“Hey, @grok, is it true?” Ever since Elon Musk’s XAI launched its generous artificial intelligence chatbot groke in November 2023, and it was rolled out for all non-Permina users, especially in December 2024, thousands of X (East Twitter) users asks BEE the question that they should check the fact on the information seen on the stage.
A recent survey conducted by a British online technology publication Tekardar It was found that 27% of Americans used Artificial Intelligence Tools Search as apps such as Openi’s chat, Meta’s Meta AI, Google’s Gemini, Microsoft’s Kopilot or Google or Yahoo, instead of Perplatity instead of engines.
But how accurate and reliable are chatbots’ reactions? Many people have Arsked this question in front of Groke’s recent statements about ‘white massacre’ in South Africa. In addition to the problematic stand of the grouke on the subject, the X users were irritated about the fact that Bot begins to talk about it when it was asked about full topics, such as in the following example:
Discussions arose around the alleged “white genocide” after the Trump administration brought white South African people to the United States as refugees. Trump said that he was facing a “genocide” in his motherland – an allegation that lashes any evidence and many are related to racist Conspiracy myth of “great replacement”,
XAI blamed the “unauthorized amendment” for the passion of the grouke with the theme “white genocide”, and said “fully invested.” But do such flaws occur regularly? Can users make sure to get reliable information when they want to do something with AI?
We analyzed the questions of this and answer for you in checking this DW fact.
Study reflects factual errors and converted quotes
Two studies conducted by the TO Center for Digital Journalism in the British Public Broadcaster BBC and the United States have found significant drawbacks this year, when it comes to the ability to accurately conveyor news reporting for the ability of generative AI chatbots.
In February, A BBC study It was found that “the answer produced by AI assistants included significant impurities and perverted materials manufactured by the organization”.
When it is asked to answer questions about the current news using BBC articles about the current news about chat, Copilot, Gemini and Perplexity, it was found that 51% were “important issues of some forms” in response to 51% chatbots.
In 19% answers, it was found that they had added their factual errors, while 13% of quotes were replaced or not present in the quoted articles.
PT Archer, director of the BBC’s generic AI program concluded, “AI assistants cannot be trusted to provide accurate news to the current and they risk misleading the audience.”
AI provides wrong answers with ‘dangerous confidence’
Similarly, Posted in Research by TO Center for Digital Journalism, Columbia Journalism Review (CJR) In March, it was found that eight generative AI search equipment in 60% of cases articles correctly identified the perfection of exerps.
Perplexity performed the best with a failure rate of “only” 37%, while Groke incorrectly presented 94% questions.
CJR said it was particularly concerned with “dangerous confidence”, with which AI devices presented the wrong answers, reporting: “Chattgt incorrectly identified 134 articles, but indicated a lack of confidence, Juston Times two hundred [total] Reactions, and never refused to reject to respond. ,
Overall, the study found that the chatbots “were normally bad to answer questions they give correct answers, instead offering incorrect or speculative answers” and that AI search equipment “cited the remaining and copied versions of the remaining and copied articles.”
AI chatbott is only good as their ‘diet’
And where does AI get their information? It is fed From different sources Like broad databases and web discovers. Depending on how AI chatbots are trained and programmed, the quality and accuracy of their answers may be different.
,One issue that has emerged recently is LLM pollution [Editorial note: Large Language Models] By Russian disintegration and promotionSo sLearly there is an issue with the ‘diet’ of LLMS, “Tomaso Caneta told DW. Male And the fact check the coordinator European Digital Media Observatory (Edmo),
“If the sources are reliable and qualitative, the answer wants the most likely to the same child.” Kineta explains that he regularly comes into the reactions that are “incomplete, not accurate, misleading or even false.”
In the case of XAI and Groke, whose owner, Elon Musk, is a fierce supporter of US President Donald Trump, a clear danger that “diet” can be controlled politically.
When AI finds it all wrong
In April 2024, Meta AI allegedly posted on Facebook in the New York Parenting Group that a handicapped in it was still academicly talented children and advised on special schooling.
Finally, Chatbot apologized and admitted that it did not have “personal experiences or children”, while Meta said 404mediaWhich was reported on the incident:
“This is a new technique and it can not always refund the response that we intend, which is the same for all the genetic AI systems. Since we launched, we have continuously released updates and reforms in our model and we continue to make them.”
In the same month, Groke misinterpreted a viral joke about a poor performing basketball player and Told users in its trending section He was investigating by the police following the allegations of the bricks of homes with bricks at Sacramento, California.
Groke misunderstood the common basketball expression, by which a player who failed to achieve any of his throwing on the target is said to “throw bricks”.
Other mistakes are less entertaining. In August 2024, Groke spread misinformation about the deadline for US presidential candidates for adding ballot papers to nine federal states after a return from the race of former President Joe Biden.
One in Public letter to Elon MuskState Secretary of Minnesota, Steve Simon, wrote that, within a few hours of President Biden’s announcement, Groke made false headlines that Vice President Kamla Harris would be disqualified to appear on the ballot in many states – untrue information.
Grake provides the same AI image for various real events
It is not just news that difficulties appear with AI chatbots; So they demonstrate severe boundaries when it comes to identifying AI -generated images.
In a quick experiment, DW asked the grouke to identify the date, location and origin of the AI of the fire in a destroyed aircraft hangar taken from a tikok video. In its response and explanation, Groke claimed that the image showed several different incidents at many different places, which range from a small airspace in Calisbury to Denver International Airport in Colorado, Tan Nut International Airport in Ho Chi Minh City in Vietnam.
In recent years, thesis places have really accessible and fire, but in question the image has not shown any of them – we strongly believe that it was generated by artificial intelligence, which Grocke seemed to recognize, despite clear errors and discrepancies in the image – including inhuman jets from the wings of the tail and fire on the airplane.
More than that, Groke recognized the part of the “tiktok” watermark that appeared in the corner of the image and suggested the fact that “its authenticity supported.” In contrast, the “more details” tab under it, Groke said what Tikkok is “a platform used for rapid spread of viral materials, which could cause misinformation when not properly verified.”
Similarly, this week, bus, Groke informed X users (Portuguese) That a viral video to show a huge anaconda in Amazon, which measures several hundred matters (more than 500 feet) in a long time, is real – despite this it is clearly generated by artificial intelligence, and even to identify a chatting watermark.
AI chatbots should not be seen as a fact-zanch tool ‘
AI Chatbot may appear as a omniscient unit, but they are not. They make mistakes, misunderstand things and even manipulation can be done. PostDoctoral Research Fellow in Felix Simon, AI and Digital News and Research Associates Oxford Internet Institute (OII),
For Canata, AI chatbot can be useful for very simple facts for investigation. But he advises not to rely completely on the subject. Both experts say that users should always have Responses with other sources.
Daniel Ebertz contributed to this report.
Edited by: Ines Eisle, Rachel Baig
