What Does AI Generated Text Mean For The Enterprise
'The Impact of AI-Generated text on the Internet' is a really interesting paper as it looks at actual data to figure out what effect this is having.
By mid 2025, about 35% of newly published websites were classified as being AI generated or AI assisted, which just goes to show the impact GenAI has had in the last few years.
The study tested six hypotheses about how people perceive the impact that AI generated text is having. These were:
- Truth decay (more factual errors)
- Epistemic islands (fewer outbound links)
- Entropy dilution (longer but lower‑density content)
- Stylistic monoculture (loss of unique writing styles)
- Semantic contraction: (Online content becomes more semantically similar)
- Positivity shift: AI generated text is much more positively sanitized
Interestingly, as per the paper, only two of these six commonly feared negative impacts were supported by the data, and these were 'semantic contraction' and 'positivity shift', whereas general public perception is that all six are having a negative impact.
Probably the biggest gap is that 75% of people believe factual accuracy is declining but the study found no statistical evidence of increased factual errors.
The other thing I found interesting in the study was that people who use AI daily are less likely to believe the negative hypothesis whereas people with a negative view of AI are more likely to believe all six are occurring. I guess people believe what they want to believe.
This seems to be the first large scale study of its type and it used Wayback Machine sampling across 2022-2025. The test used multiple AI detectors eventually selecting Pangram V, and the study combined quantative web analysis with a nationally representative survey.
This study got me thinking about AI generated text in the enterprise and the impact there.
Semantic contradiction mean that a companies blog, docs, sales collaterals etc will start to sound like everyone else's unless marketing teams differentiate.
Although the study found no statistical increase in factual errors many users believe the opposite so it shows that truth decay isn't the problem, trust decay is.
As I've said many times before, AI is a multiplier of a persons expertise not a differentiator in and of itself. Your own inputs (data, expertise, POV etc) are the differentiator that you can scale with AI.
Read the paper 👉 https://ai-on-the-internet.github.io/