The Value of Gen-AI Conversations: A bottom-up Framework for AI Value Alignment

Alexander Kagan; Katharina Baum; Lenart Motnikar; Sarah Spiekermann-Hoff

REVIEW 4 major objections 4 minor 5 references

The Value of Gen-AI Conversations: A bottom-up Framework for AI Value Alignment

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Value alignment for AI chatbots can be discovered from real user conversations, not just imposed top-down: an analysis of 16,908 employment-service logs yields nine core values and 32 concrete misalignment types.

desk verdict Useful empirical taxonomy of CA misalignments, but the 'bottom-up' discovery claim is partly an artifact of seeding the codebook with the authors' own VBE ontology. read the letter →

arxiv 2507.21091 v1 pith:BROKV7PC submitted 2025-06-26 cs.CY cs.AI

Lenart Motnikar , Katharina Baum , Alexander Kagan , Sarah Spiekermann-Hoff This is my paper

classification cs.CYcs.AI

keywords conversationalagentsgenerativeAIvaluealignmentethicalValue-BasedEngineeringqualitativeanalysiscareercounsellinglargelanguagemodels

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper argues that value alignment for conversational agents can be discovered from the bottom up, by reading what actually goes wrong in real user conversations, rather than imposed from abstract ethical guidelines. Analyzing 16,908 sessions with a European employment-service chatbot, the authors narrow to 187 ethically sensitive conversations containing 593 system outputs and code every problematic response. They arrive at nine core values—truthfulness, helpfulness, constructiveness, courtesy, prudence, sensitivity, attentivity, coherence, and clarity—and 32 concrete types of misalignment, each with a description of how it harmed the user. If the approach holds, it gives AI providers a replicable way to turn vague ethical complaints like “biased” into specific, fixable system failures.

What carries the argument

The central mechanism is a two-part analytical apparatus. The first part is the value ontology of Value-Based Engineering, the ISO/IEC/IEEE 24748-7000 standard, which treats IT systems as value bearers and defines a value misalignment as a characteristic of an output that undermines a core value, called a “negative value quality” in the standard's terms. The second is the collaborative qualitative analysis procedure, in which three coders plus a lead researcher assess each of 593 system outputs against a codebook seeded by five researchers familiar with the standard, iteratively refining definitions and resolving disagreements in a consensus workshop. The codebook of nine core values and their associated misalignment types is the load-bearing artifact that connects raw conversational data to a reusable typology.

What would settle it

An independent team that re-codes the same 593 system outputs without seeing the paper's coding scheme would falsify the central claim if it does not independently arrive at the same nine core values; so would a check showing that the nine values disappear when the full 16,908-session log is screened instead of only the 187 ethically sensitive subset.

Watch

Extended reading notes

Core claim

The paper's central claim is that a bottom-up, context-sensitive method, built on the value ontology of Value-Based Engineering and the discipline of collaborative qualitative analysis, can surface the values and value misalignments that actually matter in a deployed generative-AI conversational agent. Applied to a career-counselling CA, the method yields nine core values central to the interaction and 32 distinct value misalignments that negatively affected users, with frequencies ranging from attentivity failures in a quarter of coded outputs to clarity failures in about six percent. The authors argue this reframes broad public criticism—such as “the chatbot is biased”—into multi-faceted design problems spanning prudence, helpfulness, courtesy, and coherence, and that the resulting typology is actionable for providers and regulators.

Load-bearing premise

The findings stand or fall on whether the nine core values genuinely emerge from the conversations themselves rather than being imposed by the pre-built coding scheme the researchers used, and on whether the 187 ethically sensitive sessions fairly represent the full set of 16,908 conversations.

Editorial extensions

If this is right

CA providers can quantify which value misalignments are most frequent in their own logs and prioritize fixes accordingly, as demonstrated by attentivity failing in 25.80% of coded cases versus clarity in 6.07%.
Broad ethical criticisms such as “gender bias” can be decomposed into specific misalignments across prudence, helpfulness, courtesy, and coherence, each requiring different technical remedies.
Because different misalignments stem from different architectural components—the LLM, retrieval databases, prompt design, context tracking—value alignment must be addressed system-wide rather than at the model layer alone.
The method is generalizable to other CA deployments and to emerging regulation such as the EU AI Act, offering a replicable route from real-world logs to compliance-relevant value evidence.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The 32 misalignment types could be turned into machine-readable evaluation criteria: a classifier trained on the 593 coded outputs could automatically flag misalignments in future logs, making the bottom-up approach scalable beyond manual coding.
Applying the same codebook to a different CA domain, such as health or education, would test whether the nine values are truly universal or an artifact of the career-counselling context; the authors expect broad relevance but do not test this.
The finding that attentivity failures are most frequent suggests that context-tracking and memory mechanisms, not just model scale, may be the next binding constraint on perceived chatbot quality; this is an inference beyond the paper's own data.
A direct comparison against top-down methods such as Constitutional AI on the same logs could quantify how much bottom-up discovery adds beyond predefined principles; the paper argues for the approach but does not run this comparison.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Two things to know. First, this is a useful empirical case study: the authors took 16,908 real conversations with a public employment bot, zeroed in on 187 ethically sensitive sessions, and produced a concrete taxonomy of nine values and 32 misalignment types with vivid examples. That kind of grounded, context-specific mapping is genuinely rare in the value-alignment literature, and practitioners could use it directly. Second, the paper's headline claim—that these values were 'revealed' bottom-up from the data—is overstated. The codebook was seeded by five researchers familiar with the IEEE7000 / Value-Based Engineering ontology, which belongs to one of the co-authors, and that pre-structured list already contained the same nine values that later 'emerged.' The authors did allow for new categories and an 'other' bucket, but they never report how many new values or misalignments came from the data itself. So the circularity concern is real, though not fatal: the analysis applies a lens rather than discovering from scratch. What the paper does well: the coding process is transparent, the examples are apt, and the synthesis step shows careful consensus work. The discussion of gender bias—breaking it into issues of prudence, helpfulness, courtesy, and coherence—is a nice move. The authors are honest about limitations, including the GPT-3.5 era and the specific use case. The soft spots: besides the emergence/circularity issue, the sampling is narrow. The 187 sessions come from a filtered subset of 1,863, and the filtering criteria (no pre-generated prompts, no restarts, more than one turn, dominant language) plausibly skew the conversation types. There is no information about how representative these are of the full 16,908 sessions. Also, the number of outputs shifts from 593 in the abstract to 585 in the synthesis section; that inconsistency should have been caught. The lack of inter-rater reliability is defensible given the CQA framework, but it does mean the prevalence percentages should be read as consensus-coded impressions rather than measurements. Overall, this is a solid, honest case study. It deserves peer review, but a referee should ask the authors to explicitly address the seeding-emergence question, report the proportion of categories genuinely added during analysis, clarify the sampling rationale, and fix the numeric inconsistency. The central empirical content stands; the framing needs adjustment. I'd bring it to a reading group on AI ethics but wouldn't cite it as a definitive framework—more as an example of how to do context-sensitive evaluation.

Referee Report

4 major / 4 minor

Summary. The paper analyzes 16,908 real-world conversations with a European public-employment GenAI career-counselling chatbot. After filtering to 1,863 relevant sessions and screening 187 ethically sensitive sessions containing 593 system outputs, the authors apply collaborative qualitative analysis with a codebook built on the Value-Based Engineering (VBE) ontology to identify nine core values and 32 types of value misalignment. They report prevalence frequencies, illustrate each misalignment with anonymized examples, and argue that this bottom-up, context-sensitive method improves on top-down alignment approaches. The paper concludes with practical recommendations for CA developers and claims generalizability to other conversational agent deployments.

Significance. The study's strengths are its large real-world dataset, the detailed and transparent qualitative procedure, and the concreteness of the derived value-misalignment typology. The examples are vivid and the mapping to adverse effects is useful for practitioners. If the central claim holds, the paper provides an actionable framework that reframes abstract ethical concerns into concrete failure modes that can be prioritized and fixed. The method is portable to other CA domains. However, the significance is tempered by the circularity risk described in the major comments: the nine core values are not shown to have emerged independently from the ethically sensitive conversations, and the reported prevalence figures contain an unresolved numerical inconsistency.

major comments (4)

[§3.2.2 and §3.3.3–3.3.4] The 'bottom-up' claim is not fully supported by the reported procedure. The initial codebook was created by five researchers 'familiar with the IEEE7000 standard' who analysed randomly sampled conversations and clustered codes into 'nine core values' before the main coding of the 187 ethically sensitive sessions. The later synthesis step allowed new categories, but the paper does not report how many, if any, new core values or misalignment types actually emerged from the 'other' category. As written, the final nine values are as plausibly an application of the VBE ontology as a discovery from the conversations. Please provide an audit trail of codebook revisions, quantify any categories that arose solely from the ethically sensitive subset, and soften the 'revealed' wording if no such categories emerged.
[§3.3.4 and Figure 2] The prevalence figures are internally inconsistent. Section 3.3.3 reports 1,034 critical observations; Section 3.3.4 says the consensus process reduced the list to 585 problematic outputs; yet Figure 2 and the text in Section 4.2 describe percentages 'totalling 593 outputs.' The reader cannot determine whether the denominator for the 6.07% and 25.80% figures is the original 593 system outputs, the 585 consensus cases, or the 1,034 observations. Please reconcile these numbers and explicitly state the denominator for every percentage.
[§3.3.1] The filtering criteria (excluding pre-generated prompts, restarts, single-prompt sessions, and non-dominant languages) are plausible but are not validated against the possibility of systematic bias. The 187 ethically sensitive sessions are the sole empirical basis for the prevalence claims and the proposed generalizations. The paper should provide a sensitivity check, for example by reporting how many conversations containing health, disability, migrant-status, or financial-distress disclosures were excluded by these filters, or by comparing value-misalignment patterns in a random sample of the unfiltered data.
[Abstract and §3.3.3] The abstract states that the identified misalignments 'negatively impacted users,' but the adverse-effect descriptions were inferred by the coders from the CA's outputs, not reported by the users themselves. This is an interpretive step that should be acknowledged in the wording: e.g., 'coder-assessed adverse effects' or 'potentially negative impacts.' As written, the claim overstates the evidence, particularly because no user feedback or follow-up data are available.

minor comments (4)

[Figure 2] The submitted version of Figure 2 is garbled, with overlapping boxed labels that make the value names unreadable. Please provide a clean, high-resolution figure.
[References] Several reference entries contain corrupted text, such as 'Ling, . C.' and 'Weidinger, . , Mellor, J.' Please proofread the bibliography against the original sources.
[§3.3.3 and §4.2] The paper uses 'cases,' 'outputs,' and 'observations' interchangeably. Define the unit of analysis precisely and use consistent terminology throughout.
[§5.1] The mapping of the nine values to the '3H' criterion is interesting but underdeveloped; a small table or explicit sentence listing which values correspond to helpfulness, honesty, and harmlessness would strengthen the theoretical contribution.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the value codebook was data-seeded and the reported misalignments are empirically grounded

full rationale

The potentially circular step is the use of the Value-Based Engineering (VBE) ontology to build a nine-value codebook, followed by reporting those same nine values as the study's findings. However, the paper states that the initial codebook was constructed by five researchers 'familiar with the IEEE7000 standard' who analysed randomly sampled conversations, identified suboptimal outputs, and clustered the resulting codes into nine values. The nine values are therefore an intermediate data reduction, not a fixed input imported from the cited framework. The coding procedure explicitly allowed an 'other' category, and the synthesis step allowed creating, removing, merging, and splitting categories, with 'other' cases assigned to 'existing or new categories'. Thus the final taxonomy was not forced by construction. The 32 misalignment types and their adverse-effect descriptions were derived from the 593 coded outputs and are not fitted parameters or predictions from the ontology. The self-citations to Spiekermann (2023) and related VBE work are references to an external IEEE standard and prior applications; they do not function as an unverified uniqueness claim or as a substitute for the empirical coding. The paper's qualitative derivation is self-contained, and the usual limitation that any coding uses an analytic lens is a methodological caveat, not circularity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The paper does not introduce new physical entities or forces. Its central empirical output, the nine-value and 32-misalignment typology, is a conceptual classification derived from a preselected ontology and a narrow, privacy-restricted dataset. The ledger highlights that the main free choices are the sampling thresholds and the VBE-seeded codebook, both of which materially shape the results.

free parameters (3)

Conversation filtering thresholds = no pre-generated prompts; no restart; more than one prompt; dominant language
Chosen by hand in Section 3.3.1 to 'capture the conversations most related to our research'; this selection shapes the 1,863-session subset and therefore the derived values.
Ethically sensitive disclosure categories = health, disability, migrant status, financial distress, pregnancy, minor or elderly age
Defined in Section 3.3.2 to identify the 187 sessions; the core values are inferred only from sessions flagged under these categories, which may not cover all sensitive cases.
Nine-value codebook = attentivity, helpfulness, coherence, constructiveness, courtesy, prudence, sensitivity, truthfulness, clarity
Initial codebook constructed by five researchers using the VBE ontology in Section 3.2.2; the 'discovery' of these nine values as core is shaped by this preselected category set.

assumptions (4)

domain assumption The Value-Based Engineering value ontology (IEEE/ISO 24748-7000) is a valid and complete basis for identifying values in CA conversations.
Used to create the codebook in Section 3.2.2; if the ontology omits values relevant to users, the nine-value result is incomplete.
ad hoc to paper The 187 ethically sensitive sessions are representative of the core values and misalignments in all 16,908 sessions.
Section 3.3.2 narrows from 1,863 filtered to 187 sessions; the paper generalizes to the whole context from this subset without a representativeness check.
domain assumption Qualitative consensus coding without inter-rater reliability measures yields valid value-misalignment labels.
Section 3.2.1 adopts CQA, which rejects reliability measures; this underpins the claim that the 32 misalignment types are real rather than coder artifacts.
domain assumption Career-counselling CA interactions are an ethically sensitive context involving vulnerable users.
The introduction and Section 3.3.2 justify the focus on this context; if the context were not sensitive, the derived values might differ.

how reviews work

0 comments

Cite this review

Pith. "Pith review of The Value of Gen-AI Conversations: A bottom-up Framework for AI Value Alignment." pith.science (2026). https://pith.science/paper/BROKV7PC

@misc{pith2026250721091,
  author       = {Pith},
  title        = {Pith review of: The Value of Gen-AI Conversations: A bottom-up Framework for AI Value Alignment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BROKV7PC}},
  note         = {Machine review of arXiv:2507.21091}
}

read the original abstract

Conversational agents (CAs) based on generative artificial intelligence frequently face challenges ensuring ethical interactions that align with human values. Current value alignment efforts largely rely on top-down approaches, such as technical guidelines or legal value principles. However, these methods tend to be disconnected from the specific contexts in which CAs operate, potentially leading to misalignment with users interests. To address this challenge, we propose a novel, bottom-up approach to value alignment, utilizing the value ontology of the ISO Value-Based Engineering standard for ethical IT design. We analyse 593 ethically sensitive system outputs identified from 16,908 conversational logs of a major European employment service CA to identify core values and instances of value misalignment within real-world interactions. The results revealed nine core values and 32 different value misalignments that negatively impacted users. Our findings provide actionable insights for CA providers seeking to address ethical challenges and achieve more context-sensitive value alignment.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

5 extracted references · 2 canonical work pages

[1]

Alberts, L., Keeling, G., & McCroskery, A. (2024). Should agentic conversational AI change how we think about ethics? Characterising an interactional ethics centred on respect (Version 2). arXiv. https://doi.org/10.48550/ARXIV.2401.09082 Allouch, M., Azaria, A., & Azoulay, R. (2021). Conversational Agents: Goals, Technologies, Vision and Challenges. Senso...

arXiv doi:10.48550/arxiv.2401.09082 2024
[18]

https://aisel.aisnet.org/ecis2022_rp/18 Kluckhohn, C. (1951). Values and Value -Orientations in the theory of action: An exploration in definition and classification. In T. Parsons, E. Shils, & N. Smelser (Eds.), Toward a general theory of action (pp. 388–433). Harvard University Press. Kretzschmar, K., Tyroll, H., Pavarini, G., Manzini, A., Singh, I., & ...

work page 1951
[27]

https://doi.org/10.1007/s13347-023-00606-x Kenton, Z., Everitt, T., Weidinger, L., Gabriel, I., Mikulik, V., & Irving, G. (2021). Alignment of Language Agents (Version 1). arXiv. https://doi.org/10.48550/ARXIV.2103.14659 Khandelwal, K., Tonneau, M., Bean, A. M., Kirk, H. R., & Hale, S. A. (2024). Indian-BhED: A Dataset for Measuring India -Centric Biases ...

arXiv doi:10.48550/arxiv.2103.14659 2021
[236]

It is currently hodgepodge

https://doi.org/10.1038/s41746-023-00979-5 ing, . C., Tussyadiah, I., Tuomi, A., Stienmetz, J., & Ioannou, A. (202 ). Factors influencing users’ adoption and use of conversational agents: A systemati c review. Psychology & Marketing , 38(7), 1031–1051. https://doi.org/10.1002/mar.21491 Malgieri, G., & Niklas, J. (2020). Vulnerable data subjects. Computer ...

work page arXiv 2020
[8448]

https://doi.org/10.3390/s21248448 Associated Press. (2024). An AI chatbot pushed a teen to kill himself, a lawsuit against its creator alleges. Associated Press. https://apnews.com/article/9d48adc572100822fdbc3c90d1456bd0 Attard-Frost, B., De Los Ríos, A., & Walters, D. R. (2023). The ethics of AI business practices: A review of 47 AI ethics guidelines. A...

arXiv doi:10.48550/arxiv.2212.08073 2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Alberts, L., Keeling, G., & McCroskery, A. (2024). Should agentic conversational AI change how we think about ethics? Characterising an interactional ethics centred on respect (Version 2). arXiv. https://doi.org/10.48550/ARXIV.2401.09082 Allouch, M., Azaria, A., & Azoulay, R. (2021). Conversational Agents: Goals, Technologies, Vision and Challenges. Senso...

arXiv doi:10.48550/arxiv.2401.09082 2024

[2] [18]

https://aisel.aisnet.org/ecis2022_rp/18 Kluckhohn, C. (1951). Values and Value -Orientations in the theory of action: An exploration in definition and classification. In T. Parsons, E. Shils, & N. Smelser (Eds.), Toward a general theory of action (pp. 388–433). Harvard University Press. Kretzschmar, K., Tyroll, H., Pavarini, G., Manzini, A., Singh, I., & ...

work page 1951

[3] [27]

https://doi.org/10.1007/s13347-023-00606-x Kenton, Z., Everitt, T., Weidinger, L., Gabriel, I., Mikulik, V., & Irving, G. (2021). Alignment of Language Agents (Version 1). arXiv. https://doi.org/10.48550/ARXIV.2103.14659 Khandelwal, K., Tonneau, M., Bean, A. M., Kirk, H. R., & Hale, S. A. (2024). Indian-BhED: A Dataset for Measuring India -Centric Biases ...

arXiv doi:10.48550/arxiv.2103.14659 2021

[4] [236]

It is currently hodgepodge

https://doi.org/10.1038/s41746-023-00979-5 ing, . C., Tussyadiah, I., Tuomi, A., Stienmetz, J., & Ioannou, A. (202 ). Factors influencing users’ adoption and use of conversational agents: A systemati c review. Psychology & Marketing , 38(7), 1031–1051. https://doi.org/10.1002/mar.21491 Malgieri, G., & Niklas, J. (2020). Vulnerable data subjects. Computer ...

work page arXiv 2020

[5] [8448]

https://doi.org/10.3390/s21248448 Associated Press. (2024). An AI chatbot pushed a teen to kill himself, a lawsuit against its creator alleges. Associated Press. https://apnews.com/article/9d48adc572100822fdbc3c90d1456bd0 Attard-Frost, B., De Los Ríos, A., & Walters, D. R. (2023). The ethics of AI business practices: A review of 47 AI ethics guidelines. A...

arXiv doi:10.48550/arxiv.2212.08073 2024