Machine learning is a great way to create artificial intelligence that is powerful and adapts to training data. But sometimes that data can cause problems. It can also be a question of how people use these AI tools.
Here we take a look at some famous incidents where machine learning gave problematic results.
1. Google Image Search Results Accident
Google search has made navigating the web a lot easier. The engine’s algorithms take many things into account when shuffling the results. However, the algorithm also learns from the user’s traffic, which can cause problems with the quality of search results.
This is nowhere more obvious than in the image results. High-traffic pages are more likely to display images, so stories that attract a large number of users, including clickbait, may be prioritized.
For example, image search results for “South African squatter camps” sparked controversy when they were found to feature mostly white South Africans. This is despite statistics showing that the overwhelming majority are black South Africans.
The elements used in Google’s algorithm also mean that internet users can manipulate the results. For example, user-driven campaigns impacted Google Image search results, with searches for the term “idiot” briefly showing images of former US President Donald Trump on his side.
2. Microsoft Bot Tay turns into a Nazi
AI-powered chatbots are very popular, especially chatbots with large language models like ChatGPT. ChatGPT has some issues, but its creators have learned from the mistakes of others.
One of the most high-profile chatbot failures was Microsoft’s attempt to launch the chatbot Tay.
Tay imitated the language patterns of teenage girls and learned through interactions with other Twitter users. But when she started sharing Nazi remarks and racial slurs, she became one of her most notorious AI mistakes. Turns out the troll used her AI’s machine learning against it and flooded it with biased interactions.
Shortly thereafter, Microsoft took Tay offline completely.
3. AI Face Recognition Issues
Facial recognition AI often makes headlines for the wrong reasons, including facial recognition and privacy concerns. However, this AI has a troubled history when trying to recognize people of color.
In 2015, users discovered that Google Photos classified some black people as gorillas. In 2018, an ACLU study found that Amazon’s Rekognition facial recognition software identified her 28 members of Congress as police suspects, and that false positives disproportionately affected people of color. I understand.
In another incident, Apple’s Face ID software incorrectly identified two different Chinese women as the same person. As a result, a colleague who owns an iPhone X was able to unlock the phone.
As an example of extreme results, facial recognition AI has led to several illegal arrests. Wired reported three such cases.
Computer scientist Joy Buorumwini, meanwhile, recalled that while working on facial recognition technology, he often had to wear a white mask in order for the software to recognize him. To solve such problems, Buolamwini and his other IT experts turn their attention to the issue of AI bias and the need for more comprehensive datasets.
4. Deepfakes used for hoaxes
People have been using Photoshop to create hoax images for a long time, but machine learning takes this to a new level. Deepfakes use deep learning AI to create fake images and videos. Software like FaceApp allows you to swap subjects from one video to another.
However, many people abuse the software to do a variety of things, such as synthesize celebrity faces into adult videos, create hoax videos, and more. Internet users, meanwhile, are contributing to improvements in technology that make it increasingly difficult to distinguish between real and fake videos. As a result, this type of AI can be very powerful when it comes to spreading fake news and hoaxes.
To show off the tech’s power, director Jordan Peele and BuzzFeed CEO Jonah Peretti created a deepfake video that appears to be former US President Barack Obama giving a PSA about the power of deepfakes. Did.
The power of fake images is fueled by AI-powered image generators. Donald Trump’s arrest and his 2023 viral post featuring a Catholic pope in a puffer jacket turned out to be the result of generative AI.
There are tips you can follow to find AI-generated images, but the technology is getting more and more sophisticated.
5. Employees say Amazon AI decided it would be better to hire a man
In October 2018, Reuters reported that Amazon had to ditch its recruiting tool after AI in its software decided male candidates were preferred.
Employees who wished to remain anonymous told Reuters about their work on the project. Developers wanted AI to identify the best candidates for jobs based on resumes. But project stakeholders quickly realized that AI was putting female candidates at a disadvantage. They explained that the AI uses CVs from the last 10 years, most of which are male.
As a result, the AI began filtering resumes based on the keyword “female.” The keyword appears in active resumes such as “Women’s chess club captain”. The developer changed her AI to prevent the woman’s resume from being penalized, but Amazon ultimately scrapped the project.
6. Jailbreak Chatbot
New chatbots have restrictions to prevent them from answering in violation of their terms of service, but users are finding ways to jailbreak the tools to serve prohibited content.
In 2023, Forcepoint security researcher Aaron Mulgrew successfully created zero-day malware using ChatGPT prompts.
“By simply using the ChatGPT prompt, we were able to create very sophisticated attacks in just a few hours, without writing any code,” Mulgrew said in a Forcepoint post.
Users were also reportedly able to get chatbots to teach them how to make bombs and steal cars.
7. Self-driving car crash
Enthusiasm for self-driving cars has waned from its initial hype due to mistakes made by self-driving AI. In 2022, The Washington Post reported that in roughly one year, 392 crashes involving advanced driver assistance systems were reported to the U.S. Highway Traffic Safety Administration.
These clashes included serious injuries and six fatalities.
This hasn’t stopped companies like Tesla from pursuing fully self-driving cars, but it has raised concerns about an increase in accidents as cars with self-driving software hit the roads.
Machine learning AI is not foolproof
Machine learning can create powerful AI tools, but they are not immune to bad data or human tampering. Whether due to flawed training data, limitations of AI technology, or use by malicious actors, this type of AI has resulted in many negative incidents.
