Some journalists and writers are making themselves redundant in order to improve the quality of AI chatbots

As generative AI develops rapidly, there is growing concern that AI will take over human jobs. On the other hand, some journalists, novelists, and scholars are working for technology companies to improve the quality of chatbots, despite the possibility that their jobs may be taken away.

'If journalism is going up in smoke, I might as well get high off the fumes': confessions of a chatbot helper | Artificial intelligence (AI) | The Guardian

Chatbots using large-scale language models such as ChatGPT can perform a variety of creative tasks, from summarizing large amounts of text to composing emails, writing essays and novels, etc. However, behind the scenes, large-scale language models need to be trained, and AI development companies are contracting with novelists, scholars, freelance journalists, etc. to train the models.

According to Jack Apollo George, who actually works at an AI development company, the contractors involved in training mainly work on creating answers to fictitious questions that can be posed to chatbots. AI generates 'good' sentences, but in order to do that, it needs to be shown what 'good' looks like. Therefore, the model is loaded with human responses to fictitious questions. This reduces the frequency of the phenomenon known as 'hallucination,' in which AI explains things that are not facts as if they were facts.

'The world of AI is built on our language. Without good human language data, large-scale language models cannot improve,' George said.

In addition, Ilya Shmailov, a researcher at Google DeepMind, said, 'Training a large-scale language model based on text output by a large-scale language model rather than on text written by humans will cause a 'model collapse' in which the model loses knowledge of rare words and rare facts, known as 'minority data,' and will not work,' highlighting the importance of humans in training AI.

In fact, OpenAI has signed contracts with News Corp, the parent company of The Wall Street Journal, and other publications, to create an environment where it can collect high-quality human-written text for further training.

OpenAI announces multi-year deal with Wall Street Journal parent company News Corp - GIGAZINE

George is employed as a 'senior data quality specialist' at an AI development company, but he has mixed feelings about his job. 'High-performance large-scale language models automate our jobs, and the more they improve, the more our jobs are taken away,' he said.

Meanwhile, software engineer François Cholet said, 'About 20,000 people are employed just to create the annotated data to train large-scale language models. But without this human effort, the text output from large-scale language models would be really bad.' Initially, contractors involved in AI training were hired at low wages from people living in developing countries, but to improve the accuracy of model training, they are shifting to specialized, higher-paid jobs, sometimes for 30 pounds (about 5,600 yen) per hour.

However, Sholeh speculates that 'major technology companies have not been very clear about the manual AI training process, and in the near future, they will correct their investment in this process.' George predicts that 'as long as AI does not develop to the point where there is a shortage of new words to train, we writers, journalists, and others will not lose our jobs.'