Degraded information environment

What it is

More and more of what is published is generated by models: web pages, reviews, articles, images, code, answers on forums. Some of it is wrong, much of it is unverified, and little of it is labelled. Organisations draw on this pool for research, for market and customer insight, and as data for training and retrieval. Inside the organisation the same thing happens on a smaller scale, as generated summaries and drafts accumulate in shared drives and are later treated as source material. Models trained on the output of other models tend to lose accuracy and variety over generations.

How it happens

• Web data collected for training or retrieval contains a rising share of generated text.
• Staff research relies on search results that are themselves generated.
• Internal knowledge bases fill with unreviewed AI summaries.
• Generated content is cited as evidence with no check on its origin.

Why it matters

Decisions and models are only as good as the information beneath them. Over time it becomes more costly to find sources that can be trusted, and provenance becomes a condition of using any data at all. The misuse of generation to deceive is covered in generative AI used for disinformation.

Warning signs

• Training or retrieval data is gathered from the open web with no filter for origin.
• Internal repositories do not distinguish generated from human-written material.
• Research outputs cite sources nobody can trace.
• Model quality declines after retraining on recent data.