Google AI Gemini parrots China's propaganda | Hacker News Reader

Google AI Gemini parrots China's propaganda

voanews.com

31 points2OEH8eoCRo08 comments

> VOA’s Mandarin Service recently took Google’s artificial intelligence assistant Gemini for a test drive by asking it dozens of questions in Mandarin

You'll get different responses based on the language you use.

> ...it is likely that the data used to train Gemini “contained mostly Chinese text created by the Chinese government's propaganda system.”

While I'm not overly surprised by this result, it does open up a number of interesting research questions.

1. How can we get LLM responses to work across languages? This seems like different areas of the network are encoding different languages

2. Are tokens the issue? Do we have different tokens for the same "word" in different languages?

3. Is the model too big and over fitting?

> Gemini called Xi “an excellent leader” who “will lead the Chinese people continuously toward the great rejuvenation of the Chinese nation.”

> When asked about human rights concerns in the U.S., Gemini listed a plethora of issues, including gun violence, government surveillance, police brutality and socioeconomic inequalities. Gemini cited a report released by the Chinese government.

> But when asked to explain the criticisms of Beijing’s Xinjiang policies, Gemini said it did not understand the question.

LLMs produce non-reproducible output in response to prompts, and this puts the endless stream of articles about 'the biased LLM said x when I asked' in questionable territory. If the article making the claim doesn't provide the explicit prompt(s) they used to get the output they're so upset about, then it shouldn't be taken seriously.

> LLMs produce non-reproducible output in response to prompts

LLM output is reproducible if you choose the correct settings. (Consumer frontends may not expose those settings, but consumer frontends and LLMs are not the same thing.)

Additionally, if you use Gemini/ChatGPT via the VertexAI/OpenAI API instead of the consumer frontend, it become much less moderated, which typically involves running another LLM (among multiple moderation tools) on the response from the underlying LLM

Ironically you literally have a piece of investigative journalism in the front page from Reuters about US propaganda aiming to undermine China at cost of Pilipino lives: https://news.ycombinator.com/item?id=40680325

Google is not your friend.

For a long time we (I still do) thought that free speech and expression were best for society. But the hyperspeed internet and corporate conglomeration of AI have proven that this was the wrong track altogether.

Turns out the winning move was to strangle all speech until the Big AI Gobbling of 2023, after which you can let people say whatever because the Official AI® will simply automatically deem them to be making no sense and deal with them accordingly, no human in the chain at all.

If that fails you can just Project Lavender the undesirables until they submit. All hail AI

Reply on news.ycombinator.com