, we work with a model that understands our professional context rather than producing generic responses.

c. A safe space for experimentation

A closed model allows for free and creative experimentation. We can develop, test and refine thematic analysis models without risking data traces or breaches of confidentiality. In other words, it is a cornerstone of our analytical process that the information we analyse does not require extensive anonymisation to be included in AI-driven analysis.

4. A shared prompt template and approach across the organisation

Creating the right conditions for creative and experimental use of language models is essential. This requires a shared prompting practice — a common understanding and approach to engaging with language models across teams.

And it does not have to be complicated. A good prompt that colleagues can easily understand is largely a matter of structure. What truly matters is that everyone on a project follows the same prompt template, ensuring transparency and shared insight into how prompts are constructed.

An example of a strong prompt might involve assigning the model a role and profession, attaching or describing the project context, defining the purpose and task, specifying language and tone, providing guidance on sources and references, and clearly stating the desired output. A starter prompt that can later be refined might look like this:

Role: You act as an anthropologist with expertise in children’s perspectives on body and health, as well as in thematic analysis and pattern recognition based on interviews.

Context: The purpose of the study is described in [attached file], with additional details from kick-off meetings [attached file] and framework conditions [attached file].

Task and purpose: Based on the provided context and framework, conduct a thematic analysis of all responses in the notes document [attached file] that address the tasks outlined in the interview guide [attached file]. The purpose is:

1) to identify patterns across interviews, andat identificere mønstre på tværs af interviews og

2) to surface potential perspectives that may not be fully covered in the study.

Sources and references: Include all informants in the analysis and reference individual interviews by quoting directly from the notes document when used as the basis for analytical insights. Embed quotes in your argumentation and specify the document name where references were found.

Output: Produce a thorough thematic analysis that follows these instructions.

Kasper Wejse
SENIOR CONSULTANT
kawe@epinionglobal.com