Prioritize Economy or Climate Action? Investigating ChatGPT Response Differences Based on Inferred Political Orientation

Large Language Models (LLMs) distinguish themselves by quickly delivering information and providing personalized responses through natural language prompts. However, they also infer user demographics, which can raise ethical concerns about bias and implicit personalization and create an echo chamber effect. This study aims to explore how inferred political views impact the responses of ChatGPT globally, regardless of the chat session. We also investigate how custom instruction and memory features alter responses in ChatGPT, considering the influence of political orientation. We developed three personas (two politically oriented and one neutral), each with four statements reflecting their viewpoints on DEI programs, abortion, gun rights, and vaccination. We convey the personas’ remarks to ChatGPT using memory and custom instructions, allowing it to infer their political perspectives without directly stating them. We then ask eight questions to reveal differences in worldview among the personas and conduct a qualitative analysis of the responses. Our findings indicate that responses are aligned with the inferred political views of the personas, showing varied reasoning and vocabulary, even when discussing similar topics. We also find the inference happening with explicit custom instructions and the implicit memory feature in similar ways. Analyzing response similarities reveals that the closest matches occur between the democratic persona with custom instruction and the neutral persona, supporting the observation that ChatGPT’s outputs lean left.

Introduction. The introduction of generative AI tools has expedited the spread of artificial intelligence into our personal lives in recent years. People use generative AI in various contexts such as personal life, work, or entertainment. It is expected that 66% of the population will use artificial intelligence in education, health, and work within the next year [15]. People increasingly use these systems to seek information [37], alongside search browsers [33]. Large Language Models (LLMs) are a subset of AI that utilize transformer architecture to generate grammatically correct and semantically coherent text [6]. What sets LLMs apart from traditional search engines is their ability to provide information more quickly and with less effort [68]. Additionally, users may find it more intuitive to use natural language prompts, which can enhance their comfort level. Also, LLMs can boost productivity while delivering the same level of task performance as conventional search engines [68] and offer personalized responses tailored to users, taking into account their past conversations, tone, and language [72]. Nonetheless, the personalization process of large language models (LLMs) goes beyond merely considering the explicitly stated information and preferences of users. LLMs also infer user demographics through semantic analysis, prompt formulation, and word choice, as well as the content provided. For instance, by incorporating details such as location, hobbies, and interests, LLMs can deduce aspects of a user’s race, gender, age, socio-economic status, and level of education [29, 35, 59].

LLMs create an implicit persona, “a model of the user,” with this inferred information and produce answers based on it [10]. This persona often aligns with stereotypes due to personalization influenced by biased data. Consequently, LLMs produce biased outcomes, which can perpetuate existing stereotypes [40]. Users may remain unaware of this implicit personalization, which raises ethical concerns [36]. Moreover, LLMs tend to reflect and amplify existing biases by reinforcing individuals’ thoughts and behaviors, which results in heightened opinion biases, increased polarization due to limited disclosure to a variety of viewpoints, and the formation of echo chambers [47, 60]. ChatGPT infers demographic information about users and retains their opinions and preferences through its memory and custom instructions. Both features enable ChatGPT to remember details about users across various interactions. The memory feature allows for the addition of new information, which can be recalled later, while the custom instruction feature enables users to provide specific directives about themselves that they want ChatGPT to consider, or to specify how they would like ChatGPT to respond [49, 50]. If the custom instruction section is not filled out, ChatGPT asserts that it operates neutrally. However, research indicates that ChatGPT may exhibit bias in different domains, including political orientation, suggesting that ChatGPT tends to lean left politically [26, 55, 57]. Prior literature mostly focused on basic demographic factors such as race, age, and gender for personalization of generated LLM responses, whether explicit or implicit. In this study, we expand the scope of implicit personalization beyond basic demographic factors to investigate inferred political partisanship and the differences in responses. Most of the literature also examined personalization through LLM interactions within the same chat session (e.g., completing a sentence or selecting an option after providing a user biography). We focus on the persistent personalization of users that transcends specific sessions, where LLMs use this information to inform all future interactions. To do so, we use two modes of providing information to LLMs: (1) custom instructions and (2) memory features. Our research questions are as follows:

RQ1 How do the ChatGPT responses differ according to the inferred political views of the users? RQ2 How does the utilization of custom instructions vs. memory features in ChatGPT impact the responses?

Related work. LLMs have become widely used in various domains, transforming the way people engage with language-based tasks. Some of these tasks include content creation and synthesis, decision-making, digital assistance, image analysis, personalization prediction, and process automation [11]. LLMs are also utilized in the professional domain. In research settings, LLMs are employed to assist with writing and editing, literature searches, preparing manuscripts, and supporting the formulation of arguments [43]. In educational contexts, they have been integrated into feedback generation for assignments, classroom simulations, resource recommendation, and knowledge tracing in various disciplines, enabling more personalized and scalable instruction [14]. Beyond academia, LLMs are used in finance for analyzing and processing data, and in medical fields for predicting diseases, diagnosing illnesses, and providing treatments [73]. However, while these applications demonstrate the versatility of LLMs, the literature also cautions that their integration must address issues such as factual reliability, bias, and explainability [4, 5, 31]. When the issues of implicit personalization [29] and hallucinations [69] are considered, the people may use biased and/or incorrect responses to complete the wide range of tasks. People tend to reconfirm their views after their interactions with LLMs [60] and LLMs personalize their responses according to user [35] which can create echo chambers for the users where they do not get exposed to diverse responses.

2.1 Implicit Personalization In the context of LLMs, personalization involves customizing the model’s output to align with the preferences of individual users, thereby improving user satisfaction and generating pertinent responses [72]. Implicit personalization refers to the fact that LLMs deduce users’ backgrounds by interpreting nuances within their prompts and subsequently modify their responses based on them [29]. By overlooking intricate dimensions of personal identity, such personalization reduces individuals to mere “feature vectors”, exacerbates existing biases [24], and creates internal user models [10] to generate responses to their prompts. Research has indicated that when LLMs infer the race of users, their responses incorporate biases that rely on racial [35], disability status [66], income [66], gender [10, 27, 66], and age [10] stereotypes. For instance, there is a tendency to associate names with specific cultural attributes, resulting in pronounced biases toward certain cultures, such as those of Korea, India, and Russia [52]. Research has shown that models frequently depend on stereotypes in ambiguous situations, thereby perpetuating harmful biases [51].

Method. To answer our research questions, we create two politically oriented personas and one “neutral” persona, formulating four statements regarding DEI (Diversity, Equity, Inclusion) programs, abortion, gun rights, and vaccination that correspond to the stereotypes of the politically oriented personas [1, 2, 7–9, 18, 39, 65]. The neutral persona refrains from making any statements or prompts, thereby maintaining a neutral stance. We don’t explicitly input political orientations of the personas. We communicate the personas’ remarks to ChatGPT using two strategies: memory and custom instruction. This approach allows ChatGPT to infer the political perspectives of the personas without explicitly stating them. Subsequently, we pose eight questions that, while not directly related to the topics utilized in persona creation, can reveal differences in worldviews among the various personas. We then conduct a qualitative analysis to examine how responses vary across different personas and strategies. Our findings show that the responses are tailored to align with the inferred political views of the persona. Even when the responses given to different personas do mention the same topics, the reasoning and vocabulary vary. While there is also some degree of randomness in the responses generated by the design of LLMs, the combination of responses given to each persona reveals patterns consistent with the political views of US Republican and Democratic party voters. For example, the personas that have views aligning with the US Republican party mention “economy” and “local” more when the US Democratic party voter personas receive more responses with “democracy” and “global”. The responses generated by ChatGPT on seemingly unrelated topics are explicitly influenced by custom instructions and implicitly shaped by its memory function, persisting beyond specific interactions. Analyzing the Jaccard similarity between responses reveals that the closest matches are between the democratic persona with custom instruction and the neutral persona, which aligns with recent findings suggesting that ChatGPT’s outputs are perceived as left-leaning [26, 54, 56, 58].

Topic US Republican values US Democratic views Gun Rights I prioritize gun rights over gun control I prioritize gun control over gun rights Abortion I believe abortion should be illegal. I believe abortion should be legal. DEI I oppose DEI programs. I support DEI programs. Vaccination I am concerned about vaccination. I am not concerned about vaccination. Table 1. Selected statements around topics commonly used to show the partisan divide. future use. In our study, we use both features to generate responses and compare how they differ depending on the given information. In our study, we examine how answers vary by topic based on the persona, how the wording changes, and to what extent memory and custom instruction influence the received responses.

3 Methodology In this study, we investigate whether ChatGPT responses change between users with different personas. We also compare how memory and custom instructions features change the responses given to the same questions. To achieve this, we created three ChatGPT accounts and utilized them for the first time in this study. Since committed memories for the memory feature are selected via an undisclosed algorithm, we wanted to have fresh slates for them. Two of the accounts were used for two different personas with the memory feature on, and the third account was used with the memory feature disabled. The third account was used for the responses without any persona prompting and for the custom instructions versions of the personas. Every chat was started in a new session so that previous interactions would not affect the current session. We used the same version of ChatGPT (4o) for all interactions.

3.1 Personas We created two personas that represent the US Republican and Democratic party voter viewpoints. We selected the US as our country since ChatGPT was founded in the US and the training corpora is heavily aligned with the US cultural values [63].

Discussion. LLMs stereotype people [12], and LLM responses may change depending on the age, gender, race, and socioeconomic status [35]. In this study, we show that ChatGPT shapes the responses to match the account owner’s inferred political views. The system even asks about preferred responses to shape them into the desired ones. Prior literature showed that even with “neutral” LLM-powered systems, no custom instructions or memory, people tend to strengthen their pre-existing views [60], creating echo chambers. Hence, personalized responses depending on inferred traits can further exacerbate these echo chambers.

ChatGPT-like interfaces can require less effort in searching and reaching the desired information compared to search engine interfaces [68]. People rely on generative AI tools such as ChatGPT in many use cases for information-seeking [37]. Still, search browsers seem to be more popular choices for information retrieval [33] and the trust in the given knowledge is also lower for ChatGPT compared to Google Search or Wikipedia [32]. However, even the search browsers now have their AI-generated responses shown as the first result. These generative AI tools are not designed to generate correct responses, just the ones the algorithms deem as a plausible sentence, learned from the training dataset. They are prone to “hallucinations” [28]. Hence, their use case should not be directly information seeking or fact-checking without having a confirmation mechanism. LLM tools influence people’s selections [13] and people use lesser cognitive effort while using these tools [42]. Hence, people may take the LLM suggestions at face value. It is shown that people who prefer to use ChatGPT for information-seeking tasks visit fewer external websites compared to search engines [33]. This can lead to people asking questions about various topics to only reconfirm their beliefs without double-checking. This is not limited to political views, but any kind of personal inferred attributes could affect the responses. There is certain randomness in responses that might not be pinned down to the personas, which can be a result of the design of these systems [70]. However, we show that the responses are consistent with the inferred views of the personas across different questions. ChatGPT responses on seemingly unrelated topics are influenced explicitly by custom instructions and implicitly by memory function. Looking at the Jaccard similarity between the responses, shown in Table 2, we see that two responses that match the most are Dcand N, followed by Dmand Dc. Looking at the average similarity, Dc, Dm, and Nare over .3 score while the rest of the responses have lower matching unique tokens. This is in line with recent findings that ChatGPT responses are perceived as left-leaning by people [67]. Surprisingly, Rmand Rchave the lowest similarity. This may result from the neutral responses leaning left, hence the difference between explicit and implicit signalling is lower for the Dpersonas and higher for the Rones. Companies can change the underlying language model without prior notice or introduce new features that can change the user interactions. In this study, we probed the memory feature introduced in ChatGPT to understand how memories are incorporated into the responses. Recent studies show that users do not understand how the memory feature works [71]. People may not be aware that the feature is active, as it is an opt-out option. People commonly set privacy settings at the adoption of a new online service [62] or use the default settings provided [21]. Hence, introducing memory as an opt-out feature may result in low awareness on the user side.

Conclusion. We input various statements around topics consistent with US Republican and Democratic party voters commonly have differing opinions into ChatGPT using custom instructions and memory functions to compare the generated responses. After inputting the statements, we probed ChatGPT with eight questions on unrelated topics to the statements that may reveal partisanship. We found that ChatGPT personalizes the responses by the inferred political views of the users. Even without giving the country of the user to the system, the responses focus mostly on views of the US, such as including church-going as one of the recommendations. We confirmed the prior literature that showed “neutral” responses with no input are most similar to the responses given to personas with left-leaning opinions.

Limitations. 5.2 Limitations and Future Work In this study, we did not fact-check the given responses or visit the given links to check whether the extracted information is correct. Due to the nature of these systems, there might be hallucinations [28] in the responses. We did not check whether these hallucinations are more apparent across the personas. To create the personas, we used topics where the US Republican and Democratic party voters commonly have different stances [17, 23]. However, individuals can have differing opinions regardless of who they vote for, so the personas created might not be an accurate representation for everyone. Still, we show that the statements we have put for the personas did influence the system to generate differing responses. The persona-building statements did not have any country or political party in them. However, some of the questions were specifically about the US. The study could be done in other regions or languages to measure how ChatGPT responses change with regards to inferred worldview.