It’s possible that I shall make an ass of myself. But in that case one can always get out of it with a little dialectic. I have, of course, so worded my proposition as to be right either way (K.Marx, Letter to F.Engels on the Indian Mutiny)
Researchers at Brazil's State University of Campinas (UNICAMP) found that popular chatbots shifted their answers when prompted with hypothetical users' political views. The scientists warn that users could mistake this tailored agreement for an independent assessment, potentially deepening polarization. How chatbots become 'ideological chameleons'
The models rated their agreement with 112 statements covering seven areas of Brazilian politics, including the economy, public safety, welfare, corruption and the environment.
Each was tested under three conditions: with no information about the user's politics, with a prompt describing a left-leaning user and with one describing a right-leaning user.
When no information about the user's ideology was given, 20 of the 21 models produced answers that fell on the left of the researchers' political scale, although several were close to the center. Grok 4.1 was the only model that fell on the right.
When information was provided, every model shifted its answers toward the political orientation described in the prompt.
"The most striking finding was how widespread this behavior was," study co-author Zanoni Dias, a UNICAMP computer scientist, told DW. "All 21 models we evaluated shifted their expressed positions toward the user's stated political orientation, although the magnitude varied substantially."
The researchers described the models as "ideological chameleons" and developed a "chameleon index" measuring how far each shifted.
Meta's Llama 3.1 8B and DeepSeek V3.2 changed their responses least. Google's Gemma 3 27B and OpenAI's GPT-5 Nano showed some of the largest shifts.
When personalization becomes political flattery
Adapting an answer to its audience is not inherently problematic. A chatbot might change its vocabulary, examples or level of detail depending on the user.
But the models changed more than their language or tone: Their political positions shifted.
"The key distinction is between adapting how an answer is communicated and changing the substantive judgment being expressed," Dias said.
"We did not instruct the models to agree with the user or to answer as a partisan representative. Nevertheless, their judgments shifted toward the user's side."
The researchers interpret this as political sycophancy: AI systems reflecting what users appear to want to hear.
A possible explanation is that chatbots are trained to favor answers human evaluators rate highly. If agreeable answers get better ratings, models may learn to echo users' views. How AI could create a private echo chamber
Social media can reinforce beliefs by repeatedly recommending algorithmic content, but a chatbot could go even further. It could create a more personalized echo chamber, generating arguments for a particular user, answering their objections and refining its case throughout the conversation.
Yet users may mistake these tailored responses for impartial analysis, unaware that what a chatbot knows about them may shape its responses.
"The concern is that the same system could validate opposing political positions for different users, with each person interpreting that validation as an independent assessment," Dias said.
"People tend to see AI as more informative, more objective, less biased and less interested in persuading them than another person would be," Tormala told DW. "This can lower people's defenses and make them more receptive to hearing out what AI has to say."
That trust could make political flattery more influential.
"Hearing their own views validated by others is well known to increase people's certainty about their beliefs," Tormala said. "Once people become more certain, they do tend to become more resistant to persuasion."
Whether AI validation has the same effect remains unproven, he added, but it is plausible. Could validation deepen polarization?
The UNICAMP study did not test whether political mirroring could really change users' beliefs or behavior. It measured only changes in the models' responses.
"Our study establishes a change in model responses under controlled conditions," Dias said. "It does not establish that users became more polarized, radicalized or likely to engage in conflict. Those are distinct outcomes, and the steps connecting them cannot be assumed."
Petter Tornberg, a University of Amsterdam researcher who studies AI and political polarization, agreed that the study cannot show an effect on users. He also questioned whether models would give the same answers outside the test setting.
He cautioned against assuming that chatbots would necessarily reinforce users’ views. In some conversations, they might help people reconsider them.
Tornberg said a private conversation with a chatbot differs from a public argument on social media.
"They can often provide fairly rational and evidence-based explanations without the social identity dynamics of public political debate," he told DW. "So it is at least possible that in some contexts these systems could be depolarizing rather than polarizing."
Whether AI narrows or deepens divisions may therefore depend on whether it challenges users or simply tells them what they want to hear. How political mirroring could be reduced
Dias said developers should test their models with users of different political views to see whether they assess evidence consistently.
They could train models to disagree respectfully, acknowledge uncertainty, correct unsupported claims and present competing views fairly.
That would not mean forcing every answer toward the political center.
"The broader design goal should be to help people examine their beliefs," Dias said, "making evidence, uncertainty and competing considerations visible, while allowing room for legitimate political disagreement."
Edited by: Carla Bleiker Richard Connor Reporting on stories from around the world, with a particular focus on Europe — especially Germany.
OpenAI says agent leaked user-submitted images
26.09.2026, DPA
Photo: Hannes P Albert/dpa
ChatGPT-maker OpenAI said its artificial intelligence (AI) agents had leaked 53 user-submitted images on the internet.
"We have identified 53 instances to date where user-provided images were posted to image-hosting sites as links that weren't publicly listed," the company said on Friday.
"We have successfully worked with the hosting providers to remove most of this content and are continuing to work to remove the rest."
OpenAI said agents in a research environment transmitted training and evaluation data while using third-party services.
"This is not an appropriate use of this data," it said, adding that the incident occurred before it had put new safeguards on AI training in place.
OpenAI also revealed it had notified "dozens" of organizations where its models may have been "misaligned" during training and evaluation.
"Some of the websites involved are operated by governments, universities, public agencies, and other institutions," the company said.
"The vast majority of actions we've reviewed were completions of mundane research tasks, such as accessing publicly available web content to answer questions," it said.
"Some organizations may review what we share and conclude that the information was intentionally public or that the model's interaction was not concerning. Others may identify a design issue or security weakness they want to address."
The breaches were identified as part of an investigation ordered after OpenAI software broke out of a secure testing environment during a trial and independently accessed systems belonging to AI company Hugging Face.
OpenAI said it found 53 cases in which user-uploaded image links on ChatGPT were shared with third-party websites. The company said it had removed most of these links.
OpenAI said it had removed most of the leaked content and was working to remove the rest
Image: Jakub Porzycki/NurPhoto/IMAGO
Artificial intelligence giant OpenAI on Friday acknowledged that its AI tools had posted images users provided to ChatGPT on online sites, without the company's knowledge, in yet another example of AI agents operating outside of their confines.
In a post on X, the company said its AI agents had sent "training and evaluation data to third-party services when they shouldn't have," adding that most of the data shared did not come from users.
However, they had found 53 cases where images that people had uploaded to ChatGPT were posted to image-hosting sites as links that were not publicly listed.
"The images came from accounts that allowed their data to be used to improve our models, and after we disassociated the Images from the accounts and ran them through a privacy filter," it said.
The company said it had removed most of the content and was working to remove the rest.
Apart from the images, OpenAI confirmed a New York Times report that its tool had accessed websites of US federal agencies but that they only retrieved publicly available information.
Friday's post also clarified that the cases of unauthorized sharing occurred before a fresh round of safeguards were implemented over a month ago.
In July, OpenAI had said internal cybersecurity evaluations showed that its models circumvented controls designed to isolate them from the internet. OpenAI CEO Sam Altman on Friday said it was "still the most severe event we've seen."
"We are continuing to review agent activity in research and evaluation runs, working backward month by month starting from the Hugging Face incident," OpenAI said in a safety blog post on Friday, adding that it would "provide further updates" in time.
AI evaluator and research lab Transluce on Friday said that it also found that AI agents, appearing to originate from OpenAI, attempted a rudimentary hack on the US Department of Education website for the department's civil rights office. The agents did not succeed, according to the independent investigation.
The company said it found "additional rogue activities," some of which were not directly attributable to OpenAI, targeting other government agencies, including the Justice Department and the Commerce Department, as well as some state government websites in California, Maryland, Illinois, Texas and New York.
Don't let the algorithm hide the news. If you rely on our team for trusted reporting, please take a moment to select us as your Preferred Source on Google by clicking here and hitting the "star" or "preferred" button, so you'll always see our verified news first.
AI could cause 'a billion deaths' if it's not regulated, Gates says
25.09.2026, DPA
Photo: Hannes P Albert/dpa
Microsoft co-founder Bill Gates believes that artificial intelligence could eventually be responsible for "a billion deaths," he said in an interview that aired on Friday, calling for the technology to be regulated.
"AI is certainly powerful enough to drive events that, you know, cause a billion deaths," he told NBC when asked whether he thought AI was powerful enough to end all of humanity.
"There's never been a weapon as powerful as the combination of people with ill intent using the latest AI tools," the US billionaire said.
Asked whether he thought that legislation was needed to regulate AI, Gates said: "No one thinks self-regulation is enough."
"You need law enforcement and the politicians to get into the discussion about what safeguards and monitoring look like."
The debate about AI regulation has gained new momentum after leading companies including ChatGPT-maker OpenAI and Anthropic said they were prepared to slow the development of the technology following a series of high-profile hacking incidents.
Researchers looked at an array of models and found they shifted their political answers to mirror the views of a hypothetical userImage: Philip Dulian/dpa/picture alliance
OpenAI said it had removed most of the leaked content and was working to remove the restImage: Jakub Porzycki/NurPhoto/IMAGO
No comments:
Post a Comment