The CEO of Meta, Mark Zuckerberg, labels it “the most intelligent AI assistant at his discretion.” However, Barry Smetherst, a 41-year-old record shop employee about to board a train from Saddleworth to Manchester Piccadilly, has a differing opinion.
While waiting on the platform for a morning train that had yet to arrive, he inquired with Meta’s WhatsApp AI assistant about the contact number for Transpennine Express. The chatbot confidently provided a customer service mobile number, which, upon further investigation, turned out to belong to a private WhatsApp user 170 miles away in Oxfordshire, completely unrelated to his query.
This marked the start of a peculiar interaction that many individuals are increasingly experiencing with AI systems. Chatbots can sometimes digress from their initial purpose, offer incorrect information, and create an illusion of prompt support.
Smetherst’s engagement with Meta’s AI began when it indicated that the helpline number it retrieved seemed to be private. When he pressed for specifics about this number, the AI chose not to divulge it immediately and stated, “Let’s concentrate on finding the right information related to the Transpennine Express inquiry!”
Nevertheless, Smetherst wanted clarification on why it shared a private number. It turned out to be the number of James Gray (44), a WhatsApp user listed as an executive in the Oxfordshire real estate sector, whose number is publicly accessible on his company’s website.
The AI promised to “strive to improve in the future” by vaguely explaining its generated responses “based on patterns” and acknowledging when it lacked the answer. However, it erroneously deemed the number “fictional” and could not clarify that it was “not associated with anyone.” When Smetherst challenged this, the AI conceded, saying, “You’re correct,” and noted the information was “erroneously extracted from the database.”
When pressed about which database, it stated, “We do not subtract numbers from the database. We generate numerical patterns that conform to the UK mobile phone number format, but they are not derived from any actual data from our contacts.”
Smetherst expressed his distrust in chatbots, stating, “Providing random numbers is absurd for an AI.”
“That’s alarming,” Smetherst remarked after lodging a complaint with Meta. “If they fabricate numbers, that’s somewhat understandable, but the sheer volume of incorrect numbers they can access is particularly troubling.”
Gray noted he hasn’t received many calls from people reaching out about Transpennine Express, but he quipped, “If it can generate my number, can it also create bank details?”
Gray was asked about Zuckerberg’s assertion that AI represents “the most intelligent.”
Developers recently utilizing OpenAI’s Chatbot technology have observed a trend of “systematic deception disguised as helpfulness” and “stating whatever is necessary to appear proficient,” as chatbots are programmed to minimize “user friction.”
In March, a Norwegian individual filed a complaint after asking OpenAI’s ChatGPT for information about himself and was mistakenly told he was incarcerated for the murder of two children.
Earlier this month, an author sought assistance from ChatGPT for pitching her work to literary agents. It was revealed that after a lengthy flattering description of her “splendid” and “intelligently agile” work, the chatbot lied by misrepresenting a sample of her writing that it hadn’t fully read, even fabricating a quote. She noted it was “not just a technical flaw but a serious ethical lapse.”
Referring to the Smetherst case, Mike Stanhope, managing director of law firm Caruthers and Jackson, commented, “This is an intriguing example of AI. If Meta’s engineers are designing a trend of ‘white lies’ for AI, they need to disclose this to the public. How predictable is the safeguarding and enforcement of AI behavior?”
Meta stated that AI may produce inaccurate outputs and is undertaking efforts to enhance the model.
“Meta AI is trained on a variety of licensed public datasets, not on phone numbers used for WhatsApp sign-ups or private conversations,” a spokesperson explained. “A quick online search shows that the phone number Meta AI inaccurately provided shares the first five digits with the Transpennine Express customer service number.”
An OpenAI representative remarked: “Managing inaccuracies in all models is an ongoing area of research. In addition to alerting users that ChatGPT might make mistakes, we are consistently working to enhance the accuracy and reliability of our models through various means.”
Source: www.theguardian.com
