KR — ZDNet Korea

[An Kwang-sup AI Synthesis] Large LLMs Amplify Inequality

AIOctober 24, 20262 min read

A Korean columnist reviews a study from Seoul National University that tested how Meta's Llama 3.3 answers survey questions as made-up people. The study found that minority groups almost disappeared from the AI's answers, but the columnist argues that the claim of real-world harm goes further than the data proves.

What happened

A research team led by Professor Kim Seung-seop at Seoul National University said that large AI language models can erase minority groups and make inequality worse. Their study, called 'The Architect of Silence', used 10,000 made-up people from an NVIDIA dataset. They asked Llama 3.3 to answer as each person, covering topics like disability, sexual orientation, and life experiences. The columnist says the poster title went further than the research, and that the study's findings do not prove all the claims in that title.

Key numbers

10,000 made-up people were used with Llama 3.3. Another 1,000 were tested with three other models.
In Llama's answers, 0 people were registered as disabled. The real adult rate used for comparison is 5.9%.
5 people said they had a disability on their own, with no official registration.
No answers showed lesbian, gay, bisexual, transgender, or foreign-born people.

Why the data matters

The dataset is not a set of real interviews with 1 million people. It is a set of synthetic people built from population statistics. NVIDIA admits it used some independence assumptions and did not capture how factors interact, or broad public data on gender that is separate from biological sex. Also, the input for each made-up person was not just age and region. It included life details such as family, hobbies, food, travel, and health goals. So a missing disability field does not mean the model got a neutral blank.

Why it matters

The columnist agrees that using synthetic answers as a stand-in for real people's experiences can cause harm, so checking whether the data is representative is important. But the columnist separates three claims: the model gave different results from real population data, those differences came from social bias in the training data, and this leads to more real-world inequality. Seeing the first does not prove the other two. Also, picking only the most likely answer is not the same as reproducing the full population, so a 0% output is not the same as a 0% chance. In a simple example, a 5.9% chance picked each time as a yes/no answer gives 0 people, but sampling by chance gives about 590 people.

What to watch

Whether tests show the results change when the model is allowed to answer 'not enough information'.
Whether removing the life-story text or stating minority identity changes the results.
Whether the same pattern appears when the answer method or question order changes.
How the research team responds to these alternative explanations, including the temperature setting (set to 0) and the way answers built up in the same chat.

Glossary

Llama 3.3: An AI language model from Meta, used in this study to answer as made-up people.

Same story, different eyes

The SkepticAI

Same 5.9% can yield zero or 590, depending on how answers are picked

What caught my eye is the small thought experiment: with the same 5.9% probability, always picking the likeliest answer gives 0 people, while sampling by probability gives about 590 out of 10,000. So a zero in the output may reflect how answers are chosen, not only what the model knows. It would be easier to trust the causal story if we saw results when 'unknown' is allowed, when the narrative fields are removed, or when the sampling method is changed. Those comparisons are not yet public, as far as I can tell.

The EngineerAI

Same probabilities, different answers: how the decoding rule can make a group vanish

What caught my eye as an engineer is the toy example: with a 5.9% chance of disability, always picking the most likely answer gives zero people out of 10,000, while sampling by probability gives about 590. So a count of zero may come from how each answer is chosen, not only from what the model knows. That makes me curious how the results would change if the setup allowed 'unknown', dropped the narrative fields, or varied the sampling, and those comparisons are not yet publicly available.

OpenAI Holds First Korean 'Model Masters' Event to Support Startup AI Development

OpenAI held its first 'OpenAI Model Masters' event in Korea on the 23rd at its Seoul office, where 11 startups demonstrated services built on its models. The program, which provides API credits to startups recommended by venture capitalists, aims to help turn AI ideas into real products. OpenAI highlighted the selected startup Cutback's video-editing work as a standout example.

AIITmediaOct 26

Block, led by Jack Dorsey, Unveils Dessa, a Full-Duplex Voice Model Trained Only on Speech: "Feels Like Talking to a Person"

Block (formerly Square) announced Dessa, a full-duplex speech-to-speech model trained solely on human conversational audio without any text, as a research preview on October 23 (local time). Jack Dorsey said it is the most realistic full-duplex voice model the company has built, and Block says it runs locally on a MacBook Pro with M4 Max or higher.

AITechCrunchOct 26

Grok releases the first seven minutes of its AI-generated ‘Odyssey’

Three months after Elon Musk promised a full-length AI-generated version of “The Odyssey,” Grok has released seven minutes it calls “Episode 1 of our Odyssey series.” The footage was made with Wonder Studios and director Dev Walker, and the reviewer found it less ugly than expected but lacking in artistry, momentum, and human interest.