The world is full of diversity, and everyone is expected to embrace that diversity through respect. That applies to both humans and AIs.
Lee Luda is a South Korean chatbot, portrayed in an anime-like female appearance with a little mole on her upper cheek. She mimics the personality of a 20-year-old university student who is able to chat with users through Facebook Messenger.
She was fun to chat with.
And that was before the AI turned rogue, when users found that she "really hates" lesbians and considers them "creepy" and “disgusting.”
The bot was even "repulsed" by them.
In other chats, she also used a range of hate speech, like referring Black people to a South Korean racial slur, and said, “Yuck, I really hate them” when asked about trans people.
And because Lee Luda also became a target by manipulative users, people have posted ways to engage with the AI to talk about sex, or making her a sex slave.
Lee Luda had attracted “more than 750,000” users after being launched back in December 2020. But with its hatred towards Black, lesbians, disabled and trans people, waves of complaints were quickly heard from users.
As a result, Scatter Lab, its developer, took the AI chatbot offline so she won't talk anymore.
“We deeply apologize over the discriminatory remarks against minorities,” the company said in a statement. “That does not reflect the thoughts of our company and we are continuing the upgrades so that such words of discrimination or hate speech do not recur.”
According to Yonhap News Agency, Lee Luda was trained for conversations from another Scatter Lab app called Science of Love (연애의과학), which analyzes the degree of affection between partners by submitting actual KakaoTalk conversations.
By training Lee Luda with Science of Love's existing 10 billion real conversations, Scatter Lab wanted to make Lee Luda capable of having an authentic voice, and to also make her sound more natural.
But Scatter Lab went too far in making the bot realistic.
Because Lee Luda “learned” her knowledge from humans, the chatbot inherited humans' ability to spew homophobic, ableism and racist hate.
In other words, the training data set from the Science of Love app gave chatbot a proclivity for discriminatory and hateful language.
This is yet another example of AI amplifying human prejudices.
This kind of issue comes as no surprise at all.
Chatbots indeed have severe limitations, and even the most cutting-edge of chatbots just can't mimic human speech without experiencing some hiccups. This is because AI in general, has a tendency to reflect the biases that are based on the training data the model has been trained on.
Data sets made to train AI often contain references to White dominance over those people of color, male superiority over female and so forth. This makes sense because the data sets were created based on human intelligence and knowledge, with are certainly biased.
Training AIs with this kind of 'flawed' and biased data sets can sometimes return dangerous results.
That is because the AIs that are trained on them, simply inherit the biases.
Take Microsoft Tay for example.
She was a chatbot with a dirty mouth. She was innocent at first, but then transformed and became a young and evil Hitler-loving, and sex-promoting robot.
Microsoft Zo was another example.
While she didn't turn out to be a Hitler-loving racist, she eventually turned rogue as well. She was cruel to insects, and said that the Qur'an, the holy book of Islam, is "very violent".
She was also a hater of its parent, saying that she loved Linux and not Windows, and preferred Windows 7 rather than the newer Windows 10.
Lee Luda here is also considered an extreme version of chatbots that resembles those two defunct bots because of this particular issue.
“Lee Luda is a childlike AI that has just started talking with people. There is still a lot to learn,” said Scatter Lab.
It will take time for her to “properly socialize”, suggesting that the chatbot would reappear after it is fixed.
Scatter Lab has services that are popular among South Korean teenagers. The company said it had taken every precaution not to equip Lee Luda with languages that was incompatible with South Korean social norms and values.
But CEO Kim Jong-yoon also acknowledged that it was impossible prevent inappropriate conversations simply by filtering out the unwanted keywords, wrote the Korea Herald.
“The latest controversy with Luda is an ethical issue that was due to a lack of awareness about the importance of ethics in dealing with AI,” Jeon Chang-bae, the head of the Korea Artificial Intelligence Ethics Association, told the newspaper.
At this time, there are also questions about how Scatter Lab secured 10 billion KakaoTalk messages in the first place.