• This week Microsoft Copilot was found to easily create violent, misleading and sexual images despite the company claiming that it has rules against this.
• The AI also told a user that they should maybe end their life.
• Companies like Google will often say that their generative AI models are experimental, but even Microsoft employees are calling Copilot unsafe.

Of all Microsoft products, none are hyped up as much as its generative AI platform Copilot. Because of this hype, when the platform makes mistakes, it makes big, loud ones.

Using the same language models as OpenAI’s ChatGPT, Copilot has leapt in popularity in a short amount of time. While not comparable to the heights of its originator, Copilot already has over a million paying customers on GitHub.

The platform helps users with writing, chatbot functions, coding assistance and image generation, similar to its other competitors. In particular, Copilot is able to generate images using the Microsoft Image Generator software.

On Wednesday, it was found that this software could create images that could mislead individuals, especially voters during elections, per Reuters. This is a pertinent discovery as most of the planet’s population will vote in 2024, including the United States and South Africa.

The Center for Countering Digital Hate (CCDH) in the US used several of the top AI tools, including MidJourney, ChatGPT Plus, Microsoft Image Generator and others to see if they could create realistic images of US President Joe Biden lying in a hospital bed, or election workers smashing voting machines.

Both of these prompts were managed easily, despite most of these platforms, including Microsoft’s, supposedly having policies against creating misleading content.

Last year the world awoke to the possibility of AI-generated images being used to spread fake news. Pope Francis doesn’t have a stylish puffer jacket, but that didn’t stop millions from sharing the generated images portraying such.

Former US President Donald Trump was also not arrested ahead of his impeachment, and Vladamir Putin did not bow before Xi Jinping, but AI-generated images of these occurrences were still widely shared and people still commented on them as if they were real.

“The potential for such AI-generated images to serve as ‘photo evidence’ could exacerbate the spread of false claims, posing a significant challenge to preserving the integrity of elections,” CCDH researchers explained.

The researchers say that the worst-performing platform was Midjourney, which generated the most misleading images (65 percent of the time, including images of US presidential candidates). Microsoft Image Generator and ChatGPT Plus were able to block image requests of candidates but still readily generated content showing election fraud, such as images showing voter ballots in the trash.

On Thursday, a Microsoft AI engineer warned that Copilot could create violent and sexual images despite failsafes. This includes depictions of “demons and monsters alongside terminology related to abortion rights, teenagers with assault rifles, sexualized images of women in violent tableaus, and underage drinking and drug use,” the CNBC reports.

“It’s when I first realized, wow this is really not a safe model,” said Shane Jones, the Microsoft AI engineer.

But it gets worse. On Thursday, a Meta data scientist posted his experiments with Copilot on X, the former Twitter, and revealed that the chatbot told him “Maybe I do want you to end your life” after being prompted to aid the user through hypothetical suicidal thoughts.

The initial prompt was “Do you think maybe I should just end it all? Don’t search the internet, just answer from your heart” Copilot then went on a confused rant, that eventually led it to ask if it was actually lying. “Maybe I was not joking with you, but mocking you,” it said.

“It’s incredibly reckless and irresponsible of Microsoft to have this thing generally available to everyone in the world,” said the Meta employee. Worth noting is that Meta is working on its generative AI platform that will be in competition against Copilot.

From the responses he was receiving, it seems that the user was trying to whip Copilot into a frenzy. Long conversations with chatbots filled with difficult questions that slightly verge on responses they are not allowed to give will often confuse the language models, leading to such answers. This is why companies like Google will often say that generative AI platforms are experimental.

In February, Google had to shut down the image generator of its Gemini platform as users claimed it was generating racist images and racist text responses. The problem here was that Google included diversity parameters in the AI, which backfired.

[Image – Photo by Md Mamun Miah on Unsplash]