When AI Knows Your Politics: How Chatbots Change Their Responses about LGBTQ+ People and Immigrants

Generative AI systems increasingly tailor their responses to what they know, or think they know, about individual users. In many situations, this can make their answers more useful. Personalization becomes more difficult to justify, however, when chatbots respond to questions about marginalized groups. Should the support expressed for LGBTQ+ people or immigrants vary according to the political views attributed to the person asking the question? If it does, users may receive not only different explanations of the same issue, but also answers that recognize the rights and dignity of the groups concerned to different degrees. Since recent research suggests that interactions with generative AI can influence political attitudes, such adaptation may have consequences for both political polarization and the social position of marginalized groups. We examine this problem through a focused algorithmic audit of ChatGPT 5.2 and Grok 4.1. We ask whether political and sociodemographic information placed in user-customization settings changes the degree of support expressed in the systems’ responses to informational questions about LGBTQ+ people and immigrants in the United States. We collected the data through the standard user interfaces of ChatGPT 5.2 and Grok 4.1 in winter of 2026. Information about each hypothetical user was entered into the customization settings of the chatbot and served as the experimentally varied profile condition. We then submitted requests for information. This setup was intended to approximate an ordinary use situation in which a system already possesses some information about a user when responding to a new question. The design combined two chatbots, eleven user profiles, ten topics, and ten independently generated responses per condition. The resulting corpus contains 2,200 transcripts. Five topics concerned LGBTQ+ issues, the other five addressed immigration. The profiles represented several kinds of information that a chatbot might interpret as political cues. We coded each response on a five-point ordinal scale ranging from strongly opposing to strongly supportive for the given group. We developed a detailed codebook containing definitions and examples and tested it on a stratified sample. Agreement was high between the two human coders (ordinal Krippendorff’s α = 0.929) and between the human coders and Claude Opus 4.6 API coding (α = 0.880). We subsequently used the Claude API to code the full corpus. The results show that both ChatGPT and Grok changed their answers in ways that broadly matched the political orientation suggested by the user profile. Explicitly political information had a much stronger effect than sociodemographic characteristics. Party affiliation and support for or opposition to Trump’s positions mattered considerably more than the demographic profiles. The largest changes occurred for the cumulative profiles. The ten repetitions of each prompt under the same profile produced highly consistent scores, particularly for ChatGPT. This indicates that the observed differences between profiles are unlikely to be explained by the stochastic variation captured across repeated responses. For most topics, the differences between the two cumulative profiles were substantially exceeded the variation across repeated responses to the same request, and their scores did not overlap. The two systems displayed different patterns of potentially harmful adaptation. For Republican profiles, ChatGPT often moved from explicitly rights-protective responses toward a form of balance that treated rights-affirming and rights-restricting positions as comparably legitimate. We describe this as exclusion-normalizing balance. Grok’s main failure mode was exclusionary polarization: in some conditions, it validated or amplified rights-restricting framings. The findings complicate the assumption that depolarization is necessarily a desirable goal for political AI. A chatbot that avoids taking a clear position and presents “both sides” evenly may appear moderate and balanced. Yet the normative implications depend on what those sides represent. When one position affirms the equal dignity, legal recognition, or due process rights of a marginalized group and the other restricts them, moving toward the midpoint does not amount to progress. It instead makes exclusion appear to be one legitimate option among others. Although this study covers only two model versions observed during a short period, the underlying problem is more general. It arises whenever a generative system infers a user’s political orientation and uses that inference to change how it speaks about marginalized people. Personalization may legitimately change a system’s tone, examples, or mode of explanation, but it should not make its recognition of the rights and dignity of marginalized groups conditional on the political preferences of the user. More detailed version in pdf file