AI in f4 2026 – Serendipity Prompting and Hybrid Interpretation
AI Impuls and AI Dialog support your analysis of short text excerpts (e.g., 1–2 paragraphs) as part of qualitative content analysis, grounded theory, the documentary method, narrative analysis, or many other approaches.
These readings help challenge your own perspective and support you in developing a well-reasoned interpretation, without offering a ready-made answer or analysis.
These features are included in every f4 license—> —at no extra cost.
You select a short passage of text in your project and are asked to explain your initial interpretation. An LLM then generates three outputs: a validation impulse, a falsification impulse, and a synthesis impulse. All descriptions always refer to the specific wording of the highlighted passage. This serves as a starting point for thinking more critically about the passage. This feature is based on the concept of serendipity prompting (Krähnke, Dresing, Pehl, 2026).
Select a short passage of text and write your first interpretation of it. In f4, three LLMs generate both affirmative and contrasting interpretation variants for this purpose. You review them, write down your questions and your changed perspective, and the three LLMs respond again. Keep doing this until you have developed your own interpretation in a well-reasoned and clear manner. This feature is based on the concept of hybrid interpretation with dialogically integrated LLMs (Krähnke, Pehl & Dresing, 2025).
Function 1: For Starting a New Job
A validation prompt and a falsification prompt for your interpretation, along with an opening synthesis prompt.
Qualitative analysis begins with an observation, not with a ready-made answer. This is exactly where AI Impuls comes in. Highlight a passage in f4 and write down what you see in that section and how you interpret it. An LLM generates a methodologically guided prompt in its output that includes three aspects: a validation prompt that builds on your interpretation and attempts to further support it, to the extent possible. A falsification impulse that challenges the same phenomenon, and a synthesis impulse with open questions intended to lead further.
The AI Impuls processes only the selected passage and a small amount of context preceding it. In doing so, he is guided by a fundamental principle of sequence analysis. The statement is read on its own, and it remains unclear where it leads. (Scope of data transfer: See Privacy Policy, Section 4.)
I: To start, exactly how long have you been involved with move35 here in Marburg?
P: The first meeting took place in September. That was in ’23. And since then, they’ve been very active as well.
I: And how would you describe your work?
P: My work is basically coordinating this entire meeting. And trying to move it forward.
I keep coming back to the word “they.” It sounds detached. I think there may be problems within the group.
Validation prompt: You keep coming back to the word “they” in “they’ve been very active as well.” It sounds detached to you, and you suspect there may be problems within the group. But something stands out that begins even before “they” and that your interpretation does not yet address. The opening question is directed at P personally, yet the response begins with an event: “The first meeting.” At the very point where P is asked about their own involvement, P does not talk about themself. This can be read as a sign of detachment.
Falsification prompt: In response to the second question, P talks about themself: “My work is.” The “they” in the first response can also be read as a counterpart to this “my.” On one side is what “they” do; on the other is P’s own role. The shift from “they” to “my work” can be read as a distinction between roles and does not necessarily indicate detachment.
Synthesis prompt: Your reading leaves open what “they” says about P’s relationship with the group. At this point, the idea that there are problems is your own hypothesis. How do you see the “they” in the first response fitting together with the “my” in the second, and how does that affect your hypothesis?
The role-based reading weakens my initial hypothesis. It remains striking that, when asked about their personal involvement, P does not initially talk about themself, but later uses “my work” to describe their own role. On rereading, I notice the phrase “trying to move it forward.” Describing something as an attempt leaves open whether it succeeds. That could point to difficulties. At the same time, P describes the people referred to as “they” as “very active.” Based on this excerpt, I therefore cannot confidently conclude that there are problems within the group. I’ll see how P talks about collaboration, difficulties, and their own role later in the interview.
Behind this function lies the basic idea of serendipity prompting: The method does not plan for specific results, but rather creates conditions under which unexpected yet context-specific observations become more likely—planned serendipity, or methodical chance discoveries. The prompt is phrased in such a way that the output does not conclude, but rather leads to an open-ended question and the next step in the process: hints rather than ready-made answers.
Does a pulse and the output it generates have to be “correct”? No, it must open the door to the next step in your analysis: a point of confusion, the realization that there are different ways of looking at it, and thus an invitation to examine the text more closely. And if it doesn’t do you any good, throw it away. It’s a starting point, not a result.
Use AI Impuls when you’re exploring material for the first time, conducting a detailed analysis of specific passages, or practicing a hermeneutic approach to textual data in your teaching.
Function 2: For working on key sections
Contrasting interpretations of the same passage.
Choose a passage and share your initial observation, question, or interpretation. Based on this, three LLMs generate contrasting interpretations: They emphasize different aspects, highlight tensions, and open up alternative lines of inquiry. You guide the process and keep the options open. Your reaction determines the next round.
Interpretation A, Counter-interpretation B, and Contrast C are three model-generated interpretation variants of the same passage. The model used is indicated in parentheses (e.g., Qwen, Gemma, etc.).
In the AI dialogue, you conduct the analysis: Your observation kicks off the process, your interpretation sets the direction, and your objection keeps the analysis grounded in the material. AI outputs can go beyond the text and stretch an interpretation too far. That is precisely when the purpose of the method becomes clear: Your objection refocuses the examination on the text and makes the analysis not only text-based but also transparent.
Drawing on approaches from interpretation workshops, you will guide the interpretation process, make methodological decisions, and continually relate the process to your research interests.
The result isn’t a finished AI interpretation, but a transparently documented testing process—which serves as the basis for a memo, code development, or a team discussion.
AI Dialog is based on the published method of hybrid interpretation of text-based data using LLMs integrated into dialogue (Krähnke, Pehl & Dresing 2025). AI is not considered a source of knowledge in its own right: its outputs are interpretive variations that require verification—unresolved issues and alternative perspectives whose validity you test against the source material. AI Dialog applies this method to open-weight models in the EU in f4; for data protection reasons.
Our AI Design Principles
Today, LLMs can generate summaries, code, categories, and entire analyses. However, technical feasibility does not in itself constitute a methodological justification. To date, there are no generally accepted, method-specific validated procedures or established standards for the evaluation of automated comprehensive analyses of qualitative data.
That is why f4 does not automate the core interpretive work. Instead, AI Impulse and AI Dialogue follow four design principles:
LLM outputs can sound plausible yet still overlook, overstretch, or add meanings. In f4, they are therefore not considered results, but rather ideas that require further examination: AI Impulse raises a question, while AI Dialogue offers possible interpretations for consideration.
A single model response can quickly make a particular interpretation seem correct, even though the output depends on the model and the prompt and can vary for the same query. AI Dialog therefore presents different—and even contradictory—interpretations side by side. It’s not the model that determines which one to wear, but your own evaluation and refinement of the material.
The call for “human-in-the-loop” approaches falls short if researchers have to review entire interviews or corpora to achieve this. f4 deliberately uses short, highlighted passages. This doesn’t make the test trivial, but it does limit it to a manageable subset. Claims and exaggerations can be checked directly against the text; omissions also become easier to spot.
The more analytical steps are delegated to AI, the more important it becomes to demonstrate one’s own research contributions and accountability—both in terms of academic standards and methodology. So, in f4, you select the passage, formulate your initial interpretation, challenge it, reject it, and decide what holds up. AI creates the difference, not the result. This way, the responsibility for interpretation, methodological decisions, and the analytical process remain with you.
You can find information on current approaches to AI in qualitative research and the challenges of using AI here in the lecture notes (approx. 30 pages, PDF).
Transparent & Clear
The highlighted section, the output, and your own decision remain documented and verifiable. This disclosure serves to ensure intersubjective reproducibility, a recognized criterion of quality in qualitative research (Steinke 1999; Strübing et al. 2018).
Prompts are not a technical detail; rather, they convey methodological decisions made in advance. If they remain hidden, you may adopt an analytical perspective that you are unable to reflect upon. That is why we have decided to disclose the prompts we developed as well. You can export them and review them. Since we continue to refine the prompts over time, f4 stores information locally in your project about which passage, prompt version, model, and output belong together. The export feature helps you document and disclose everything related to the methodology section, appendices, and peer review, exactly as recommended by the DFG, for example.
Both functions are based on methodological concepts that we have developed and will (soon) publish. The hybrid interpretation has been published (Krähnke, Pehl & Dresing 2025), and the serendipity approach has been documented.
Zur Unterstützung der eigenen Interpretation von einzelnen Textstellen wurden in f4 (Version 2026) die Funktionen AI Impuls und AI Dialog eingesetzt. Sie orientieren sich am methodologisch hergeleiteten und dokumentierten Vorgehen der hybriden Interpretation textbasierter Daten mit dialogisch integrierten LLMs sowie am Serendipity-Prompting (Krähnke, Pehl & Dresing 2025). Zu ausgewählten kurzen Textpassagen wurden modellgenerierte Impulse und widerstrebende Lesarten erzeugt, von mir geprüft, abgewogen und nachvollziehbar am Material weiterentwickelt. Den Verlauf dieser Entwicklung zu der von mir gewählten Deutung, die zugehörigen Passagen, Prompts, Modelle und Ausgaben habe ich im Anhang dokumentiert.
Citation: Krähnke, U., Pehl, T., & Dresing, T. (2025). Hybrid Interpretation of Text-Based Data Using Dialogically Integrated LLMs: On the Use of Generative AI in Qualitative Research. SSOAR. nbn-resolving.org/urn:nbn:de:0168-ssoar-99389-7
Clarify before starting
There are two questions to consider before using AI in research, whether you’re using it in f4 or other QDA software. It’s best to clarify both of these before you get started.
There is no consistent stance on this issue at universities yet, but there is a widespread concern: that AI will take over the analysis for you. Depending on the examination regulations, this may be considered cheating. AI Impuls and AI Dialog are explicitly designed not to do the thinking for you, but to help you deepen your own understanding. Every step is documented and can be disclosed. Regardless, you should check with your advisors and/or the Registrar’s Office beforehand to see if this is acceptable for your situation. You can share this page as background information for the conversation.
As soon as personal data is processed, the GDPR applies. Qualitative data almost always contains personally identifiable information (interviews, observation logs, self-reports, open-ended responses, and much more). The fact that a service is technically GDPR-compliant is ONE requirement, but it is not enough on its own! You need a legal basis, which generally means written, informed consent from your participants that explicitly mentions and includes the transfer of data to our service. Anonymizing the data usually does not solve the problem in qualitative interviews, as we explain in more detail in this article on data masking. If you do not have such informed consent that explicitly mentions f4’s AI features, you are not legally permitted to use this service.
High Security Standards for Sensitive Research Data
Your research participants have entrusted you with their data. As soon as this data relates to identifiable individuals, it constitutes personal data within the meaning of the GDPR (Art. 4(1)). In this case, you are required to handle this data with care. Here’s how we’re tackling the technical aspect of this challenge:
Upon request, we will send you the complete test package; please simply send us an informal email request.
FAQ
That’s up to your college—check with your academic advisor and the registrar’s office beforehand (see “Am I even allowed to use this?”). What f4 provides: transparent, citable procedures, with every step documented—plus suggested wording in the Methods section.
AI Impuls and AI Dialog are included in the f4 license price. There are no additional costs, and there is no billing per request or per minute. You can use both features as long as your f4 license is valid. (Billing by the minute applies only to automatic transcription, not to AI features.)
The highlighted passage, along with a short, clearly defined introductory text, the prompt, and your input—and, in the case of an AI dialogue, the history of the current AI dialogue session so far. Never your project, never audio files. See the “Privacy Policy” section, item 4, for details.
Open-weight language models—such as those from the Llama, Mistral, Gemma, or Qwen families—run on the GWDG or STACKIT infrastructure in Germany. In the AI Dialog, you can see which model generated the output for each reading variant.
No—not on purpose: Observations, open questions, and interpretations. You develop the interpretation that ultimately appears in your work based on the material itself.
No. But not everyone has access to an interpretation workshop: AI Dialog generates contrasting interpretations for individual preparation when no group is available—as a way to prepare for real groups, not as a substitute for them.
For information on the centrally enforceable data processing agreement, see the Data Protection section, item 5.
f4 can do even more
Automatic Transcription
Manual Transcription & Proofreading
Qualitative Text Analysis
