ChatGPT suddenly issues a sweeping ban, leaving global AI writers completely cut off from their core available resources overnight.

The golden age of AI writing is collapsing!  

As new generations of models keep getting more powerful, the articles they generate are getting worse and worse.  

They know when to list a few suggestions, where to insert subheadings, and never forget to add a concluding sentence at the end.  

But after reading the whole piece, you can barely retain a single sentence in your mind.  

The more powerful the model is, the less human-like the writing becomes

Recently, Adam Hunt, a postdoctoral researcher at the University of Cambridge whose research focuses on evolutionary psychiatry and human evolution, has also noticed this phenomenon.

Hunt used to be very optimistic about AI. But on July 28, he confessed in a long post that he has become increasingly pessimistic now.

There was once a popular metaphor that depicted model capabilities as a circle with a sharp spike.

First, code and math capabilities surpass humans, then as the model scales up, capabilities in every dimension grow gradually, and AGI is finally achieved.

In reality, the spike representing code and math is indeed growing longer and longer, while language expression and simple reasoning capabilities are not improving synchronously.

From the perspective of reinforcement learning, this outcome is almost inevitable.

Earlier generations of LLMs seemed to "become smarter across the board" because their training corpora covered all fields, ranging from poetry to academic papers, spanning the entire internet.

However, that is just a byproduct of the training data, not the model's real "comprehension ability".

Chain-of-thought and web search features were subsequently introduced to extend the lifecycle of these models, but this path eventually hit a wall.

Shortly afterwards, AI companies eager to turn the tide set their sights on code.

The reason is simple: there is an endless supply of training data for code.

All commits, reviews, and PRs on GitHub are readily available. Moreover, similar to mathematics, code has clear feedback signals: a program works if it runs successfully.

When it comes to writing articles? What can you use to automatically score "how good a piece of prose is"? Without corresponding rewards, RL will not optimize in that direction.

As a result, all training resources are shifted to code, and language expression capabilities can only rely on the legacy left over from the pre-training phase.

The legacy can only get thinner and thinner, and even regressions may occur to accommodate performance improvements in other fields.

This is exactly where Goodhart's Law is easily manifested.

When a metric is taken as the training target, it can no longer accurately reflect the real capability. Laboratories optimize models using benchmark tests, then use similar tests to prove that the models have become stronger.

The accuracy on the leaderboard will continue to rise, but the actual user experience may stagnate or even decline.

This time, even imitation is restricted

While capabilities are degrading, another thing is also happening.

A Reddit user posted a complaint that he had been using ChatGPT to write a book for several months with a premium model subscription, then took a break for a while.

When he came back, the prompt that had always worked suddenly returned a rejection.

The prompt itself is not complicated: specify a writer, ask the model to add more dialogues, enrich details, and avoid fragmented sentences.

In the past, as long as the name of the writer was specified, the model could generate content close to the writer's style, without extra explanations on sentence length, rhythm, and perspective.

Now, this kind of prompt no longer works.

Ars Technica tested prompts targeting Stephen King, J.K. Rowling, Amy Tan, and Charles Dickens, while Engadget tested prompts targeting Agatha Christie, and ChatGPT rejected all imitation requests.

If you don't name specific writers, but only ask for abstract features such as "suspense intensity", "narrative rhythm", and "dialogue density", the model can still output content, but the style is no longer the same as the original writer's.

The golden age of AI writing has collapsed

Nowadays, this controversy has shaken users' basic expectations for AI writing.

Models do not necessarily write better and better, and prompts no longer remain effective permanently.

Once the platform changes its rules, the entire writing pipeline built around a specific model can collapse instantly.

However, AI writing will not disappear, it will just return to the position it should be in: a tool.

It can help look up information, organize structures, and revise sentences, but it cannot replace the author to decide what to write and why to write.

In other words, human experience, judgment and desire for expression are becoming increasingly valuable.

References: https://arstechnica.com/ai/2026/07/chatgpt-stops-cloning-famous-writers-voices-but-may-capture-a-similar-feeling/

This article is from the WeChat official account "AI Era", author: ASI Revelation, published with authorization from 36Kr.