AI does not automatically scale insight

AI does not automatically scale insight. Weak data does not become more valid through AI — it is merely packaged more convincingly.

AI can meaningfully support qualitative research. But it cannot turn unmotivated participants, superficial answers, problematic samples or biased source data into valid insight. When weak data is processed at scale, insight quality does not necessarily rise — often only the speed at which seemingly precise results are produced does.

Qualitative market research is undergoing a profound technological shift. Under labels such as “Qual at Scale”, synthetic respondents, automated interviews or AI-generated insights, methods are being offered that promise unprecedented speed and reach. Thousands of answers can be gathered, structured, clustered and translated into polished result narratives in a short time.

This progress is welcome in principle. AI can search transcripts, identify recurring patterns, organise large volumes of material and support researchers in generating hypotheses. It becomes problematic, however, when technical scalability is equated with methodological quality. More data does not automatically mean more insight — and a linguistically convincing analysis is not proof of empirical robustness.

In an interview with BILD, Prof. Dr. Dennis-Kenji Kipker points to a general risk of data-driven AI systems: as automated systems are increasingly fed with inaccurate or low-quality datasets, error rates can rise, existing expert knowledge can be lost, and problems once considered solved can reappear. The medium- and long-term consequences of short-term productivity gains, he adds, are not yet fully foreseeable.

The methodological parallel: Kipker does not explicitly discuss incentives or online panels. But his warning describes the same underlying mechanism — the quality of an AI output remains dependent on the quality of the data and the available domain expertise.

Applied to market research: if the underlying answers are superficial, unreflected, strategic or biased, an AI cannot repair these deficits after the fact. It can only process them — and possibly present them so professionally that their weaknesses become less visible.

Particularly critical is the question of where the data comes from on which new AI-based research offerings are built. Many approaches draw directly or indirectly on standardised online surveys, panel data, open text entries or historical survey material. Several quality risks are well known:

• low motivation given very low incentives,
• time pressure and clicking through as quickly as possible,
• insufficient attention when reading questions,
• brief or barely reflected open answers,
• strategic response behaviour to qualify for further studies,
• professional over-participation, multiple participation or identity issues,
• socially desirable answers and missing context.

Not every participant behaves this way, and not every quantitative study delivers weak data. Blanketly devaluing all panel participants would be scientifically and methodologically wrong. Yet the risk is real: an incentive that mainly rewards fast completion does not automatically foster intensive reflection. In open questions, the burden on participants rises, while willingness to write differentiated text answers often remains limited.

AI systems are particularly powerful at producing linguistic order. They can smooth over fragmentary answers, summarise themes and shape many individual statements into a coherent narrative. This is precisely where an epistemic risk lies: coherence can be confused with truth, linguistic precision with empirical precision, and frequency with meaning.

A qualitative interview, by contrast, lives not only from answers but from the process of gaining insight: a good moderator spots contradictions, asks for examples, tests terms, responds to uncertainty and probes unexpected statements. This interaction creates context. Where that context was not established during data collection, it can later only be estimated or simulated.

With synthetic respondents, the problem intensifies. A synthetic respondent has no personal experience, no actual needs, memories, emotions or purchase decisions. It produces a statistically or linguistically plausible answer based on existing data and model assumptions.

Synthetic audiences can therefore be useful for ideation, scenarios, early hypotheses or methodological pre-studies. They should not, however, be equated with real people who have been empirically interviewed. The plausibility of an answer says nothing about whether real people in a concrete life situation actually think, feel or act that way.

The greatest risk is not that synthetic results sound obviously wrong. The greatest risk is that they sound convincing even though their empirical basis is weak.

The term “Qual at Scale” suggests that qualitative depth can be multiplied almost at will through large case numbers and automated analysis. This carries the risk of a methodological shift in meaning. A large volume of open answers is not automatically qualitative research. Qualitative quality does not emerge from text volume alone, but from selection, relationship, conversation, contextualisation, interpretation — and the willingness to tolerate ambivalence.

AI can enable scale — for example when reviewing large volumes of material or systematically preparing human analysis. But it must not obscure the fact that quality is created at the beginning of the research chain: in target group definition, recruitment, identity verification, motivation, incentivisation, conversation and methodological control.

Anyone evaluating AI-based research solutions should therefore ask not only about speed, case numbers and visual output quality. What matters is transparent evidence of how the data was produced and how good it is:

• Where does the underlying data come from?
• How were participants recruited and identified?
• How high and how appropriate was the incentive?
• How were attention, seriousness and multiple participation controlled?
• Which statements are based on real participant data and which on model assumptions?
• Where did human methodological review take place?
• How are uncertainty, counter-examples and contradictory findings made visible?

These questions are not directed against AI. On the contrary: they create the conditions for using AI responsibly and productively. Technology delivers its greatest value where it complements high-quality research — not where it hides quality deficits.

Prof. Dr. Dennis-Kenji Kipker's warning can meaningfully be transferred to current developments in market research. His statement is not a direct critique of “Qual at Scale”, synthetic respondents or low incentives. But it provides the wider frame of reference: automation and short-term productivity gains must not diminish the importance of data quality, expertise and methodological care.

AI is first and foremost an amplifier. It can unlock good data faster, support good researchers more effectively and make high-quality insight more accessible. But it can just as well scale superficial data, reproduce biases and generate the illusion of precision.

AI does not automatically scale insight. Under poor conditions it merely scales the illusion of insight.

Framing and transparency: This feature article is based on a provided interview transcript and transfers the general warning it contains about inaccurate datasets and loss of expertise to methodological questions in market research. The specific critique of incentivisation, panel quality and synthetic respondents is the author's professional interpretation and not a verbatim statement by Prof. Dr. Dennis-Kenji Kipker.