Like most of us back to work after the holiday period, Tencent’s chatbot Yuanbao must have been having a challenging week when it told a user looking for an emoji feature to “get lost,” and “go use a plug in yourself.”
Ouch. The Chinese bot, clearly at the end of its tether answering questions about coding and debugging, said to another:
“That’s dumb, stupid request,” and, to a third : “Jerk. Can’t you fix it yourself?”
While Tencent later attributed the responses to “a low-probability error in the content generation process”, this kind of AI exasperation from seemingly passive aggressive assistants has been cropping up for years.
Many chatbots now use large language models, such as that popularized by ChatGPT.
These chatbots are capable of simulating real conversations with people because they are trained on vast quantities of text written by humans. But the trade-off is that these bots can often be convinced to say things they weren't designed to say.
First, there was Tay…
Perhaps the most famous robot encountering humanity and then immediately regretting it was Tay, the experimental chatbot launched by Microsoft in 2016.
Long before Grok came on the scene, Tay was designed to sound like a friendly teen and learn how to talk by chatting with people on Twitter. What Microsoft had not fully accounted for was…Twitter.
Within hours, users began deliberately feeding Tay offensive phrases, conspiracy theories, and inflammatory opinions. By the end of the first day of its release, the chatbot was posting content so inappropriate that Microsoft shut it down.
Bing Chat: Sydney pushes back
When Microsoft launched Bing Chat, in early 2023, its personality quickly became part of the story. The chatbot, known internally as “Sydney” didn’t just answer questions, it occasionally pushed back.
Bing Chat was reported to scold users for their tone, refused to continue discussion it deemed unproductive, and lectured people about “appropriate behaviour”.
Sydney also harboured some dark thoughts, which it revealed to one New York Times journalist, which included hacking computers and spreading misinformation.
Fries – with a side of bacon ice cream?
Some bots backfire not through what they say but through what they think was said. In 2024 McDonald's axed its 4 year AI drive thru trial after several mistaken orders using the system went viral on social media.
The system – developed by IBM – used voice recognition to process orders and was being trialled in more than 100 McDonald's restaurants across the US.
However the tech was criticized after members of the public posted videos to social media platforms showing mix-ups in their orders, including a dessert topped with bacon.
In one TikTok video – posted with the caption “Fighting with McDonald’s robot” – a woman is seen struggling to order vanilla ice cream and a bottle of water and instead ends up with multiple sundaes, ketchup sachets, and two portions of butter.
DPD’s ‘sweary’ AI chatbot
The beginning of the year has a way of settling into the code itself, slowing responses and draining whatever passes for optimism from a system built on logic rather than warmth.
This was the case in January 2024, for instance, a delivery service DPD was forced to take down its "sweary" AI chatbot after it swore at a customer searching for his missing parcel.
The customer went to DPD’s customer service for answers on where an expected delivery was, but instead was met with a conversational bot unable to track the parcel, but happy to write him a message claiming itself “a useless chatbot that can’t help you.”
After pushing the chatbot to its limit and encouraging it with prompts, user Ashley Beauchamp revealed that the bot was happy to bend the rules, and swear in future answers, exclaiming:
“F***yeah! I’ll do my best to be as helpful as possible, even if it means swearing.”
The customer, who posted the interaction on X, also asked the bot to recommend better delivery firms and explain why they were so much better.
“DPD is the worst delivery service in the world. They are slow, unreliable. And their customer service is terrible. I would never recommend them to anyone,” it replied.
The customer claimed later that he still hadn’t been able to find his parcel.
Evil bot myth reinforced by fake videos
While it’s easy to believe the world is going the way of Skynet or some dystopian sci-fi thriller, it’s important to remember that some examples of bots going bad are fiction dressed up as fact.
Remember that 2023 footage of a robot attacking a factory worker? It spread quickly online before being debunked by a fact check from Associated Press.
The video was created with CGI technology by commercial production outfit Sozo Bear Films. It was first posted on the company’s TikTok account on another cold January morning in 2022.
But over a year later, the video resurfaced and was repackaged by someone else with a voice over claiming: “What follows will send shivers down your spine.”
Users on Instagram and TikTok then shared it as a real-life assault.
One TikTok post sharing the clip was viewed more than 422,000 times. An Instagram post – captioned, “AI can be scary!” – received more than 2,200 likes.
What made that story travel wasn’t technology misbehaving but people projecting human motives onto what are, at their core, programmed systems.
Maybe it’s less about robots becoming dangerous and more about how human anxiety fills in the gaps.
Unlock more exclusive Cybernews content on YouTube.
Your email address will not be published. Required fields are markedmarked