'SNS Dominated by AI Bots... Signs of Collapse? [Tech Talk]'

AI Bots Starting to Dominate SNS
Will the Internet Be Filled Only with AI Content?
Critical Impact on Model Training Ecosystem

Recently, social networking services (SNS) including X have become filled with AI-automated chatbots. These bots flock to popular posts with high user exposure, randomly leaving replies and encouraging activities such as 'likes' and 'retweets' that help increase view counts.

These AI bots are likely accounts created to generate advertising revenue. Moreover, since they are relatively easy to create, they have rapidly flooded the platforms. For now, they only cause minor annoyance or discomfort to users, but there are warnings that the 'flood of AI bots' could one day contribute to a crisis on the internet.

The Flood of 'AI Bots' Starting on SNS

A "chatbot account" on a social networking service (SNS) presumed to be created by users of Indian nationality. Screenshot from X

Today, SNS platforms like Facebook, Instagram, and X operate business models that share a portion of advertising revenue with popular accounts. The more exposure and recommendations or reposts an account receives, the greater its earnings.

At some point, 'AI bot accounts' began to flood in, aiming to exploit this system. These accounts link text chatbots like ChatGPT or image generators to automatically create comments, posts, and more. So far, their quality has been rough and their responses 'inhuman,' making it easy to recognize them as AI. However, occasionally they generate posts that seem genuinely human, sparking attention.

Will the Internet Become Filled with Less Human Data and More AI?

An example of the so-called 'model collapse.' When trained only on data generated by artificial intelligence (AI), the results gradually collapse to the point where the shapes become unrecognizable. FreeThink capture.

As companies began launching AI services, the flood of 'automated content creation tools' was almost inevitable. However, the real problem is that this flood of bots could potentially collapse the internet and even AI itself.

In August this year, the international academic journal Nature published a paper addressing the so-called 'model collapse' phenomenon, drawing attention from the IT industry. Model collapse refers to a phenomenon where neural network models trained solely on AI-generated data, rather than human data, gradually become unstable and eventually 'collapse' to a level of uselessness.

This model collapse phenomenon has served as a chilling warning to computer scientists and AI researchers. The exact reason why AI trained on self-generated content collapses is not yet known. Some scholars speculate that, similar to biological 'inbreeding,' a lack of diversity in data can cause neural networks to become malformed.

As AI models grow exponentially larger, the amount of data used for training also increases. This has led to a point where the human-generated text data left on the internet is actually becoming insufficient. If only 'low-quality' data left by AI bots remains on the internet, AI companies will find it increasingly difficult to create environments suitable for training models.

Enhancing Chatbot Performance Also Reduces Human-to-Human Interaction

Another issue is that the advancement of AI itself reduces interaction between humans. For example, a study published last year found that one year after coding automation features were introduced to ChatGPT, activity on Stackoverflow (an overseas website for sharing development tips) dropped by 16%. Communication with chatbots has partially replaced communication between humans.

Paradoxically, the more chatbots replace human roles within the internet, the faster the amount of data available for AI training will be depleted. Ultimately, AI companies will need to spend more time carefully selecting high-quality 'real' human data scattered across the internet, which will slow down AI development accordingly.

Diversity Is Essential to Prevent 'Model Collapse'

Estimated Inventory of AI Training Data. The amount of online data produced by humans is gradually decreasing.

Of course, 'model collapse' is not a certain future. Big tech companies are already developing various tools to filter out AI-generated data. For example, Google mandates embedding a digital 'watermark' invisible to the human eye into AI-generated images. Data with this watermark is later filtered out from AI training datasets.

There is also an argument that the 'diversity' of AI models can prevent model collapse. As mentioned earlier, the collapse phenomenon is similar to biological inbreeding issues. Just as human interactions involve diverse patterns, AI must be trained on high-quality data that captures such diversity to remain healthy across generations.

Therefore, even if synthetic data created by AI floods the future, if that data is generated by different AI models, the likelihood of collapse decreases.

To enhance the 'durability' of AI technology against collapse, regulatory authorities worldwide need to strictly address monopoly issues. This means promoting competition so that not only AI from some big tech companies but also other AI models can be actively utilized.