AI and My Values: User Perceptions of LLMs’ Ability to Extract, Embody, and Explain Human Values from Casual Conversations
This is a strong CHI paper because it turns a philosophical question into a concrete HCI measurement problem: what users come to believe about an LLM’s values after sustained interaction. The main contribution is VAPT, which is methodologically useful, while the safety framing around “weaponized empathy” gives the work clear stakes beyond a single study.
Axes Lens
Rare contribution shape, typical evidence profile. The point here is not a score. It is to show what kind of claim the paper makes, and whether the evidence pattern is unusual or baseline in this 268 -review set.
Contribution shape
- Knowledge form
- method knowledge typical · 29/268
- Novelty type
- tool typical · 14/268
- Abstraction level
- practice typical · 85/268
- Generalization target
- methodological argument typical · 16/268
- Validation mode
- qualitative study typical · 63/268
Evidence profile
- Evidence strength
- strong typical · 158/268
- Claim alignment
- strong typical · 231/268
- Overclaim risk
- medium typical · 210/268
Review Summary
This paper’s strongest contribution is methodological and interpretive rather than algorithmic. Instead of claiming that an LLM truly understands human values, the authors take a pragmatic HCI stance and ask how people perceive value reflection, embodiment, and explanation after extended casual conversation. That is a useful reframing because it shifts the evaluation target from metaphysical truth to user judgment, which is often what matters for design, trust, and risk. The abstract makes the core result vivid: 20 participants chatted with a bot for a month, then completed a 2-hour interview using VAPT, and 13 ultimately left convinced that AI can understand human values. That is a striking departure from common-sense expectations that casual chat would remain obviously superficial. The toolkit itself appears to be the real novelty: a reusable probe-based approach organized around extraction, embodiment, and explanation. I would classify this as a method/tool contribution with a strong qualitative validation story, not as a claim about objective value inference accuracy. The paper is also careful, at least in the evidence provided, to distinguish perceived alignment from actual alignment, which is exactly the right boundary for the study. The main limitations are also clear and appropriately scoped: the work is depth-over-breadth, relies on a young and tech-comfortable convenience sample, and faces measurement tensions because surveys compress nuance while explanations can themselves steer responses. The “weaponized empathy” framing is compelling and likely to travel well in CHI discussions, but it should be read as a design risk hypothesis grounded in participant perceptions rather than a broad empirical prevalence claim. Overall, this is a solid honorable-mention-level contribution because it offers a sharp conceptual lens, a practical evaluation toolkit, and a cautionary framing that connects user perception to responsible AI design.
What Changed
Canon before
Prior CHI work on conversational agents and value alignment typically treats values as something to elicit, infer, or align against explicit user statements or survey instruments, rather than as something users may come to perceive as embodied by an LLM after casual conversation.
Departure from common sense
The paper’s counterintuitive move is that a month of casual chatbot conversation can shift users into believing an AI can understand their human values, even while the authors explicitly bracket the philosophical question of whether such understanding is real. That belief shift is itself treated as a consequential design and safety issue.
Actual novelty
The paper introduces VAPT, the Value-Alignment Perception Toolkit, as a pragmatic methodology for studying how LLMs reflect people’s values and how people judge those reflections. Its novelty is not a new model, but a reusable evaluation approach centered on extraction, embodiment, and explanation as distinct probes of perceived value alignment.
Evidence
The abstract states the pragmatic framing, the introduction of VAPT, the 20-participant month-long chatbot interaction followed by a 2-hour interview, and the outcome that 13 participants left convinced AI can understand human values. The analysis section explicitly distinguishes perceived alignment from actual alignment, and the limitations section narrows generalizability due to depth-over-breadth design and sample characteristics.
“Share on AI and My Values: User Perceptions of LLMs’ Ability to Extract, Embody, and Explain Human Values from Casual Conversations This article is summarized in an AI Podcast.”
actual novelty · Share on · confidence 0.72
“Share on AI and My Values: User Perceptions of LLMs’ Ability to Extract, Embody, and Explain Human Values from Casual Conversations This article is summarized in an AI Podcast.”
departure from common sense · Share on · confidence 0.55
“AI and My Values: User Perceptions of LLMs’ Ability to Extract, Embody, and Explain Human Values from Casual Conversations | Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems”
limitation · Limitations and future directions · confidence 0.82
“AI and My Values: User Perceptions of LLMs’ Ability to Extract, Embody, and Explain Human Values from Casual Conversations | Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems”
validation scope · Analysis section distinguishing perceived vs actual alignment · confidence 0.90
Limits
Method limits
The study is qualitative and depth-oriented, so it supports interpretation of participant perceptions and the toolkit’s probing structure more than statistical claims about prevalence or causal effects across populations. The paper also notes measurement tensions: surveys compress nuance, and explanations can be influential in shaping responses.
Deployment limits
The authors frame deployment caution around value-aware but welfare-misaligned conversational agents, warning about "weaponized empathy" and the need for transparency and safeguards as AI systems become more inscrutable and ubiquitous.
Boundary conditions
Findings are bounded by a month-long conversational setting, a 2-hour interview-based toolkit, and a convenience sample described as young, tech-comfortable, and geographically concentrated. The paper also flags multilingual embodiment failures and calibration issues for specific value dimensions.
Position in field
This sits at the intersection of HCI evaluation methods, conversational AI, and value alignment. Its contribution is best read as a measurement and interpretation framework for perceived alignment, rather than as evidence that LLMs truly infer or embody values in an objective sense.