pith. sign in

arxiv: 2606.20064 · v1 · pith:7K5CXB6Xnew · submitted 2026-06-18 · 💻 cs.HC

AI Conversational Interviewing: Scaling Up Semi-Structured and In-depth Interviews

Alexander Wuttke , Max Melchior Lang , Christopher Klamm , Quirin W\"urschinger , Frauke Kreuter This is my paper

Pith reviewed 2026-06-26 15:50 UTC · model grok-4.3

classification 💻 cs.HC
keywords AI conversational interviewingpublic opinion researchsemi-structured interviewsmigration policyopen-ended responsessurvey methodsconversational datasocial science methods
0
0 comments X

The pith

AI systems can scale semi-structured interviews to reveal reasoning and mental models that standardized surveys miss.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tests whether an AI can lead open-ended interviews with probing to collect public opinion data at large scale. In a migration policy study with 571 respondents, the conversational transcripts exposed considerations and subgroup differences in thinking that a full battery of closed-ended questions did not capture. Participants who finished rated the AI interviews as favorably as the standardized survey across modes, though completion rates varied. The work positions the method as a practical addition that combines depth with reach while releasing data and code for others to use.

Core claim

The authors claim that AI Conversational Interviewing supplies scalable open-ended data whose transcripts surface considerations and reasoning beyond the reach of standardized survey items, including markedly different mental models of migration among subgroups that hold similar attitude levels. Participant evaluations of the AI interviews reached or exceeded those of the standardized survey in voice, chat, and free-choice modes, although completion itself varied by condition.

What carries the argument

AI Conversational Interviewing, an AI-led system that conducts adaptive open-ended probing in voice, chat, or free-choice modes to gather detailed respondent reasoning on a policy topic.

If this is right

  • Researchers can gather open-ended public opinion data from hundreds of respondents without relying on trained human interviewers.
  • Analyses can distinguish reasoning patterns among subgroups that register similar scores on attitude scales.
  • The method produces respondent evaluations at or above those of standardized surveys in multiple delivery modes.
  • Open release of transcripts and pipeline materials allows direct testing on new topics and populations.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The approach could be applied to other domains where attitude measures conceal varied underlying beliefs, such as economic policy or technology adoption.
  • Automated analysis of the resulting transcripts might identify reasoning clusters more efficiently than manual coding.
  • Longer-term studies could track whether the mental models surfaced by AI interviews predict behavior better than attitude batteries alone.

Load-bearing premise

An AI system can conduct effective semi-structured probing and maintain conversational quality comparable to trained human interviewers without introducing systematic bias or limiting response depth.

What would settle it

A direct side-by-side comparison in which human interviewers elicit significantly greater response depth or less patterned bias than the AI system on the same migration questions.

Figures

Figures reproduced from arXiv: 2606.20064 by Alexander Wuttke, Christopher Klamm, Frauke Kreuter, Max Melchior Lang, Quirin W\"urschinger.

Figure 1
Figure 1. Figure 1: Comparative Strengths: Attitude Measurement Approaches populations (Conrad and Schober, 2008; Fowler and Mangione, 1998). Especially in web-based administration, standardization delivers major benefits: surveys are scalable, cost-efficient, and are, if carefully designed, comparable across respondents, countries, and time points. They also allow sample sizes large enough for subgroup analysis, experiments,… view at source ↗
Figure 2
Figure 2. Figure 2: Overall system architecture and participant flow. 7 [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Voice Interface in different states. Note: A = Entering the ID; B = Before starting the interview; C = The interviewer speaking; D = The respondent speaking. 2.6 Pre-Tests Before the main field period, we conducted multiple pre-tests with members of the research team and Prolific respondents to assess the technical viability of the study and refine the im￾plementation. These pre-tests prompted several modi… view at source ↗
Figure 4
Figure 4. Figure 4: Participant flow 3.2 Evaluation of the technical setup Technical accessibility constitutes a foundational prerequisite for successful interview comple￾tion, particularly because even respondents experienced with online survey panels may be un￾familiar with browser-based conversational interfaces — let alone voice interactions mediated by microphone and speaker input. Overall, 80.3% of respondents reported … view at source ↗
Figure 5
Figure 5. Figure 5: Evaluation of the AI conversational interviewer [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Evaluation of AI Conversational Interviewing. 3.4 Comparing evaluations of AI Conversational Interviewing with standard￾ized surveys How does this experience compare with participating in a standardized survey? Our pre￾registered success criterion specified that evaluations of the AI interview should not fall sub￾stantially below evaluations of the subsequent standardized survey. Across both text and voice… view at source ↗
Figure 7
Figure 7. Figure 7: Comparison of AI Conversational Interviewing and the standardized survey [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Overall evaluation across interview modes. 3.5 Interview duration and content Turning from self-reported evaluations to the textual data generated by the conversational interviews, [PITH_FULL_IMAGE:figures/full_fig_p014_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Analytical framework for extracting issue repertoires and argument structures from conver￾sational interviews. The classification draws on two taxonomies developed for this study: 12 topical areas (e.g., housing, labor market, humanitarian concerns) and 10 rhetorical frames (e.g., evidence: anec￾dotal; evidence: testimonial; reasoning: causal; assertion: claim), both described in detail in Appendix Section… view at source ↗
Figure 10
Figure 10. Figure 10: Issue topic distribution by party identification across all moves in the collapsed sequences (consecutive duplicates and the residual other category excluded). Each bar sums to 100%; segment labels indicate the within-party share of each issue. 3.8 Frame analysis Beyond topical content, we examine the rhetorical structure of respondents’ reasoning through argument type analysis (Soltani and Romberg, 2023;… view at source ↗
Figure 11
Figure 11. Figure 11: Argument-type distribution by party identification 4 Discussion This study asks whether large language models can conduct semi-structured interviews at the scale of standardized survey research, and what such interviews add to the public-opinion toolkit. Three findings anchor our conclusions. First, the collected textual public opinion data surface considerations, arguments, and reasoning structures that … view at source ↗
Figure 12
Figure 12. Figure 12: Perceived waiting time for voice condition. F Technical problems during the AI interview Respondents indicated, for four interview tasks, whether they experienced no, minor, or major problems. Tabulations below use the full dataset of all respondents, including respondents who could not be linked to an interview transcript or who discontinued early. Cell entries are percentages with respondent counts in p… view at source ↗
Figure 13
Figure 13. Figure 13: Average words per response by absolute turn position, separately for voice (VAPI, blue) and text (chat, red) interviews. Lines show means; shaded bands denote ±1 SD. The leftmost region (turns 1–2) corresponds to greetings and identification. The lower panel reports the number of interviews contributing to each turn position. Text responses peak early and decline steeply, while voice responses remain stab… view at source ↗
Figure 14
Figure 14. Figure 14: Attitutdes towards immigration I Text analysis: Taxonomy and validation This appendix provides details on the taxonomy of topics and argument types and on the validation procedures and results for coding the topics and argument types. The issue taxonomy comprises 12 substantive categories organized into seven thematic groups, as reported in [PITH_FULL_IMAGE:figures/full_fig_p071_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Inter-Annotator Agreement J Text analysis separated by voter groups This appendix documents the text analysis results for all voter groups and not just the selected groups of voters that are discussed in the main text [PITH_FULL_IMAGE:figures/full_fig_p074_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Issue topic distribution by party identification across all moves in the collapsed sequences (consecutive duplicates and the residual other category excluded). Each bar sums to 100%; segment labels indicate the within-party share of each issue. 74 [PITH_FULL_IMAGE:figures/full_fig_p074_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Argument-type distribution by party identification across all moves in the collapsed se￾quences. Each bar sums to 100%; segment labels indicate the within-party share of each argument type. K Sequence analysis This appendix reports an additional analysis on the collected interview data that examines the sequences of topics respondents raised during the interview, separated by party affiliation. 4 people d… view at source ↗
Figure 18
Figure 18. Figure 18: Issue topic trees (depth 2) for the two rhetorically most distinctive parties. Node radius is proportional to the share of sequences passing through the node at that position (entry share at the root, joint entry × transition share at depth 1); edge labels report the transition probability. Vertical dots below a node signal that the topic chain continues at deeper levels (branches with transition probabil… view at source ↗
Figure 19
Figure 19. Figure 19: Most likely issue sequence per party, obtained by a greedy Markov walk through the topic-transition matrix (no self-loops; previously visited topics are excluded). The percentage under the first node is the share of sequences starting with that topic; subsequent percentages are the transition probability from the previous node. K.2 Framing analysis The analysis of argument-type sequences uncovers distinct… view at source ↗
Figure 20
Figure 20. Figure 20: Argument-type strategy trees (depth 2) for Die Linke. Node radius is proportional to the share of sequences at that position (entry share at the root; joint entry × transition share at depth 1). Vertical dots below a node signal that the chain continues at deeper levels (branches with transition probability ≥ 10%, consecutive duplicates collapsed). To make these rhetorical signatures concrete at the respo… view at source ↗
Figure 21
Figure 21. Figure 21: Most likely argument-type sequence per party, obtained by a greedy Markov walk through the argument-type transition matrix (no self-loops; previously visited moves are excluded). The percent￾age under the first node is the share of sequences starting with that move; subsequent percentages are the transition probability from the previous node. 79 [PITH_FULL_IMAGE:figures/full_fig_p079_21.png] view at source ↗
read the original abstract

Public opinion research has long faced a trade-off between depth and scale: standardized surveys enable large-scale measurement but restrict respondents to researcher-defined categories, obscuring the diversity of unexpected considerations that underlie public sentiment. More conversational interviews provide richer insights through open-ended probing, but their reliance on trained human interviewers has kept them difficult to scale. This study introduces AI Conversational Interviewing as a method for collecting open-ended public opinion data at scale, pursuing three objectives: to demonstrate the analytical value of conversational text data for questions beyond the reach of closed-ended items; to assess the method's practical viability through participants' own evaluations; and to inform implementation by experimentally comparing voice-based, chat-based, and free-choice interview modes. We conducted a study combining an AI-led interview with a standardized survey on migration policy among 571 respondents recruited via Prolific and Payback Panel. The findings establish AI Conversational Interviewing as a viable and valuable addition to the social-science toolkit. The conversational transcripts surface considerations and reasoning that a comprehensive standardized battery does not capture such as markedly different mental models of migration among subgroups with similar attitudes levels. Among respondents who completed the interview, evaluations of the AI interview were at or above those of the standardized survey across modes, although completion itself varied by condition. By releasing open data and open-source pipeline materials, the study contributes to a growing literature on harnessing artificial intelligence to expand the methods of public opinion measurement.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 1 minor

Summary. The manuscript introduces AI Conversational Interviewing as a scalable method for semi-structured public-opinion interviews. It reports results from a 571-participant experiment on migration policy that pairs an AI-led interview (voice, chat, or free-choice modes) with a standardized survey battery; the central findings are that the conversational transcripts reveal considerations and mental models not captured by closed items, that completers rate the AI interview at or above the survey on participant evaluations, and that completion rates vary by mode. The study releases open data and pipeline code.

Significance. If the method can be shown to produce human-comparable probing depth and thematic coverage without systematic bias, it would meaningfully expand the toolkit for combining scale with open-ended insight in opinion research. The planned open release of data and code is a concrete strength that supports reproducibility and further testing.

major comments (3)
  1. [Abstract, objectives and findings sections] Abstract, objectives and findings sections: the claim that AI Conversational Interviewing is a 'viable and valuable addition to the social-science toolkit' and scales semi-structured interviews rests on an implicit assumption of comparability to trained human interviewers, yet the experiment contains no human-interviewer arm; the only comparison is to a standardized survey, leaving open whether observed differences in surfaced considerations arise from conversational format, AI artifacts, or both.
  2. [Findings section] Findings section (as described in abstract): no quantitative metrics of probe frequency, response elaboration, or inter-rater coded thematic diversity are reported to ground the assertion that transcripts 'surface considerations and reasoning that a comprehensive standardized battery does not capture'; without these, the analytical-value claim cannot be evaluated against the survey baseline.
  3. [Methods] Methods (implied by soundness note): the manuscript provides no details on the AI interview protocol, transcript coding scheme, statistical tests for mode differences, or handling of non-completers; these omissions directly affect assessment of the viability and mode-comparison objectives.
minor comments (1)
  1. [Abstract] Clarify in the abstract whether the AI system uses a fixed prompt template or dynamic probing logic, as this bears on replicability.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive comments, which help clarify the positioning and presentation of our work. We respond to each major comment below.

read point-by-point responses

  1. Referee: Abstract, objectives and findings sections: the claim that AI Conversational Interviewing is a 'viable and valuable addition to the social-science toolkit' and scales semi-structured interviews rests on an implicit assumption of comparability to trained human interviewers, yet the experiment contains no human-interviewer arm; the only comparison is to a standardized survey, leaving open whether observed differences in surfaced considerations arise from conversational format, AI artifacts, or both.

    Authors: We agree that the absence of a human-interviewer comparison arm means we cannot claim direct comparability to trained human interviewers. The study demonstrates value relative to standardized surveys by showing additional considerations in the conversational data and comparable or better participant evaluations. We will revise the relevant sections to remove any implication of human comparability and explicitly frame the contribution as scaling open-ended interviewing beyond surveys, while noting the lack of human benchmark as a limitation for future research. revision: yes

  2. Referee: Findings section (as described in abstract): no quantitative metrics of probe frequency, response elaboration, or inter-rater coded thematic diversity are reported to ground the assertion that transcripts 'surface considerations and reasoning that a comprehensive standardized battery does not capture'; without these, the analytical-value claim cannot be evaluated against the survey baseline.

    Authors: The manuscript currently supports the claim with qualitative illustrations of distinct mental models. To address this, we will add quantitative metrics in the revision, including the count of unique considerations identified through thematic coding of transcripts compared to survey responses, and report inter-rater agreement for the coding process. This will provide a more robust evaluation of the analytical value. revision: yes

  3. Referee: Methods (implied by soundness note): the manuscript provides no details on the AI interview protocol, transcript coding scheme, statistical tests for mode differences, or handling of non-completers; these omissions directly affect assessment of the viability and mode-comparison objectives.

    Authors: We will revise the Methods section to include fuller details on the AI interview protocol, the coding scheme for transcripts, the statistical tests applied to mode differences, and the approach to non-completers. The open release of data and code will also aid in assessing these aspects. revision: yes

Circularity Check

0 steps flagged

No circularity: purely empirical study with no derivations or self-referential predictions

full rationale

The paper reports an empirical experiment (n=571) contrasting AI-led interviews against a standardized survey on migration policy, evaluating participant satisfaction, completion rates, and qualitative transcript content. No equations, fitted parameters, model predictions, or derivation chains exist. Claims rest on direct data collection and open-ended analysis rather than any reduction of outputs to inputs by construction. No self-citations are invoked to justify load-bearing methodological uniqueness or uniqueness theorems. The absence of a human-interviewer arm is a methodological limitation but does not constitute circularity under the defined criteria.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The paper is an empirical methods contribution; it rests on domain assumptions about AI interviewing capability and respondent behavior rather than mathematical derivations or new entities.

axioms (1)
  • domain assumption AI systems can conduct effective semi-structured interviews that elicit open-ended responses comparable in quality to those from trained human interviewers
    This premise underpins the viability and analytical-value claims in the abstract.

pith-pipeline@v0.9.1-grok · 5803 in / 1265 out tokens · 26237 ms · 2026-06-26T15:50:27.110852+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

32 extracted references · 3 canonical work pages

  1. [1]

    Can Generative AI Improve Social Science?

    issn: 1047-1987, 1476-4989. doi: 10.1093/pan/mpt049. url: https://www.cambridge. org/core/product/identifier/S1047198700013942/type/journal_article (visited on 07/30/2024). Bail, Christopher A. (2022). “Can Generative AI Improve Social Science?” In: Big Data & Society 9.1, pp. 1–7. doi: 10.1177/20539517221080146. url: https://journals.sagepub. com/doi/ful...

    work page doi:10.1093/pan/mpt049 1987

  2. [2]

    Does prompt formatting have any impact on llm performance?

    doi: 10.48550/ARXIV.2411.10541. url: https://arxiv.org/abs/2411.10541 (visited on 02/27/2026). Helfferich, Cornelia (2011). Die Qualität qualitativer Daten: Manual für die Durchführung quali- tativer Interviews. ger. 4. Auflage. SpringerLink Bücher. Wiesbaden: VS Verlag für Sozialwis- senschaften. isbn: 978-3-531-17382-5 978-3-531-92076-4. doi: 10.1007/97...

    work page doi:10.48550/arxiv.2411.10541 2026

  3. [3]

    Why open-source generative AI models are an ethical way forward for science

    doi: 10.18653/v1/2023.argmining-1.15 . url: https://aclanthology.org/2023. argmining-1.15/. Spirling, Arthur (2023). “Why open-source generative AI models are an ethical way forward for science” . In: Nature 616.7957, pp. 413–413. Spradley, James P. (2016). The ethnographic interview . eng. Long Grove, Illinois: Waveland Press, Inc. isbn: 978-1-4786-3207-...

    work page doi:10.18653/v1/2023.argmining-1.15 2023

  4. [4]

    Ich bestätige, dass ich die Studieninformationen gelesen habe und Gelegenheit hatte, die Informationen zu prüfen

  5. [5]

    Ich stimme der Erhebung und Speicherung meiner Daten in anonymisierter Form zu, die zu Replikationszwecken der Forschungsgemeinschaft zur Verfügung gestellt werden können

  6. [6]

    Ich stimme der Teilnahme an der Studie zu

  7. [7]

    Ich bestätige, dass ich mindestens 18 Jahre alt bin

  8. [8]

    Ich stimme zu, dass Daten auch außerhalb der EU verarbeitet und gespeichert werden können

    Ich erkläre mich einverstanden, dass im Rahmen dieser Studie Daten von V API, Literal.AI, Qualtrics und OpenAI verarbeitet und gespeichert werden. Ich stimme zu, dass Daten auch außerhalb der EU verarbeitet und gespeichert werden können. Q1: Datenschutzinformationen. Die Survey-Seite enthielt ausführliche Datenschutzinfor- mationen gemäß Art. 13 DSGVO. In...

  9. [9]

    Gespräch via Text-Chat,

  10. [10]

    Beim Text-Chat werden Sie Ihre Antworten über die Tastatur eingeben

    Gespräch via Sprachein- und Sprachausgabe. Beim Text-Chat werden Sie Ihre Antworten über die Tastatur eingeben. Das Gespräch via Sprachein- und Sprachausgabe nutzt Mikrofon sowie Lautsprecher/Kopfhörer. Je nach Auswahl wurde anschließend die entsprechende Text- oder Voice-Oberfläche mit demselben Instruktionstext wie in den direkt zugewiesenen Bedingungen...

  11. [11]

    dem Gesprächspartner Antworten geben,

  12. [12]

    problems_txt

    vom Gesprächspartner Fragen erhalten. problems_txt. Nur für Befragte mit kleinen oder großen Problemen: Welche Probleme sind während Ihres KI-gestützten Gesprächs aufgetreten? Nennen Sie bitte alle Probleme. Q91. Hat sich der KI-Interviewer während des Gesprächs in irgendeiner Weise anstößig, ver- letzend oder in anderer Weise grob unangemessen geäußert?

  13. [13]

    Falls “Ja” bei Q91: Hinweis auf die Möglichkeit der Kontaktaufnahme mit Prof

    Ja, und zwar folgende Äußerung: Q94. Falls “Ja” bei Q91: Hinweis auf die Möglichkeit der Kontaktaufnahme mit Prof. Dr. Alexander Wuttke. ai-accuracy . Manchmal traut man sich in einem Interview nicht frei zu sagen, was man eigentlich denkt. Inwiefern reflektieren Ihre Äußerungen gegenüber dem KI-Gesprächspartner Ihre ehrlichen Gedanken, Gefühle und Erfahr...

  14. [14]

    Sollte häufiger eingesetzt werden

  15. [15]

    Im Folgenden möchten wir noch einmal mehr über Ihre politischen Ansichten erfahren

    Sollte nicht häufiger eingesetzt werden B.8 Block: GLES Q45. Im Folgenden möchten wir noch einmal mehr über Ihre politischen Ansichten erfahren. q42. Parteipositionen zum Zuzug von Ausländern. Matrix für CDU/CSU, SPD, Bündnis 90/Die Grünen, Die Linke und AfD auf einer 11-Punkte-Skala von Zuzug von Ausländern erleichtern bis Zuzug von Ausländern einschränk...

  16. [16]

    Minderheiten sollten sich den deutschen Gepflogenheiten anpassen

  17. [17]

    Der Wille der Mehrheit sollte immer Vorrang haben, auch wenn Minderheitenrechte betrof- fen sind

  18. [18]

    Einwanderinnen/Einwanderer sind im Allgemeinen gut für die deutsche Wirtschaft

  19. [19]

    Die deutsche Kultur ist durch Einwanderinnen/Einwanderer bedroht

  20. [20]

    wirklich deutsch

    Einwanderinnen/Einwanderer erhöhen die Kriminalitätsrate in Deutschland. q75. Parteiidentifikation: SPD, CDU/CSU, CDU, CSU, Grüne, FDP, AfD, Die Linke, BSW, andere Partei (offene Nennung) oder keiner Partei . 38 q126. Wichtigkeit verschiedener Kriterien, um “wirklich deutsch” zu sein:

  21. [21]

    in Deutschland geboren sein,

  22. [22]

    deutsche Vorfahren haben,

  23. [23]

    deutsch sprechen können,

  24. [24]

    Vierstufige Skala von überhaupt nicht wichtig bis sehr wichtig

    sich an deutsche Traditionen und Gepflogenheiten halten. Vierstufige Skala von überhaupt nicht wichtig bis sehr wichtig . q27a. Einwanderinnen/Einwanderer sollten verpflichtet werden, sich der deutschen Kultur anzupassen. Fünfstufige Zustimmungsskala. q166. Matrix zu autoritären, nationalistischen und fremdenfeindlichen Aussagen mit fünfstu- figer Zustimm...

  25. [25]

    Was Deutschland jetzt braucht, ist eine einzige starke Partei, die die Volksgemeinschaft insgesamt verkörpert

  26. [26]

    Wir sollten endlich wieder Mut zu einem starken Nationalgefühl haben

  27. [27]

    Die Bundesrepublik ist durch die vielen Ausländer in einem gefährlichen Maß überfremdet

  28. [28]

    Auch heute noch ist der Einfluss von Jüdinnen/Juden zu groß

  29. [29]

    Wie in der Natur sollte sich in der Gesellschaft immer der Stärkere durchsetzen

  30. [30]

    B.9 Block: End comparison

    Die Verbrechen des Nationalsozialismus sind in der Geschichtsschreibung weit übertrieben worden. B.9 Block: End comparison. Vergleich zwischen KI-Gespräch und standardisiertem Fragebogen auf fünf Di- mensionen mit den Antwortoptionen KI-gestütztes Gespräch , Standardisierter Fragebogen, Kei- ne Präferenz :

  31. [31]

    ich habe mehr meine individuelle Meinung ausdrücken können,

  32. [32]

    Wir führen ein Gespräch über Politik und Gesellschaft, mit besonderem Fokus auf Migrationsthemen

    bevorzuge ich für zukünftige Interviews. survey_evaluation. Gesamtbewertung des klassischen Fragebogens auf einer 7-Punkte-Skala von sehr schlecht bis sehr gut . survey-accuracy . Inwiefern reflektieren die Äußerungen im klassischen Fragebogen die ehrli- chen Gedanken, Gefühle und Erfahrungen? Skala von 1 = völlig bis 8 = überhaupt nicht . device. Gerätet...

    2015