“The only thing that can truly make them better is more stuff, but when the majority of stuff being created isn't good, or interesting, or even written for a human being, ChatGPT or Claude's models will learn the rotten habits of rotten content.”
Our reading:
Only more data can truly improve the models, and when most new content is poor or not written for humans, ChatGPT and Claude will learn its bad habits.
Assessment: Unresolved
What happened
Did the forecast come true?
Outcome: Unresolved
We checked, and the evidence so far doesn’t settle it.
Judged on a plain reading. They named no measure, so we used the reading a typical informed reader would take. It is spelled out under “How we judged it”.
Evidence up to . Checked .
Why
Graphite estimated AI authorship at 50.9% of sampled English-language articles in Q4 2025 and 49.9% in Q1 2026 [e4]. These estimates do not establish that most new internet content is poor or intended for machines [e4]. The April 14, 2024 preprint version reports defects from generated training data in experimental models [e1]. The April 29, 2024 study reports avoiding collapse when real and synthetic data accumulate [e2]. Neither study establishes the predicted acquisition of defects in ChatGPT or Claude [e1, e2]. OpenAI's September 12, 2024 disclosure attributes reasoning improvements to reinforcement learning, questioning the claim that only additional training content can improve performance [e3]. It does not provide the controlled training-content comparison needed to settle that premise [e3]. The evidence therefore leaves the governing conditional forecast unresolved [e1, e2, e3, e4].
Caveats
- AI-authorship estimates do not measure the claimed quality or intended-audience condition [e4].
- The experiments do not connect that condition to training and inherited defects in subsequent ChatGPT or Claude models [e1, e2].
- Avoiding model collapse does not establish the absence of lesser defects [e2].
- OpenAI's disclosure does not fully isolate training-content quantity from other changes [e3].
- The forecast specifies no deadline.
Evidence (4 items)
Context only: arXiv; Shumailov and colleagues,
use of model-generated content in training causes irreversible defects
Why it matters: The experiments support the possibility of training-content feedback causing defects. They do not establish the internet-wide antecedent or an observed outcome in deployed ChatGPT or Claude models.
- Where
- Abstract
- Publisher
- arXiv; Shumailov and colleagues
- Kind of source
- Preprint, not peer reviewed
- About events on
- Published
Context only: arXiv; Gerstgrasser and colleagues,
Nevertheless, we find that model collapse is avoided when accumulating data.
Why it matters: Keeping original real data while accumulating synthetic generations avoided collapse in the tested settings. This limits an inevitability reading, but avoiding collapse does not prove that models inherit no lesser defects.
- Publisher
- arXiv; Gerstgrasser and colleagues
- Kind of source
- Preprint, not peer reviewed
- About events on
- Published
Context only: OpenAI,
This process dramatically improves the model’s ability to reason.
Why it matters: OpenAI attributes reasoning gains to reinforcement learning and separately reports gains from additional inference computation. This challenges the premise that additional training content is the only route to better performance, but does not isolate every training-data change or test the forecast's antecedent.
- Publisher
- OpenAI
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- About events on
- Published
Context only: Graphite,
before returning to 49.9% in Q1 2026.
Why it matters: The study classifies sampled English-language articles by likely AI authorship. It measures neither all new internet content nor whether that content is poor or intended for machines, so it cannot establish the antecedent.
- Publisher
- Graphite
- Kind of source
- Independent measurement
- About events on
- Published
How we judged it
- Kind of claim
- Conditional forecast
- Deadline
- None given. We don’t add one.
Conditions
- when the majority of content created isn't good or written for humans (their words)
Where we had to interpret
- This combines an unconditional claim about how models improve with a conditional forecast about their future behavior. The conditional forecast governs the overall assessment.
- In context, 'more stuff' means additional training content. 'Better' means materially improved performance on useful tasks.
- The antecedent concerns a majority of newly created internet content being poor or intended primarily for machines. AI authorship alone does not establish either property.
- 'Learn the rotten habits' means that training on this content causes identifiable defects in subsequent model outputs, such as repeated errors or degraded quality. It does not require total model collapse.
- Controlled studies of other models can establish a possible mechanism, but cannot alone establish the forecast's outcome for ChatGPT or Claude.
Right if: Evidence establishes the stated majority condition and shows that subsequent ChatGPT or Claude models acquired identifiable output defects through training on that content. Supporting the additional exclusivity premise would also require evidence that improvements depend on additional training content rather than other changes.
Wrong if: A controlled improvement in a named model with training content held fixed would contradict the exclusivity premise. To contradict the conditional prediction, evidence must establish the antecedent and directly test whether exposure to the specified content causes the predicted defects; general benchmark improvement alone is insufficient.
We wrote this test after the outcome was known.
What would settle it
Representative measurements must establish the majority condition. Dated training records and controlled tests must then show whether exposure causes the predicted defects in ChatGPT or Claude. A separate controlled experiment holding training content fixed could test the exclusivity premise, but would not alone settle the conditional forecast.
We found 31 claims in this post. See them all.
Full record
Source and dates
Where's Your Ed At, Are We Watching The Internet Die?, paragraph 35.
- Said
- We read the source
- live page, full text, public
- Evidence up to
- Checked
We read this wording on 2026-10-28. The post may have been edited after 2024-03-11.
Where it was said (1)
Said once in the posts we read.
: Where's Your Ed At, Are We Watching The Internet Die?
The only thing that can truly make them better is more stuff, but when the majority of stuff being created isn't good, or interesting, or even written for a human being, ChatGPT or Claude's models will learn the rotten habits of rotten content.
- Where
- paragraph 35
- Published
- 2024-03-11
- We read it
- , live page
- Access
- full text, public
History
Versions
- Current: Unresolved. Evidence up to 2026-10-28.
What happened to this record
PublishedBy us
Published after both AI reviews and the final check passed.
Held backBy us
The final check failed: a cited source didn’t say what the write-up said.