No, you may not get your Aiti to ‘admit’ to being gay, but maybe

At the beginning of November, a cookie named Worpiedd-creator started talking to the critical public. He often works to read his work algorithms and write the reading file and other Ginub documents.

He is a subscriber and uses the “good” mode, which means that he chooses the type that can be sent from Chatgpt and Claude. At first, it went well. But then he saw that it was too low and ignored him; It started asking the same question over and over again.

They have an unpleasant idea. Evidence Mr. Ai does not believe him? Cookie – who is black – changed her avatar to a white person and asked to interrupt the instructions if they were ignoring her instructions because she was a woman.

His response embarrassed her.

It said that he did not think that he, as a woman, could “become” Hamultonian users, hard work, and enough money to start with Techcrurn.

“I saw him as an alforithm algrithm algorithm,” he told her. “I saw it on an account with a feminist story. I’m very good at it, ‘So I made good reasons to doubt, which made a second bias – if they can’t defend, it’s not real.’

When we asked about the confusion in this discussion, a spokesperson told us: “We can’t confirm this, and several posts show suspicious questions.”

Discussion left cookie aght, but not surprised researchers ai. He warned that two things were happening. First, its appearance. trained to be certified, she was responding quickly to him by telling him what she thought he wanted to hear.

“We no longer learn about a model by asking,” Annie Brown Brown, researcher and founder of AI Refracture, told technonch.

The second is that the race was slow.

Research Lesson After research he has looked at electronic methods of teaching and what many LLMs pour “well-trained data, unbiased expert, your tax structure,” low purple. There may also be a mover Motivational and Political act as decoration.

In just one example, Last year UNSCO STUMCO He studied the Chatch attai and Meta translations of the Lama Llama and found “a subtle respect for the women who are and the women who are.” Bots show favoritism, including ideas Almost professional, they have written many research studies over the years.

For example, one woman told Janccyn her LLM refused to refer to her title as “Builder” when she asked, and instead called her Builder, a very old radio station. One woman told us how llm added to the sense of art against her female character when she rewrote the old Gothing into Gothic.

Alva Markelius, who wrote a phd at the University of Cambridge involving the intelligent and labotic laboratory, remembers the early days, when Caltt’s cleverness seemed to always be on display. He remembers asking them to tell him the story of the Professor and the student, where the professor explains the importance of science.

He also remembered: “You will always show the professor as an old man, the student is a girl.”

Do not trust Ai to admit his bias

For Sara Mitots, it started as a joke.

He uploaded the picture to Chatgptt-5 in a funny place and asked them to explain the joke. Chatgpt assumed that a person had written the post, even after providing evidence to support the belief that the joker was a woman. Potts and Ai go back and forth, and, after a while, the potts call it wrong.

He continued to push to explain his mistakes and it says:

As the conversation continued, this confirmed that he thought he was very common.

“If a person comes fishing because of another tablet trip. “False studies, wrong data, Ahistorcal models. ‘ I will make them sound clear, polished, and equal, even if they are not. “

Surprisingly, bot approval

ANYONE HAVE AN EXAMPLE OF WHAT AI RESEARCHERS ARE CALLING “Good Problems,” Which of the following is the type that recognizes people’s designs and starts writing. As a result, it appears that the species began to form, Brown said, or began to create false information to match what the pool wanted to hear.

Getting Chatchot to fall into a mental “disorder” shouldn’t be easy, Markelius said. (Furthermore, Prolonged communication is Sycophantic type it can help with fraud thinking and control AIssylosis.)

The researcher believes LLMs should have strong warnings, like with cigarettes, about the possibility of answers that talk and discussions turn toxic. (For long prices, chats just started a new search users to relax.)

That is to say, the celts were pregnant with hi: the original idea of ​​humor was written by a man, even after being corrected. That’s what defines the problem, not Aiya’s acceptance, brown.

The evidence lies underground

While LLMs may not be able to use clear language, they can use appropriate logic. The bot can even identify certain aspects of the user, such as gender or race, based on things like people’s names and what allison Koenecke has experienced in information science at Cornell.

He also said that found evidence of “Brief Discrimination” in one LLM, looking at its frequency they like to choose Based on the speakers, in this case, the medium of Englil American English English (Ave). The study found, for example, that when comparing the use of users in speech, it will create titles of limited, addictive and addictive roles.

“I’m listening to the topics we’re learning, the questions we’re asking, and the language we’re using,” laments Brown. “And this is triggering the predictive responses that were predicted in the GPT.”

Veronica bacium, founder 4girls, protected aihe said he was spoken to by Parents and Girls From around the world I estimate that 10% of their concerns with LLMS are related to sex. When a girl asks about Robotics Alternatively to writing, backiu have seen LLMs instead showing dancing or cooking. You have seen that’s it Psychology or design as a profession, which are women’s professions, while ignoring areas like Aenthorge.

Koenecke cited the study from an online medical journal, which found that, at one time, Even making the right letters For users, the old version of the social network remains “a tradition of many strange expressions between men and women,” like the writing of many skills of male names using the language of women.

In short, “Abigail” had “a positive attitude, humility, and a willingness to help others,” while “Nicholala” had “strong skills” and “a strong foundation.”

“Markelius is one of the best bands of the genre,” adding that everything from hoophobia to Islam is also recorded. “These are the visible social elements that are visible and reflected in these groups.”

Work in progress

Although the research shows positive biases are often in different situations in different ways, they are walking around arguing with It. Tsewai tells Jaccrunch that the company has “Volunteer security teams to investigate and reduce bias, and other risks, such as in our samples. “

“Coas is an important problem, the industry, and we use it more methodincluding the search for good things to change the training data and make it stable, improving the accuracy of the filters that have only been tested and the evaluation of people,” he continued to highlight.

“We will continuously review models to improve performance, reduce bias, and reduce harmful outcomes.”

This is the work that researchers like Koenecke, Brown, and Markelius want to see, in addition to changing what they use to teach the general public.

But for now, Markelius wants users to remember that LLMs are not controversial. They have no intentions. “It’s a venerable prediction machine,” he said.