"Bard and ChatGPT will perish without your assistance"

2023-08-14

In order to maintain large-scale language models and make them better than previous versions, they require human-created content.

Google and OpenAI openly acknowledge that in order to improve Bard and ChatGPT, they need your data. Recently, The Guardian published an article in which Google stated that copyright laws should be modified to allow generative AI systems to scrape internet content.

The company urges Australian policymakers to support a "copyright regime that allows for appropriate and fair use of copyright-protected content to train AI models in Australia on a wide and diverse range of data," while also providing an opt-out option for entities unwilling to have their data used for AI training.

On the other hand, in the debate over unauthorized web scraping, OpenAI has introduced an automated web crawler called GPTBot. This bot can collect publicly accessible data to train AI models, and OpenAI guarantees to execute this process in a transparent and responsible manner.

As generative AI becomes increasingly popular, the demand for data is also growing. ChatGPT and Google Bard, which are based on large-scale language models, rely on a large amount of text, images, and videos.

OpenAI states that GPT-4 will learn from approved, curated, and publicly available data sources, which may include publicly accessible information. OpenAI recently obtained a trademark for GPT-5, and its success depends not only on the computational power of GPUs but also on the quality of the training data.

However, the availability of data from OpenAI and Google still remains uncertain due to public awareness of their web scraping practices, which has sparked strong opposition.

Human content is the lifeline

To maintain large-scale language models and make them better than previous versions, human content is needed. The question here is whether companies should pay for this content or simply obtain it from the internet. Given the current situation, it is not surprising if future human-generated content is sold at a high price.

OpenAI is not only training GPT-4 on human content but has recently started training it on the dataset created by ChatGPT. However, this training cannot continue for too long as it eventually leads to model degradation. When the model learns from data generated by other models, this degradation process occurs. Therefore, the accuracy of the true data distribution gradually decreases.

Anyone who attempts to write poetry with ChatGPT will easily discover that it has been trained on the highest level of poetry collections and prose. Unfortunately, OpenAI did not seek permission from the authors.

Last month, 8,000 writers, including Margaret Atwood, Viet Thanh Nguyen, and Philip Pu, signed a petition calling on AI companies to stop the unauthorized or unlicensed use of writers' works. They believe that the hard work behind any form of art deserves recognition and should be attributed to the relevant creators.

However, when it comes to copyright of AI-generated works, ownership issues arise. Typically, the initial ownership of a work belongs to its creator under copyright law. However, since there have been no legal or copyright office rulings regarding AI creations, it remains uncertain who the true creator is.

A twist in the story

Currently, both OpenAI and Google are proceeding cautiously. They have shifted the responsibility of sharing data to the creators. Google states that creators should be able to choose not to have their works used by generative AI. A Google spokesperson pointed out that they hope to develop community-developed standards similar to the robots.txt system, which would allow creators to opt out of having parts of their websites crawled.

Similarly, OpenAI mentioned in a blog post that if you do not want GPTBot to access your website, you can block it by adding GPTBot to your website's robots.txt file. This means that website owners need to take proactive measures to prevent OpenAI from accessing their websites instead of allowing them to use their content for training. This is the first step OpenAI has taken to allow people on the internet to choose not to have their information used to train their large-scale language models.

The question is: is this approach correct? It is worth noting that users are initially defaulted to an opt-in state and need to manually set themselves to an opt-out state. This may be because convincing individuals to give up their privacy rights can be challenging for creators of large-scale language models.

In addition, OpenAI has taken measures to avoid legal disputes, such as recently partnering with The Associated Press to obtain real-time data that can be freely used to train their models. By doing so, they ensure access to a vast and diverse dataset without infringing copyright laws or facing potential legal issues.

Engaging in a compensation battle with these companies may not yield any results due to the lack of proper legal support and the time and cost involved. Therefore, if you rely on ChatGPT or Bard to complete tasks such as writing emails or coding, the cost you face is sacrificing your own data.

RECENT AI NEWS

RECENT AI TOOLS