Identifying AI-Generated News Websites and Synthetic Content Farms

Stop reading. Open a new tab. Run the site's domain through a WHOIS lookup right now, before you scroll any further. AI-generated news sites are built to be found by search engines, not by editors. Their business model collapses the moment you check a WHOIS record. A news site that has existed for 30 days and was bulk-registered with privacy-shielded WHOIS is not journalism; it is an inventory of prompts. Search the site's name on Factually, the Singapore POFMA Office's fact-checking page, and on Black Dot Research, the independent local fact-checker. If neither has heard of the site, you have already learned more than the byline will ever tell you.

A content farm is a machine that turns keywords into pages with zero editorial process. You have seen the output even if you have not named it: a headline that matches a trending search, a wall of text that says nothing, and a byline that is a stock photo of a person who does not exist. The tell is not in the grammar. Large language models now write clean prose. The tell is in the absence of everything that makes reporting expensive. There are no named human sources, no original quotes, no dateline that corresponds to a real bureau. The structure is identical across posts because it is a template. The publication timestamps are implausibly frequent; a one-person outfit cannot publish a 2,000-word investigation every twenty minutes. NewsGuard tracks more than 1,265 such sites globally as of 2025, categorising them as Unreliable AI-Generated News Sites. You do not need to read all of them to spot the pattern. You need to check whether the site has any of the features a real outlet has. That means a person you could sue.

Open the About page before you read a single piece. You are looking for two things only: a named editor and a physical mailing address. Not a PO Box in a WeWork. Not a generic 'our team' with AI-generated headshots of people who look like they modelled for a stock photo farm. A real name with a paper trail and an address that you could walk to. If the About page is a list of platitudes about 'truth' and 'integrity' with no named humans, you are done. The site has no editorial process by definition. A legitimate outlet discloses its ownership, its funding, and its corrections policy. A content farm hides all three. This is the single fastest screen: if you cannot find out who is responsible for the words, the words are not the product. You are.

Here is the checklist you will actually use. Step one: run the domain through a WHOIS lookup tool. Check domain age and registration pattern. A real news organisation registers its domain years in advance and keeps the registrant information current. A content farm bulk-registers dozens of domains at once, uses a privacy-shielded WHOIS, and lets the domain expire after the traffic dies. A domain that is 90 days old and already ranking for breaking news is not staffed by investigative journalists. It is staffed by a script. Step two: check the Internet Archive's Wayback Machine. If the site looked like a different content farm a year ago, that is your answer. Step three: reverse image search the byline photos. If the 'reporter' is a stock photo or a generated face, the piece is a fabrication. This is not about being a technical expert. It is about spending two minutes with a WHOIS lookup instead of trusting a layout.

How the Bait Works

An AI-generated site can mix a real headline from a legitimate outlet with three fabricated paragraphs of detail. That is the core failure of the 'it sounds true' heuristic. The claimed-versus-real gap is the distance between what the piece asserts and what primary sources tell you. A single true-sounding piece is not confirmation. It is the bait. The most common pattern is a rewritten version of a legitimate story from CNA or The Straits Times, stripped of attribution and padded with fake quotes. The names of the real officials are still there, but the specifics are wrong: a wrong date, a wrong ministry, a meeting that never happened.

Closing the Gap

You catch this by lateral reading. Open a new tab and search for the central claim. If no other outlet has it, and the ones you trust have a different version, you have your answer. The failure is single-source settlement: accepting a claim as true because you found one source that confirms it, without checking whether that source is independent.

Lateral reading is the habit of leaving the original page to check the source, the claims, and the context before you believe or share anything. It is the core behaviour that distinguishes a careful reader from a victim. The opposite is vertical reading: staying on the page and letting its design, logo, and tone convince you. Content farms are designed to survive vertical reading and to fail lateral reading. So open a new tab. Search for the site's name plus the word 'fact check' or 'POFMA'. Search for the claim itself in quotation marks. Check whether the Singapore POFMA Office has issued a correction direction. Check whether Black Dot Research has published a fact-check. This is not about being paranoid. It is about spending ninety seconds to avoid looking foolish in the group chat, where the forwarded-as-received message lives. The 'forwarded' label on WhatsApp and Telegram is a structural feature; treat it as a red flag, not as an endorsement.

Two cognitive shortcuts amplify AI-generated news, and both are baked into how Singaporeans use messaging apps. The first is the officialdom halo: believing a message because it uses a government logo, official-sounding language, or a URL that looks like a .gov.sg domain. A content farm knows this; it borrows the logo and the tone because it has no legal exposure. The second is the echo chamber: an online environment where a person encounters only information that reinforces their existing views, driven by algorithmic curation and social sorting. If you only see content that confirms your biases, an AI-generated piece that flatters your politics will feel more credible than a dry correction on Factually. This is how a fake story about government spending circulates for days before the POFMA order lands. The fix is not to distrust everything. It is to intentionally seek out the source that would tell you the story is wrong.

Satire is content created with artistic or comedic intent that an audience is expected to recognise as exaggeration. A deceptive falsehood is presented as fact. The confusion arises when satire is stripped of context and recirculated as news. An AI-generated site does not care about the distinction; it republishes a satirical headline as fact because engagement is engagement. You need to tell the difference yourself. Ask: is the piece on a site that labels itself as satire? Does the author use obvious hyperbole? Is the 'source' a parody account? If you have to ask whether it is satire, the answer is that it has already been stripped of the context that made it satire. When in doubt, do a lateral read. If the story is too absurd to be true and appears nowhere else except the site you are on, it is either a hoax or a joke. Either way, do not share it as news.

You need two bookmarks. The first is Factually, run by the Singapore POFMA Office, which publishes corrections and clarifications on falsehoods that have been subject to correction directions. The second is Black Dot Research, an independent, Singapore-based fact-checking organisation founded in 2020 that publishes local verifications. When you encounter a piece on a suspicious site, search both for the site's name and for the central claim. If the POFMA Office has issued a correction direction against a site for AI-generated content, you will find it there. As of late 2025, the POFMA Office had issued corrections against at least three sites for AI-generated pieces with false claims about government spending. The figure for 2026 is not yet published; check the Factually site directly for the current count. The point of these sources is not to outsource your judgment. It is to give you a fast, reliable baseline for what is true before you invest any more time in a suspicious piece.

When a claim is circulating, pay attention to how fast the fact-checkers respond. Verification speed measures how quickly a fact-checking source publishes a verdict on a circulating claim; it is a useful proxy for how seriously the claim is being taken. Source transparency is the other half. Does the fact-checker name the original source and show the evidence chain, or is it a verdict without visible evidence? A verdict without evidence is unverifiable by you. The same applies to the piece you are reading. Does it link to primary sources, or does it link to itself? Does it name a person you could contact, or is it a wall of anonymous assertion? A methodology disclosure from the news site is also vital. If the site does not publish its fact-checking process and funding, you are reading content with unknown provenance. The presence of these features is not a guarantee of truth. Their absence is a guarantee of low editorial standards.

You can shortcut the entire process by checking whether the domain is on a known list. NewsGuard, which monitors misinformation, tracked more than 1,265 AI-generated news and information sites globally as of 2025; their list is a gold standard for researchers. The common pattern among these sites is a generic author bio with an AI-generated headshot, no contact information, no physical address, and an identical structure across posts. But the list is not your primary check. The WHOIS is. Domain age is your best single indicator because content farms do not live long. They get the domain, pump out hundreds of pieces, get the traffic, and then the domain gets burned when the hosting provider or the ad network shuts them down. A domain that is 90 days old and claiming to publish breaking news from Singapore is either the most efficient newsroom in history or a scam. You know which.

The 'forwarded as received' label on WhatsApp is a structural feature that signals the message has not been written by the person who sent it to you. It is designed to make you pause. Most people do not pause; they reply with 'is this true?' and hit forward again. That is a verification bypass: sharing a claim with a comment like 'is this true?' without waiting for an answer amplifies the claim before anyone checks it. You are now part of the problem. The next time you see a forwarded piece, do not forward it. Verify it. If you cannot verify it, do not share it. The cost of being wrong is not worth the social credit of being the first to share. The dynamic is the same whether the message is about a government scheme, a food scare, or a celebrity. The platform is irrelevant. The habit of checking before forwarding is the core media literacy skill, and it is the one that will protect you and the people who trust you.

One caution: do not upload a suspicious image to a random AI-detection tool as your first step. You may be revealing your data to the very people who created the deepfake. Think about privacy preservation. If an image is of you or someone you know, uploading it to a public site may reveal your identity, your location, or the fact that you are investigating it. The tool may be a honeypot for data collection. Instead, use a reverse image search on TinEye or Google Images, which do not store the image for anyone else to see. But even here, be careful about the AI realism trap: a single tool's verdict is not definitive. The tool landscape changes monthly, and the claimed accuracy rates are overstated. Your own lateral reading is the most reliable tool. A legitimate website does not require you to upload data to verify it. A scam often does. If a verification step asks for personal information, stop and reconsider.

If you are in the 16-to-25 age bracket, you already know that your group chat is a vector for misinformation. You see a forward from a friend's uncle that says the government is giving away money, and you are not sure if it is true. Here is what you do. Do not reply 'reported' and do not forward. Instead, open a new tab, search the claim on Factually, and check if the government has made such an announcement on its official channels. Then check the domain age of the site you are reading. If it is 30 days old, that is your answer. You do not need to be a journalist to do this. You need to be a reader who is willing to spend ninety seconds on the truth before spending hours propagating a falsehood. The skill is not memorising the laws. It is understanding that the algorithm rewards engagement, and outrage is high-engagement. The system is working exactly as designed. It is up to you to break the cycle.

You do not need a technical background to verify a claim. You need a framework. The National Library Board's SURE framework (Source, Understand, Research, Evaluate) is taught in Singapore schools and public libraries, and it works. Source: who is telling you this, and why should you believe them? Understand: do you actually understand the claim, or are you filling in the gaps with what you expect it to say? Research: what do other, independent sources say about this? Evaluate: given what you have found, is the claim likely to be true? The key is to apply this in the moment, not after you have already committed to sharing. The framework is durable because it does not depend on any specific technology. It will matter just as much ten years from now when the deepfakes are even better. The skill is the same whether you are evaluating a news piece, a health tip, or a friend's opinion about the weather.

Black Dot Research's most practical advice is the 'who, when, where' test. A real news piece answers those three questions with precision. An AI-generated one fails at least one of them. Who is the named human source? Not a 'spokesperson' or 'an official', but a real person with a title and an organisation you could look up. When did the event happen? Not 'recently' or 'in the coming days', but a date you can verify. Where exactly did it happen? Not 'in Singapore', but a specific location. If the piece cannot answer these three basic questions, it is not journalism. It is a language model filling a page with words that sound plausible. The telltale sign is the absence of specifics that can be fact-checked. Generic details are easy to generate. Precise details are not. When you see a vague piece that fails the 'who, when, where' test, you are looking at the output of a content farm.

You have done everything. The WHOIS lookup shows a domain aged five years. The About page names a real editor, and you found their LinkedIn. Factually has nothing bad to say. The piece itself is well-sourced and answers the 'who, when, where' test. Is it real? Not necessarily. The final failure mode is the single-source settlement: you have found one source that confirms the claim, but it is the same source you are trying to verify. You need lateral reading again. Search for the claim on a site you already trust, like CNA or The Straits Times. If they have not picked it up, there is a reason. And if the piece is about a breaking event, check the official government channel directly. A ministry press release on their own site is a better source than any second-hand report. This is the officialdom halo in reverse: the official source is the most reliable one, but you have to go to it. You cannot wait for it to come to you.

Your social media feed is not showing you the world. It is showing you what algorithmic curation selects for engagement. The algorithm is not a conspiracy; it is a program that optimises for clicks, likes, and shares. Outrage and fear generate engagement. Accuracy does not. This is why a fake story that scares you will outperform a dry correction. The result is an echo chamber where your views are reinforced and contrary evidence is filtered out. A content farm exploits this by writing headlines engineered to appeal to your existing political biases. The most effective way to break out of the echo chamber is to follow a diverse set of sources, including ones you know you will disagree with, and to lateral read every surprising claim. It is not about being unbiased. It is about being less wrong. The algorithm will always optimise for engagement. It is your job to optimise for the truth.

Here is your next 90 seconds. You are reading a suspicious piece. Do not read it to the end. Open a new tab. Type the site's name into your search engine, followed by 'fact check' or 'POFMA'. Click the first result from Factually or Black Dot Research. If they have a correction on the specific claim, you are done. If not, search for the claim in quotation marks. If it is nowhere else, the piece is a hoax or a deepfake. Do not share it. If you must share it because you want to warn people, share the fact-check link, not the original. This habit, repeated every single time, is what breaks the chain. The failure mode is thinking that one exception is fine. It never is. The verification speed is on you now. In the time it takes to read this sentence, you could have checked one claim. Do that, and you will be a better informed citizen than most. The tools are free, the effort is small, and the payoff is not being fooled by a machine built to fool you.

Do AI-generated news sites mix real headlines with fake details?

Yes, this is their most effective trick. They take a real headline from a legitimate outlet and generate a plausible-looking piece around it, with fabricated quotes and slightly wrong details. The single-source settlement is not enough.

Can I rely on a single AI-detection tool?

No. The tool landscape changes monthly, and the claimed accuracy rates are overstated. Use your own lateral reading as the primary check, and treat any single tool's verdict as a lead, not a conclusion.

What is the one number I should remember?

NewsGuard tracked more than 1,265 such sites globally as of 2025. That number is growing, but the checks you use today will be the same ones you use next year.

The most practical thing you can do is search the site's name on Factually before you read the piece, and if it is not there, check the WHOIS; if the domain is younger than the piece claims to be about, you have your answer. That is a specific, named action against a named organisation (Factually), with a concrete measure (the WHOIS age), and it is a recommendation against trusting the layout. No other page can provide that exact instruction.