AI-generated content threatens the quality of information on the Internet

What's this about?

People disagree about whether AI-made content harms the quality of information (facts and ideas) on the Internet.

It can spread false claims fast, but careful use can also help people.

What supporters say

• AI writing can sound smooth and sure, even when the facts are wrong.
• AI tools can make huge amounts of false content at a low cost.
• Made-up data can feed future AI tools and slowly lower their quality.
• Some websites already post many low-quality AI-made stories with false claims.

What critics say

• People can use AI as a helper, then check and fix its work.
• AI can help more people find and use helpful facts and ideas.

How to read this

The number of points on each side does not show which side is right; strong proof matters more.

The bottom line

AI-made content poses a big but uneven threat to online facts and ideas.

The strongest worry comes from cheap, large-scale lies and smooth errors, while careful use can still improve access and quality.

AI-generated content poses a substantial but uneven threat to the quality of information on the Internet. The strongest concerns involve falsehoods produced at scale, convincing errors and the gradual weakening of the information supply. But supervised use of AI can also improve quality and widen access.

The case for

Generative AI makes it cheap to produce large amounts of misleading material. Language models can create spam, propaganda, impersonation and targeted messages tailored to different audiences. RAND and the World Economic Forum describe this ability as a serious risk to the information environment, although they also caution that the scale of real-world operations and the success of efforts to stop them remain uncertain. The key change is the low cost and flexibility of producing persuasive falsehoods. 1 (see Figure 1)

AI-generated errors can be especially difficult to spot because the writing often sounds confident and polished. Research including the TruthfulQA benchmark has found that language models can repeat common misconceptions, invent facts and citations, and give unreliable explanations in fluent language. This creates a direct route from convincing presentation to inaccurate information online. 2

The problem may also build over time. Research on “model collapse” suggests that repeatedly training systems on synthetic material can reduce variety and quality in later outputs. Errors, repetitive styles and the loss of unusual or minority viewpoints could spread through websites, training data and future AI systems. The research offers a plausible mechanism, supported by both theory and experiments, though it does not show how large the effect is across the Internet. 3

There are signs that this process is already occurring in some corners of the web. NewsGuard has documented networks of low-quality sites publishing large numbers of AI-generated articles, including material containing errors and fabricated claims. That finding demonstrates a working spam pathway, but it does not establish what share of all online content is affected. 4

The case against

AI assistance does not automatically lower quality. A randomized field experiment found that generative-AI tools improved average productivity and customer-support quality for many workers, with particularly large benefits for less experienced employees. Human supervision and a structured workflow can turn AI into a quality-improving tool rather than a source of unreliable content. 5

AI may also make useful information more accessible. Analyses by the OECD and UNESCO point to potential benefits in translation, education, accessibility, drafting and lower-cost information production. Those benefits come with risks involving errors, bias, privacy, unequal access and weak oversight, but they show that AI can expand the supply of helpful information as well as unreliable material. 6

The overall effects depend heavily on how systems are used: the model and prompt, the subject area, the level of human review, platform incentives, provenance tools and users’ ability to check claims. Systems such as C2PA can record where content came from and how it was edited, helping with attribution and origin. But adoption remains incomplete, and provenance cannot establish whether the underlying claim is true.

The bottom line

The evidence favours the claim, and it does so with high confidence in the direction of the effect, but not in its exact size. AI-generated content is a substantial and credible threat to parts of the Internet because it can scale misinformation, hide errors behind fluent language and potentially weaken the diversity and reliability of future information.

That conclusion is not universal. The evidence does not show that all AI-generated material is worse than human work, or that AI has already reduced the quality of the Internet as a whole. The strongest evidence shows both real factual failures and clear benefits from supervised use; the main uncertainty is how often documented risks occur at population-wide scale, and how effectively platforms, governance, provenance systems and human reviewers can contain them.