If in your business, they bet everything for chatbots, we have bad news: they lie. More than 60% of the responses they offer are wrong, although the machines say it with total security. If they question them too much, they change their version in flight, and if they question them too much, they destroy. The promise of an AI that simplifies access to information collided with a more gray reality.
From the journalism Center of Columbia journal published a study by comparing eight IA engines. Klaudia Jaźwińska and Aisvarya Chandrasekar have developed a strict methodology and have reached several conclusions, all rather dark: chatbots are not good to “refuse to answer questions that could not answer precisely and offered in incorrect or speculative place”. The most ironic is that Premium chatbots have given incorrect responses more safely than free models. That is to say to lie better.
In all cases, the majority avoided the preferences of the robot exclusion protocol, copied articles and linked without “precise citation guarantee in the responses”. AND Which one who lies? The least used seems to be.
Homepod Siri analysis and the best sound are already at home
Grok 3, Elon Musk's Ai, is the least reliable of all
For the study, they use, they say, to the following methodology: they analyzed the precision of several chatbots when identifying information in press articles, by evaluating them in three aspects, Find the right article, identify the appropriate editorial and provide the precise URL. To do this, they selected ten articles of twenty different publishers, extracted from text fragments and made a total of 1,600 consultations on eight different chatbots. The results were evaluated manually and classified according to their degree of precision, from “correct” to “completely incorrect”.
The results revealed that more than 60% of the responses were incorrect, with notable differences between platforms: Perplexity failed in 37% of consultations, while GROK 3 reaches an alarming error of 94%. The fastest … The fastest inventions. The chatbots not only gave bad answers, but they did it with worrying security, without resorting to prudent sentences or recognizing their limits. The worst part is transported by premium models, such as Perplexity Pro and Grok 3, with a higher error rate than their free versions due to their tendency to offer definitive responses even when they did not have reliable data.
Another worrying aspect was the apparent violation of directives Robots.txtwhich allow publishers to block automatic trackers. Although chatbots should not access the blocked content, some and perplexity – the performance of which is to identify the items they should not have viewed, raise suspicion regarding compliance with these standards.
And, as we said at the beginning, chatbots also tended to cite incorrect sources and, in many cases, provided that broken or invented links. This shows that the license agreements between AI and media companies do not guarantee greater precision. Even in cases where there is a formal collaboration, Like over time and San Francisco ChronicleChatbots have given inconsistent results. Perplexity fails time payment walls and Los Angeles Times as if they were sidewalks.
Apple was forced to keep up with a rhythm that I didn't want to
We said a few days ago that Apple had a problem, not to go behind his competitors, but to stain his credibility with promises he had to postpone 2026, like the new Siri. But there is an appreciation that I want us to keep in mind: Maybe Apple does not arrive late at the career of Modal ChatbotsBut his competition still sells a dish without cooking.
Those who arrived have serious holes. According to the latest NewSecure study, the Deepseek service, thus applauded and promoted by its speed, has serious security gaps in terms of data protection and processing. Claude, for example, is stellar in hers, but is not integrated into complex tools. Copilot was promoted as something futuristic, but he turned out to be a normal and current chatpt. Gemini progress slow but of course, although every day I use it less because it makes me more answers “I can't do this” or “I'm not scheduled for this task”.
Advanced language models, which are already present as an additional layer of conviviality in several basic IT processes. It is clear that Apple Intelligence still has a long way to go, but even the famous perplexity the AI is well stopped in this exercise, which shows that a few years ago, for everyone, for everyone, Until a certain balance, sustainability and informative reliability.
In Applesfera | Use China Depseek safely in Mac: this is the only way to avoid China data
In Applesfera | When Apple Intelligence is available in Spanish and USA: everything we know
