Manifesto:
Human-Centric AI in Empirical Research

Preamble

We dedicate ourselves to the empirical research of media environments – from the quantitative content analysis of public opinion and the structuring of complex TV and radio broadcasts to qualitative document analysis and the examination of pop-cultural artifacts. The integration of Artificial Intelligence (AI), particularly Large Language Models (LLMs) and multimodal architectures, marks a qualitative leap in the efficiency and depth of our methodology. However, we do not view AI as a replacement for the researcher, but rather, in the spirit of the Extended Mind Theory, as our methodological prosthesis: it undertakes the syntactic structuring of vast amounts of data to create cognitive space for the true essence of science – human expertise.

Driven by this conviction, we commit ourselves to the following six principles and the operational guidelines derived from them.

Part 1: The 6 Principles of Epistemological Responsibility

1. Epistemological Sovereignty: We delegate tasks, never meaning

AI is the master of probability and syntax; humans are the masters of meaning (semantics) and intention (pragmatics). When AI adapts coding schemas or fills in variables, it provides statistical suggestions. Defining what a code means within the research context is exclusively the responsibility of the human project leadership.

2. AI as an Extended Mind: Automation to sharpen expertise

We use AI-supported tools not to replace researchers, but to relieve them of mechanical burdens. The AI can perform preparatory work (prefill): structure, suggest, optimize - but the human performs the validation, the in-depth analysis, and the interpretation.

3. Protection of the Anomaly: Resistance against the statistical mainstream

AI tends to smooth data and converge toward the statistical mean. However, in PR and media research, the most relevant insights often lie in the deviation (the "shitstorm", the innovative framing, the biographical anomaly). Our researchers commit to actively searching for the anomalies that the algorithm discards or standardizes as "noise".

4. Contextual Grounding: Translating from the data vacuum into social reality

An algorithm calculates correlation coefficients, an AI assistant generates text modules for reports – but no algorithm, no AI assistant can know what this means for the social, political, or historical reality of an investigated entity. The reverse translation of bare numbers into a discursive, power-critical, or PR-strategic reality is the core of our human research performance.

5. Ethical Research Design: The human as the final authority

Empirical social research is an ethical act. The decision of which data sources are trustworthy, which biographical details must be handled sensitively, and how results are formulated to be methodologically sound and socially responsible cannot be calculated. The human researcher remains the ethical guardian of every project.

6. Primacy of the Research Design: Initiation through human research interest

The use of AI is never initiated by the AI itself or dictated by its mere technical capabilities. Any initial integration – be it through prompts, RAG systems, or agents – is derived exclusively from the research interest and the investigation design of the researchers. AI may not be used for its own sake or primarily for economic or purely technical motives, but always with scientific consciousness as a targeted methodological instrument.

Part 2: Operational Guidelines for Everyday Research (Code of Practice)

To anchor these principles in everyday research, strict Human-in-the-Loop (HITL) guidelines apply to all phases of our projects (from preparation to reporting).

A. Adaptation of Coding Schemas and Research Design

• Use of AI: We use AI, for example, to extract existing documents, structure them, and adapt them to research questions. AI serves as a tool that complements, rather than replaces, human expertise.
• Human Expertise (Review): An AI-generated or adapted document, such as a coding schema or report, must never enter the field unchecked. Senior researchers verify definitions for theoretical foundation (discriminant validity, construct validity) and adapt categories to the specific situational context of the current research interest.

B. Data Collection, Coding, and Structuring

• Use of AI: Multimodal models handle the primary structuring of unstructured data. This includes: transcription of TV/Radio (e.g., via Whisper), automated generation and clustering of screenshots, and the transformation of transcripts into codable structures. The AI also generates "prefills" for coding masks.
• Human Expertise (Review):
  • Manifest vs. Latent Variables: Manifest variables (e.g., length of the segment, broadcaster, present persons) may be completely filled out by the AI following successful reliability tests.
  • Latent, interpretive variables (e.g., irony, political framing, emotional tonality) are merely suggested (prefilled) by the AI. The human coder must review these suggestions, evaluate them critically, correct them if necessary, and finally approve them.
  • The control of transcription quality is carried out on a sample basis by trained researchers to prevent systematic distortion (bias) concerning dialects or technical terminology.

C. Analysis, Evaluation, and Reporting

• Use of AI: Automated, established mathematical-statistical algorithms create initial data distributions, calculate correlations, and test for significance. The AI can then make initial interpretive assessments. LLMs are used to draft preliminary versions for certain passages of the final report (e.g., descriptive summaries).
• Human Expertise (Review):
  • Checking Causality: Through its statistical analysis, the AI finds correlations that might be interpreted causally. The researchers verify the causality. No statistical correlation is included in a report without researchers confirming its methodological and logical validity.
  • Editorial Authority: Text suggestions from the AI serve as a writing aid, not as final truth. The wording, the nuance of conclusions, and the strategic recommendations for PR or media clients are authored by the human team.
  • The final report bears the signature, the intuition, and the contextual knowledge of the study authors.

Final Provision

We see ourselves as pioneers of a hybrid research methodology. Through the transparent and rule-based use of AI, we maximize the breadth and speed of our analyses. Through the mandatory, methodologically anchored human review process, we guarantee the scientific depth, validity, and social relevance of our research results.