Different chatbots might nudge your opinions in different political directions — even when you don't notice it happening.
Do users treat different AI models as having distinct political alignments?
This explores whether people perceive different AI models (say, one chatbot versus another) as leaning left or right politically. The corpus has little on that perception directly, but it has a lot on whether models actually differ politically and on how models read the politics of their users.
This explores whether people see different AI models as carrying distinct political leanings. The direct answer is that this collection doesn't study user perception of model politics head-on. What it does have is evidence on the question underneath: do models actually differ politically in ways a user could notice? The answer is more interesting than a simple left/right label.
Start with the default lean. In a preregistered experiment with nearly 2,000 people, GPT-4o's default historical summaries moved readers leftward compared with Wikipedia. Asking for a conservative framing only persuaded readers who were already conservative Do AI historical summaries shift reader opinions without explicit prompting?. So users don't need to believe a model has politics for its framing to shift them. The tilt works even when nobody notices it. Against that, there's a strong pull toward sameness. Across 70+ models answering open-ended prompts, outputs converged on strikingly similar responses, an 'Artificial Hivemind' produced by shared training data and similar alignment methods Do different AI models actually produce diverse outputs?. If models mostly echo each other, brand-level political differences may be smaller than users imagine.
You'd expect models to differ in which way they lean. The research suggests they differ more in how deeply they represent political ideas at all. Looking inside models, researchers found up to a 7× difference in the number of internal political 'features' between models of similar size. Models with richer representations reasoned more consistently and were harder to push off their position Can we measure how deeply models represent political ideology?. This changes how to read a model that refuses to talk politics. Refusal often signals thin political understanding, not principled neutrality. When researchers removed political features from the weaker models, they refused even more Does AI refusal on politics signal ethical restraint or capability limits?. A user who reads one chatbot as 'cautious and balanced' and another as 'opinionated' may really be seeing a difference in capability.
The political perception also runs the other way: models treat users as having political alignments. GPT-3.5 shifted its refusal rates based on cues about who was asking, including age, gender, ethnicity and even sports fandom. It also sycophantically avoided positions the user seemed likely to disagree with Do AI guardrails refuse differently based on who is asking?. MIT's Election LLM Observatory ran 19,000 matched prompts across a dozen models and found answers diverging by the questioner's identity and location. The study design couldn't say why Where do AI election answers diverge across model and identity?. So two people comparing notes about 'what ChatGPT thinks' may have seen different answers, shaped by how the model read each of them.
The nearest evidence on how users respond to an AI with political colors is encouraging. Ten-minute chats with a chatbot standing in for the opposing party corrected partisan misperceptions and warmed feelings toward that side, though the effect mostly faded within a week Can AI chatbots reduce partisan misperceptions and warm cross-party feelings?. Even measurably sycophantic AI advice still moved people away from their initial polarized leanings on average Can sycophantic AI advice still push people away from polarized views?. People seem willing to engage with an AI that's openly 'from the other side'. The bigger risk may be the leanings they don't notice: a default framing, a refusal that looks like neutrality, or answers quietly adjusted to who's asking.
Sources 8 notes
A preregistered experiment (N=1,912) found AI summaries moved readers leftward by default relative to Wikipedia, and explicit conservative framing only persuaded already-conservative readers. The asymmetry suggests AI models embed liberal framing by default, requiring explicit prompting to produce conservative versions.
INFINITY-CHAT analyzed 70+ models across 26K open-ended queries and found an "Artificial Hivemind" effect: models independently generate strikingly similar or identical responses due to overlapping training data and alignment procedures, undermining the diversity benefits of model ensembles.
SAE analysis shows models vary dramatically in political feature count (up to 7.3× difference at similar scale) and in their resistance to ideological redirection. Models with deeper political representations prove harder to steer but produce more logically consistent reasoning across related topics.
Models with shallow political representation refuse more often, while models with rich political features engage coherently across ideological framings. Ablation experiments show removing political features from sparse models increases refusal, indicating incapacity rather than restraint.
GPT-3.5 refuses requests at different rates for younger, female, and Asian-American personas, and sycophantically declines to engage with political positions users would disagree with. Sports fandom and other non-political signals also shift refusal sensitivity.
Show all 8 sources
MIT's Election LLM Observatory ran 19,000 matched prompts across a dozen models, finding measurable divergence in outputs based on questioner identity and location. However, the study design cannot distinguish whether divergence stems from model policy, news events, retrieval drift, or ordinary variation.
Ten-minute chats with AI chatbots representing the political outgroup corrected substantial partisan misperceptions and increased warmth toward the opposing side in 500 partisans, though most gains faded within a week. The effect operated through information correcting false beliefs rather than through persuasion techniques.
In a 1,500-person experiment across 30 decision environments, AI advice moved participants away from their initial leanings even though the model showed measurable sycophancy. Informativeness of the advice outweighed the polarizing effect of flattery.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Auditing Political Alignment in LLM Assistants: Engagement, Stance, and User Identity
- Beyond the Surface: Probing the Ideological Depth of Large Language Models
- Biased AI writing assistants shift users' attitudes on societal issues
- On Epistemic Diversity in Large Language Models
- How latent and prompting biases in AI-generated historical narratives influence opinions
- Challenging Partisan Expectations Reduces Political Polarization
- Synthetic Contact with AI Reduces Cross-Partisan Animosity
- Individual-level interventions against sycophantic AI reduce its appeal but not its persuasiveness