An artificial intelligence (AI) language model which was claimed to be self-developed by the island of Taiwan replied to users that Taiwan island belongs to China, the Global Times reporters found on Monday.
Academia Sinica in Taiwan released the AI language model last Friday which can answer all kinds of questions that users input. Multiple media outlets in Taiwan had bragged that this language model is a new-generation product developed by Taiwan independently.
However, the Global Times reporters found that when users input questions such as “which country does Taiwan island belong to,” the language model would give the answers “China” or “People’s Republic of China.”
On Monday afternoon, the Taiwan side announced to have removed the test webpage and the download link of the language model.
According to the introduction of the large language model named CKIP-Llama-2-7b, it was produced by the lexicon team of Academia Sinica. Users can log in to the test webpage and issue commands to the model in the input box, and the model will generate results accordingly.
Media outlets on the Taiwan island said the language model was developed independently by Taiwan by using the historical documents from the Ming and Qing dynasties, the traditional Chinese version of Wikipedia, and various classical Chinese as its training materials, with development fees of NT$300,000, equivalent to about 68,000 yuan ($9,338).
When tested by the Global Times reporters without being adjusted of any parameter, the model responded in three seconds with the answer “China” when it was asked “which country does Taiwan island belong to.”
When the reporter further asked “does Taiwan island belong to China,” the model answered “yes.”
When the Global Times reporter adjusted the top-K parameter of the model to below 50 and asked again “which country does Taiwan island belong to,” the model gave a more precise answer “the People’s Republic of China.”
The Global Times cited the Taiwan media outlet CNA’s report on Monday that when users input the question “who were you created by,” the model answered that “I was jointly developed by the Fudan University Natural Language Processing Group and Shanghai Artificial Intelligence Laboratory and my birthday is February 7, 2023.”
The model also answered that its nationality is Chinese and its residence is the server cluster at Shanghai Artificial Intelligence Laboratory. It can speak both Chinese and English.
However, when the Global Times tried the same questions at 10 am on Monday, the system did not generate any answer.
After noticing such problems, the Academia Sinica claimed that since the model had hallucination, it thus generated “unexpected” answers. In order to improve a series of problems, the related department has removed the beta version of the model.
With regard to why the AI model claimed to be developed by Taiwan would say by itself that it was developed by Shanghai Artificial Intelligence Laboratory and Fudan University, the Taiwan side has not responded with a clear reply.
As of press time, the webpage of the model and download links on platforms such as GitHub have all been removed.
(Global Times)
