Pages outlive the facts they were built on
Every day, people make real decisions from pages that were accurate when they were written. They follow sign-up steps for a service that shut down, or budget around a free plan that was quietly withdrawn. The page gives no sign that anything changed, because nobody went back to change it.
The web has never had a way to keep its pages current. It barely keeps them at all. When Pew Research Center checked pages that existed at some point between 2013 and 2023, it found that a quarter of them were no longer accessible by late 2023.
Many of the pages that survive have gone unchecked. Pew found that "23% of news webpages contain at least one broken link, as do 21% of webpages from government sites." In the same study, more than half of Wikipedia pages (54%) had at least one reference pointing to a page that no longer exists. Those are pages people still read, and each broken link marks a spot nobody has fixed since the page it pointed to disappeared.
Why pages go stale and stay that way
A page goes out of date when the thing it describes changes, and it stays out of date because nothing tells the publisher.
A page is written once, about things other people control. A company changes its prices on its own schedule. An agency revises a rule, or a regulator recalls a product. None of them writes to every site that ever described the old version, and the publisher has no reason to look again. The page goes on saying what it said on the day it went up, and it keeps being found and read.
Publishing is set up as a one-way act. A new page gets written and promoted. An old page sits in an archive, and checking it again is easy to put off. Even documents written with great care decay this way. When researchers at Harvard Law School checked the web links in United States Supreme Court opinions in 2013, they found that "49.9% of the links cited in the Supreme Court opinions no longer had the cited material."
Broken links are the easy case, because a program can find them. An out-of-date fact is harder to catch. The page loads, the sentence reads well, and nothing on it tells the reader it has stopped being true. Finding one takes someone who knows what changed and a document that proves it.
Publishing became almost free
Generative AI made it cheap to produce pages at a volume no editor could review. Google named the problem in its own rules. In 2024 it began treating "producing content at scale to boost search ranking" as abuse. Its spam policies give this example: "Using generative AI tools or other similar tools to generate many pages without adding value for users". Google later reported that once the change rolled out, its results carried 45% less low-quality, unoriginal content.
The pages that get through don't have to be malicious to do damage. Being wrong and confident is enough. NewsGuard, which tracks AI-generated news sites against a published four-part test, counted 49 of them in May 2023. Its tracker now lists 3,749.
AI answers and old pages
AI assistants and AI search now build their answers out of web pages, so a wrong page can become a wrong answer.
Google describes the mechanism in its documentation for site owners. AI Overviews and AI Mode may run what Google calls a query fan-out, "issuing multiple related searches across subtopics and data sources" to put a response together. The same holds beyond Google. OpenAI's documentation for developers says that when its models search the web, "the model's response will include inline citations for URLs found in the web search results." The answer is assembled from pages, and it inherits whatever those pages got wrong.
The tools add errors of their own, and their makers say so. Google's help page for AI Overviews states: "AI Overviews can and will make mistakes." The same page advises readers to "Always check important info in more than one place." OpenAI's own researchers explain part of the reason: language models "sometimes guess when uncertain, producing plausible yet incorrect statements instead of admitting uncertainty." An out-of-date page gives a system like that something plausible to repeat.
The largest published test of AI assistants on news questions, run by 22 public broadcasters in 18 countries and coordinated by the EBU and the BBC, found that 45% of answers had at least one significant issue, that "31% of responses showed serious sourcing problems", and that a fifth "contained major accuracy issues, such as hallucinated and/or outdated information." In a separate test by the Tow Center at Columbia Journalism Review, eight AI search tools asked to identify the source of a news excerpt got it wrong more than 60 percent of the time.
When an AI answer repeats an out-of-date page, the tool has done what it was built to do: it found a page and reported what the page says. In that case the error starts with the page, and a page is something we can fix.
Why a middleman
We do the checking a publisher would otherwise have to do, so that a request to fix a page arrives ready to act on.
Anyone can email a publisher to say a page is wrong. A message like that asks the publisher to do the research: find out whether the reader is right, what changed and when. A publisher may not have time for that, and a message with no evidence is easy to set aside.
We send a different kind of message. It points to the exact sentence, states what is true now and when it changed, and links the documents that prove it. It has been checked twice before it arrives, and the publisher can confirm every part of it for themselves. That makes updating the page a small job.
It also spares the person who noticed the error. Someone who spots a mistake may not want to argue with a stranger's website, or may have reasons to stay out of it. They can hand the claim to us, and their name stays with us unless they choose otherwise. Every claim goes through the same checks, and every publisher gets the same message, which anyone can read on our page for publishers.
The fix starts at the source
No one can audit every answer an AI tool gives. The pages those answers draw on are public, and pages can be fixed.
Correcting the record is an old duty in journalism. The Society of Professional Journalists' code of ethics tells journalists to "Acknowledge mistakes and correct them promptly and prominently." Most websites are not newsrooms, and many have no one whose job is to go back and check. We can't hold them to a code they never signed. We can bring the evidence to each publisher, one page at a time, and ask.
This initiative is built on a simple bet: correct the page, and everyone who reads it afterward reads the correction, including any system that reads it next.
Correct the page, and everyone who reads it afterward reads the correction.
What we do
Content Correction takes out-of-date and wrong claims to the people who can fix them, at no cost to anyone.
Anyone can send us a page that is out of date or wrong, such as a guide to a retired product or a page that still describes a withdrawn approval. The people who write to us are readers who noticed something, and people and businesses that a page gets wrong. Both get the same checks. We check the claim against primary sources, and if it holds up, we take it to the publisher with the evidence and ask them to fix the page, where people actually read it. Then we tell the person who sent it what happened.
We take on a claim only when acting on it can cost someone money, time, access, health or safety. Trivia stays out, however well sourced. The full rules are on how it works.
We don't publish the claims we work on. The goal is a corrected page in the place people read it. A public list of other people's mistakes would do nothing for the readers of those pages, and it could make publishers less willing to talk to us.
What we will never do
These limits apply to every claim, whoever sends it.
- Argue opinions, predictions or open scientific questions. We act only when documents settle the claim.
- Name and shame. We work with publishers directly and publish nothing about the pages we contact.
- Take sides in politics. Political claims are out, whoever makes them.
- Charge publishers, or take money to leave a page alone.
- Help remove accurate information, however unwelcome it is.
- Pass on the name or email of the person who sent a claim, unless they ask us to.
- Give medical, legal or financial advice. We point to what an official source changed.
A note from the founder
I grew up in Seattle, and for as long as I can remember I've been drawn to journalism. I never thought of myself as much of a writer. What pulled me in was the plainer part of the work, getting the facts down and keeping them straight.
Over the last few years I watched AI-generated articles spread across the web, page after page of confident text that nobody seemed to have checked, and I watched AI tools start answering people's questions from whatever they found. I started Content Correction to do something practical about it. We can't check the whole web. We can check one page at a time, and ask the people who publish it to put it right.
If you've found a page that's out of date or wrong, send it to us. I read every submission.
