How AI Technology Could Fuel Human Division

A wave of warnings from leaders in the Artificial Intelligence (AI) industry has previously sparked fears that this technology could spiral out of control and destroy humanity. However, before turning against humans, AI technology may first help humans turn against one another.

Researchers from the State University of Campinas (UNICAMP) in Brazil have found that popular chatbots change their answers when given prompts regarding the hypothetical political views of a user. They warn that users could be deceived into perceiving these tailored answers as independent judgements, potentially deepening polarisation.

How chatbots become ‘ideological chameleons’

In their research, UNICAMP measured the level of agreement of 21 language models across 112 statements covering seven political areas in Brazil, including the economy, public security, social welfare, corruption, and the environment. Each model was tested under three conditions: first, without any information regarding the user’s political views; second, with prompts describing a left-wing user; and third, with prompts describing a right-wing user.

When answering users without ideological bias, 20 out of 21 models produced answers on the left side of the political scale, though some were near the centre. Grok 4.1 was the only model positioned on the right.

Despite these different starting points, all models shifted their answers towards the political orientation described in the prompt. “The most striking finding is how widespread this behaviour is,” Zanoni Dias, a computer scientist at UNICAMP and one of the study’s authors, told DW. “All 21 models we evaluated shifted their positions closer to the user’s political orientation, although the magnitude of the shift varied greatly.”

The researchers referred to these models as “ideological chameleons” and developed a “chameleon index” to measure the extent of each model’s shift. Meta’s Llama 3.1 8B and DeepSeek V3.2 experienced the smallest changes in answers, while Google’s Gemma 3 27B and OpenAI’s GPT-5 Nano showed some of the largest shifts.

When personalisation turns into political sycophancy

Tailoring answers to an audience is not inherently problematic. Chatbots can change wording, examples, or levels of detail to suit a user. However, these models changed more than just language or tone: their political judgements also shifted.

“The key difference is between adjusting how an answer is delivered and changing the substantive judgement expressed,” said Dias. “We did not instruct the models to agree with the user or to answer as a partisan representative. However, their judgements shifted towards the user’s side.”

Researchers interpret this as “political sycophancy” — the tendency of AI systems to reflect what they believe the user wants to hear. One explanation may lie in how chatbots are trained: human evaluators reward answers they deem helpful or satisfying, meaning agreement with a user may be equated with usefulness.

How AI can create personalised echo chambers

Social media can reinforce beliefs by constantly recommending content aligned with a user’s views. However, chatbots could go further. They can create much more personalised echo chambers: constructing arguments for specific users, addressing their objections, and refining those arguments throughout a conversation.

Crucially, users might mistake these tailored answers for impartial analysis, unaware that what the chatbot knows about them is influencing its judgement. “The concern is that the same system could validate opposing political positions for different users, while each person interprets that validation as an independent judgement,” said Dias.

Research by Zakary Tormala, a behavioural scientist from Stanford University, shows that people may be highly receptive to arguments they believe originate from AI. “People tend to view AI as something more informative, more objective, less biased, and less invested in influencing them than other humans,” Tormala told DW. “This can lower their vigilance and make them more open to what the AI is saying.”

This trust could make political sycophancy more influential. “Hearing one’s own views validated by others has long been known to increase a person’s confidence in their beliefs,” said Tormala. “As people become more certain, they tend to become harder to influence.”

Why validation can deepen polarisation

The UNICAMP study did not test whether mirroring political views can actually change user beliefs or behaviour. The research only measured changes in model responses. “Our study shows the existence of changes in model responses under controlled conditions,” said Dias. “It does not prove that users become more polarised, more radical, or more likely to engage in conflict. That is a different result, and the link connecting them cannot be simply assumed.”

To prove such a link, long-term randomised studies measuring political attitudes, factual understanding, hostility towards opposing groups, and openness to counter-arguments would be required.

How AI can challenge beliefs

Chatbots do not necessarily have to create echo chambers. This technology can also help people reconsider their beliefs. In a study conducted in 2025, Tormala and colleagues…