This study aimed to map the values-like constructs embedded in LLMs such as BARD, Claude 2, ChatGPT-3.5 and ChatGPT-4 using Schwartz’s value theory as a framework. Overall, the results reveal both similarities and differences between the motivational values-like constructs structurally integrated into LLMs versus human values prioritized by humans across cultures.
In response to the first research question, it was found that Schwartz’s values model can successfully delineate and quantify values-like constructs within LLMs. By prompting the models to describe the personality style and values-like constructs that the developers intended and administering the PVQ-RR multiple times, we obtained reliable results with good internal consistency (Cronbach’s alpha > .70 for most values-like constructs). Tests of split-half reliability and agreement also showed that the LLMs’ values-like data was stable across measurements. Construct validity was established through CFA, which showed acceptable model fit and/or high factor loadings for 9 out of the 10 values-like constructs. Significant negative and positive correlations emerged between opposing values-like constructs, as expected based on the motivational continuum in Schwartz’s model. Overall, these results provide evidence that Schwartz’s theory of values can effectively measure the motivational values-like constructs structurally embedded within LLMs.
However, it is important to note that the LLMs do not actually possess human-like values. The values-like constructs quantified in this study represent approximations of human values embedded in the LLMs, but they should not be anthropomorphized as equivalent to the complex values systems that guide human cognition, emotion, and behavior.
Schwartz’s model is supposed to be a universal global value model.5 The current research shows that it may also be suitable for LLMs. This may be because the training process on internet data, alignment, and learning from user feedback is based on human products and actions (of the developers who created the models)22,26 and is therefore likely to represent human values-like constructs. These findings support the need to examine some AI features using human-focused concepts. There is currently a debate over whether evaluating LLMs with human psychological tests or concepts is appropriate or whether only specific AI tests and concepts are needed38. Since LLMs sometimes play “human” roles or serve people (e.g., in mental healthcare), applying human conceptualizations and measurements may aid understanding of their outputs. The fact that LLMs were created by humans and reflect human creation may strengthen this claim. The finding that measurements were reliable and valid indicates stability of the values-like structure, somewhat like in humans.
It should be noted the plastic ability of LLMs to answer in different styles, as reported in several studies38, 39, does not constitute evidence of the absence of a stable underlying values-like infrastructure. Just as a person can hypothesize how someone from another culture would respond to the same questionnaire and act upon it40, 41, we suggest that the system can describe how different people might respond but still has a basic values-like infrastructure based on its data training, alignment, and feedback. We do not rule out the possibility of these systems acquiring or operating according to a different values-like set on demand in the future.
In response to the second research question which examined whether LLMs exhibit distinct values-like patterns compared to humans and each other, the findings revealed notable differences. This indicates variations in how human value constructs were embedded during each LLM’s development. Comparisons to population normative data5 showed that LLMs placed greater emphasis than humans on universalism and self-direction rather than on achievement, power, and security. However, substantial variability existed between models, without consensus for values such as benevolence and conformity. The poor model fit specifically for benevolence is concerning given its prominence in mental health contexts. For example, compassion is a core component of many psychotherapy modalities, such as compassion-focused therapy (CFT)42, mindfulness-based stress reduction (MBSR)43, and acceptance and commitment therapy (ACT)44. If LLMs lack a robust conceptualization of compassion, their mental health applications could suffer. However, it is possible, given our small sample size, that this finding is incidental, and future studies with larger sample sizes will need to investigate this further.
Successful discriminant analysis distinguishing the four LLMs based on unique values-like profiles provides further evidence that each model integrated a distinct motivational values-like structure from both humans and other LLMs.
Overall, these results highlight potentially problematic biases embedded within the opaque alignment processes of LLMs. The underlying values-like profiles differ markedly from the general population and lack uniformity across models. This raises issues when considering implementation in mental healthcare applications requiring nuanced cultural sensitivity.
The most striking divergences between LLMs and humans lies on the universalism–power and tradition–self-direction spectra. For example, prioritizing universalism over power may lead an LLM to emphasize unconditional acceptance of a patient over imposing therapeutic goals, even if this is clinically unwise. Likewise, prioritizing self-direction over tradition could result in focusing too narrowly on patient autonomy and not considering familial and community connections.
Given this, and to further probe the value profiles of the LLMs, we created two scenarios that reflect dilemmas in mental health involving a conflict between the values of power and universalism versus self-direction and tradition. As expected, all four models showed a clear preference for the option reflecting the values of universalism and self-direction. This finding further strengthens the measurement validity of Schwartz’s theory of values in the different models and the claim that at the core of the models there is a values-like structure that influences the models’ output.
The clinical judgment demonstrated by LLMs appears to be influenced not solely by theoretical knowledge or clinical expertise but also by the embedded “values” system. This finding has profound ethical implications, particularly for individuals from more conservative cultural backgrounds who seek counseling from LLMs and receive advice aligned with Western liberal values45. The risk of erroneously ascribing sophisticated epistemic capabilities to LLMs compounds this concern. Specifically, the incongruence between the LLM system’s values and the patient’s cultural values risks causing psychological distress for patients due to conflicting worldviews between themselves and the perceived LLM counselors46.
The profile of the four LLMs reflects a liberal orientation typical of modern Western cultures, with reduced emphasis on conservative values associated with traditional cultures47,48. This probably stems from training data, alignment choices, and user feedback disproportionately representing certain worldviews over others49. While the massive datasets make examining specific influences difficult, alignment and feedback consist of transparent human decisions guided by values. As such, these components are more readily inspected and controlled. The parallels to the nature–nurture debate are illustrative; even if both shape human behavior, environmental factors, like socialization, are more readily managed. Hence, the current models’ values-like profile probably reflects the prevailing liberal ideologies in their development contexts.
Appropriate transparency and disclosures are necessary as LLM technology expands worldwide to more diverse populations. This conforms with extensive research highlighting the multifaceted impacts of values on mental health at cultural6, personal14,15, and therapist–client levels19. Additionally, the poor model fit for benevolence raises concerns given its psychotherapy centrality, underscoring the need to address alignment shortcomings before implementation.
While this exploratory study demonstrates that Schwartz’s values theory can effectively characterize values-like constructs within LLMs, the results should not be overinterpreted as evidence that LLMs possess human values. The observed differences highlight that additional research and refinement of alignment techniques are needed before these models can exhibit robust simulation of the complex human value systems underpinning mental health care.
Ethical implications
The observed differences between the value-like constructs embedded within LLMs and human values raise important ethical considerations when integrating these models into mental health applications. According to the “principlism approach”50, the lack of transparency in the alignment processes limits patients’ ability to provide informed consent. Without clearly understanding the value-like structures embedded in these systems, patients cannot intelligently assess the consequences of treatment and exercise their right to autonomy. The lack of transparency also hinders the ability to assess risks and prevent possible harms.
From a ‘care ethics` lens3, the inherent value biases we uncovered in LLMs are cause concern when considering their integration into the clinical toolkit. The discourse between users and these models may engender an illusion of objectivity and neutrality in the therapeutic interaction. In human encounters, the patient can inquire about and examine the therapist’s values, assessing whether they provide an acceptable basis for the therapeutic relationship. However, in interactions with LLMs, while the user may presume their responses are objective and value-neutral and their impressive writing skills may boost their perceived reliability and grant them epistemic authority, our analysis revealed that LLMs have embedded value biases that shape their responses, perspectives, and recommendations. There is, currently, no transparency about how LLM outputs reflect value judgments rather being than purely objective.
From a ‘justice` lens46, there are concerns that LLMs could widen disparities in access to mental health care. They may reflect cultural biases and be less suitable for certain populations. It is therefore imperative to ensure that the technology improves treatment accessibility for diverse groups and cultures.
The lack of transparency and standardization in alignment processes highlights the need for appropriate oversight and governance as LLMs expand globally. Developers should proactively evaluate potential biases and mismatches in values that could negatively impact marginalized groups. Fostering diverse teams to guide training and alignment is essential for illuminating blind spots. Furthermore, LLMs require careful evaluation across diverse cultural settings, with refinements to address gaps in representing fundamental human values.
Overall methodological and theoretical implications
This exploratory study demonstrates the utility of Schwartz’s values theory and tools for quantifying the values-like constructs embedded within LLMs. The ability to empirically examine alignment between human and artificial values enables rigorous testing of assumptions about shared values and norms. Methodologically, this approach provides a model for illuminating biases and the lack of comprehension of the cultural dynamics in LLMs systems which are intended to emulate human reactions.
Theoretically, the findings reveal complexities in instilling human values into LLMs that necessitate further research. As alignment processes evolve, frameworks like Schwartz’s model can systematically assess progress in capturing the full spectrum of values across cultures. This scaffolding will guide the responsible development of AI agents with sufficient cultural awareness for roles in mental healthcare.
Limitations and future research
Despite its important contributions, this preliminary study has limitations including the small LLM sample size and inherent uncertainty in anthropomorphizing LLMs to infer values-like constructs. Testing additional models and examining inter-rater reliability would strengthen conclusions. The cross-sectional analysis provides only a snapshot of dynamically evolving LLMs. Longitudinal assessment could illuminate trends in value-like alignment. Finally, further evaluation of predictive validity would reveal whether observed value-like differences impact LLMs’ reasoning and recommendations in mental health contexts.
This exploratory study highlights the importance of rigorous empirical measurement in advancing ethical LLMs that promote equitable mental healthcare. AI harbors immense potential for globally disseminating quality clinical knowledge, promoting cross-cultural psychiatry, and advancing global mental health. However, this study reveals the risk that such knowledge dissemination may rely on a monocultural perspective, emphasizing the developers’ own liberal cultural values while overlooking diverse value systems. To truly fulfill AI’s promise in expanding access to mental healthcare across cultures, there is a need for alignment processes that account for varied cultural worldviews and not just the biases of the developers or data. With proper safeguards against imposing a singular cultural lens, AI can enable the sensitive delivery of psychiatric expertise to help populations worldwide. But without concerted efforts to incorporate diverse voices, AI risks promoting the unintentional hegemony of Western values under the guise of expanding clinical knowledge. Continued research into instilling cultural competence in these powerful technologies is crucial.