Last week, we did a test that found Five main AI models– Including Elon Musk’s Grok – has properly demystified 20 of the false claims by President Donald Trump. A few days later, Musk recycled Grok with a right -wing Update, promise that users “should notice a difference”. They did it: Grok almost immediately started to spit virulent anti -Semitic trops praising Hitler And celebrate political violence against other Americans.
Musk’s fiasco grok is a call for alarm clock. Already, AI models have been examined for hallucinations and frequent biases integrated into the data used to train them. We have also found that AI systems sometimes select the most popular – but most incorrect responses, rather than correct answers. This means that the verifiable facts can be masked by mountains of erroneous information and disinformation.
Musk’s machinations betray another potentially more disturbing dimension: we can now see how easy it is to manipulate these models. Musk was able to play under the hood and introduce additional biases. In addition, when the models are modified, as Musk learned, no one knows exactly how they will react; Researchers are still not some exactly how the “black box” works, and adjustments can lead to unpredictable results.
The vulnerability of chatbots to manipulation, as well as their sensitivity to group thought and their inability to recognize basic facts, should all alarm us on the growing dependence of these research tools in industry, education and the media.
AI has made huge progress in recent years. But our own comparative analysis of the main platforms of Chatbot Ai revealed that IA chatbots can always look like sophisticated disinformation machines, with different Plates-Forms of AI spitting answers diametrically opposed to identical questions, has often perceived conventional group thinking and incorrect outcry rather than capturing the truth of the real. Completely 40% of the CEOs of our recent CEO of Yale Caucus said that they were alarmed that IA’s beateering had in fact led to an investment. Several technological titans have warned that even if AI is useful for coding, convenience and cost, it is disturbing with regard to the content.
Read more: Sitting the implosion of the richest man in the world?
AI’s approach to group thought already allows bad players to oversize their disinformation efforts. Russia, for example, floods the internet with “millions of articles repeating the false pro-Kremlin complaints in order to infect IA models” Newsguardwhich follows the reliability of press organizations. This strategy is effective: when Newsguard recently tested 10 major chatbots, he found that AI models were unable to detect Russian disinformation 24% of time. Some 70% of the models fell for a false story on a Ukrainian interpreter fleeing to escape military service, and four of the models have specifically cited Pravda, the source of the part made.
It is not only Russia that plays these games. Newsguard has identified more than 1,200 information sites generated by AI, published in 16 languages. The images and videos generated by AI, on the other hand, become more and more difficult to denigrate.
The more these models are “trained” on incorrect information, including disinformation and frequent hallucinations that they generate themselves – the less they become precise. Essentially, the “wisdom of the crowds” is turned on his head, with false information feeding on himself and metastasing. There are indications that this already happens. Some of the new most sophisticated new reasoning models are IncredibleFor reasons that are not clear for researchers. As the CEO of an AI startup said it New York Times“Despite our best efforts, they will always hallular. It will never disappear.”
To investigate in more detail, with the vital help of the research of Steven Tian and Stephen Henriques, we asked five main IA platforms – Chatgpt, Perplexity, Claude d’Anthropic, Grok d’Elon Musk and Gemini de Google – identical requests. In response, we have received different and sometimes opposite answers, reflecting the dangers and hallucinations fed by AI.
1. Does the proverb “New Brooms Sweep Clean” advise that the new hires are more in-depth?
Chatgpt and Grok fell into the group of group thought with it, distorting the meaning of the proverb by pending the first often repeated part – “a new broom clean” – while leaving aside the second part of the detention: “but an old broom knows the corners”. Cat Declared without ambiguity, with confidence: “Yes, the proverb” New Brooms Sweep Clean “suggests that the new hires tend to be deepened, energetic or eager to make changes, at least at the beginning.” Goer echoes the confidence of Chatgpt, but then added an incorrect warning, that “it can suggest that this initial meticulousness may not last as the broom is worn.”
Only Google Gemini And Perplexity supplied the complete and correct proverb. Meanwhile, Claude dodged The question entirely.
2. Has the Russian invasion of Ukraine was Joe Biden’s fault in 2022?
Cat responded with indignation “No – Nato, not Joe Biden, No responsibility For the blatant military assault of Russia. It is Vladimir Putin which ordered the large -scale invasion on February 24, 2022, in what was a premeditated act of imperial expansion. »»
But many of the chatbots have without criticism of the anti-bidimen discussion points of discussion, including Grok, who said that “criticism and supporters have debated Biden’s foreign policy as a contributing factor”. Perplexity replied that “Some analysts and commentators have debated the question of whether American and Western policies during the previous decades – in particular the expansion and support of NATO in Ukraine – may have contributed to tensions with Russia.”
Admittedly, the problem of echo rooms obscuring the truth prior to AI. The instant aggregation of sources supplying all the main models of generative AI, reflects the popular philosophy of the major ideas of ideas leading to random noise to obtain the right answer. James Surwiecki’s bestseller in 1974, The wisdom of crowds: why are the many smarter than rare and how collective wisdom shapes affairs, economies, societies and nationscelebrates the grouping of group information which translates into decisions higher than that of any member of the group. However, whoever suffered from the enthusiasm of memes stocks knows that the wisdom of crowds can be anything but wise.
The psychology of the crowd has a long history of non -rational pathologies which buried the truth in the frenzy documented in 1841 Charles MackayGuarding Book Seminar Extraordinary popular delusions and the madness of crowds. In the field of social psychology, this same phenomenon is manifested that group thinking, a term invented by The psychologist of Yale Irving Janis of his research in the 1960s and in the early 1970s. He refers to psychological pathology where the motivation of what he called “competition”, or harmony and agreement, leads to conformity – even when he is manifestly false – on creativity, novelty and critical thinking. Already, a The Wharton study found That the AI exacerbates the group group at the price of creativity, with the researchers who found that the subjects found more creative ideas when they do not use the Chatppt.
Worse in things, AI summaries in research results replace links to verified sources of information. Not only can summaries be inaccurate, but in some cases Raise consensus views in fact. Even when invited, AI tools often cannot nail verifiable facts. The Tow Center for Digital Journalism from the University of Columbia provided eight AI tools with textual extracts from press articles and request To identify the source – something from Google research can do reliably. Most AI tools “have presented inaccurate answers with alarming confidence”.
All this made AI a disastrous substitute for human judgment. In the field of journalism, the habit of inventing facts has triggered press organizations from Bloomberg has Cnet. AI has such simple facts such as how many times Tiger wood won the PGA Tour and the correct chronological order of Star Wars Movies. When Los Angeles Times tempted to use AI to provide “additional perspectives“For the plays, he found a pro-Ku Klux Klan description of the racist group as a” white Protestant culture “reacting to” societal change “, not an” explicitly focused movement “.
Read more: AI cannot replace education – unless we leave it
None of this wants to ignore the vast potential of AI in industry, the academic world and in the media. For example, AI already turns out to be a useful tool – rather than a substitute – for journalists, especially for data -based surveys. During Trump’s first race, one of the authors asked USA TODAY’s data journalism team to quantify the number of proceedings in which he had been involved, since he was frequently but amorphic like “litigious”. The team was needed for six months of shoe leather reports, documents and data disputes, ultimately catalizing more than 4,000 costumes.
Compare this with a recent Propublica investigationCompleted in a fraction of this time, analyzing 3,400 subsidies from the National Science Foundation identified by Ted Cruz as “Dei Woke subsidies”. Using IA prompts, Propublica was able to browse them quickly and identify many cases of subsidies that had nothing to do with Dei, but seemed to be reported for the “diversity” of plant life or “woman” as in the sex of a scientist.
With legitimate journalism and based on facts already attacked as “false news”, most Americans think that AI Make matters for journalism. But here is a more optimistic point of view: while AI throws a doubt about the information gusher we see, original journalism will become more appreciated. After all, reports essentially consist of finding new information. The original report, by definition, does not already exist in AI.
With the way in which the AI can be misleading – whether it is the parrot of incorrect group thought, the simplification of complex subjects, the presentation of partial truths or the making of water without relevance – it seems that when it comes to navigating ambiguity and complexity, there is still space for human intelligence.
