Verifying AI Chatbot Outputs and Identifying Hallucinations

Stop trusting fluency. A large language model has no database of verified facts. It has a statistical model of what text should follow other text, and it generates the most plausible continuation of your prompt. That is why the same model that writes a perfect essay on the Monetary Authority of Singapore will, in the next response, invent a statute, cite a court case that never existed, or attribute a quote to a person who never said it. This failure mode has a name: hallucination. A hallucination is not a glitch. It is a structural property of the technology. When a chatbot is uncertain, it does not say 'I do not know.' It fabricates something that sounds confident, because confidence is what the training data taught it to produce. Treat every factual claim as a lead, not a finding. The question is what to do next. The answer is a repeatable workflow: isolate the specific factual claim, open new tabs to search for the claim on a primary source or an established fact-checking organisation, and apply a verification framework before you believe, share, or act on anything. The tools for that workflow exist, are free, and work, but only if you use them in the right order and with the right level of scepticism. The method is not complicated. It is disciplined. Discipline is the only thing that separates a person who uses a chatbot as a research assistant from a person who gets misled by one.

Before you can verify an answer, know what you are up against. IBM Research defines an AI hallucination as a phenomenon where a large language model generates text that is nonsensical, unfactual, or ungrounded in the provided source data. That definition distinguishes a hallucination from a typo or a misunderstanding. The model is not misremembering. It is generating text with no basis in any source it was trained on. The rate at which this happens is higher than most people assume. According to the Vectara Hallucination Leaderboard, which tests general-purpose chatbots on summarisation tasks, between 3% and 27% of outputs contain factual errors, depending on the task domain and model version. That is a wide range, and the variance is not random. A model that performs well on a simple summarisation task can fail catastrophically on a question about legal precedent or medical dosing. The Stanford University Human-Centered AI programme measured even worse numbers for legal research, finding that 17% to 33% of cases cited by general-purpose AI tools were fabricated or incorrect. In the medical domain, a study published in JAMA Internal Medicine found that 10% to 20% of AI-generated health answers contained clinically significant errors. OpenAI's own work on GPT-4 found that 20% of outputs about real people contained factual errors. You do not need to memorise these numbers. You need to understand what they mean. If you ask a chatbot a question in a domain where accuracy is life-or-death, legal, medical, financial, biographical, the odds that the answer contains a fabricated detail are not one in a thousand. They are one in five, or worse. That is not a rare event. That is a coin flip with weighted dice.

Hallucinations In The Singapore Context

Black Dot Research, the independent fact-checking organisation based here, has documented cases where AI chatbots fabricated local laws, official policies, and domestic statistics when queried directly. A chatbot might confidently state that a specific section of the Penal Code was amended last year, or that a particular grant provides a certain amount of money, and be completely wrong on all three counts. It will also generate plausible but false Singlish phrases, invented local place names, and cultural references that sound right to a non-Singaporean but are pure invention to anyone who lives here. The National Library Board's S.U.R.E. programme, which stands for Source, Understand, Research, Evaluate, has published guidance specifically warning users that AI chatbot outputs must be verified against authoritative sources like NLB eResources, official agency websites, and established news organisations. The education system has taken note as well: AI literacy modules are now part of primary and secondary school curricula under the EdTech Masterplan 2030, precisely because the skill of checking an AI's work is now as fundamental as reading and writing.

The realism of AI-generated content has outrun the average person's ability to detect it. The AI realism trap is the outdated mental model that says fabricated media looks fake, blurry, with disfigured hands, unnatural eye contact, and garbled text. That model is obsolete. Current generative AI tools like Midjourney, DALL-E, and Sora can produce photorealistic images and video that pass casual inspection, and the visual tells change with every model update. By the time you learn what a Midjourney v5 hand looks like, v6 has fixed it. The result is that millions of people are confidently sharing content they would have dismissed as obviously fake two years ago. The trap is not that people are naive. It is that their detection heuristic is calibrated for a world that no longer exists. The same applies to text. A chatbot's output sounds authoritative because it was trained on authoritative-sounding text. It uses the right jargon, structures sentences the way a subject-matter expert would, and never hedges unless you ask it to. That is not a sign of knowledge. It is a sign of good pattern matching. The only way to break the spell is to stop evaluating whether the output sounds true and start evaluating whether it is true. That is a different task, and it requires a different set of tools.

Misinformation Versus Disinformation

This is also why the distinction between misinformation and disinformation matters here, even though both end with you believing something false. Misinformation is false information shared without intent to deceive. Disinformation is false information shared deliberately. When you share a chatbot's hallucinated answer because it sounded plausible, you are participating in a misinformation chain. When you share it knowing it might be false because you want to reinforce an existing belief, you are participating in disinformation. The AI does not care which one you are doing. It just produces the next most likely token. Understanding the difference helps you understand your own failure mode. The fallback position, 'I did not know it was wrong', is not a defence. You do know it might be wrong, because you know that AI chatbots hallucinate at a rate of 3% to 33%, depending on the domain. When you share a chatbot answer without verification, you are not the victim of technology. You are the vector for it.

Here is the method. It is the same one that fact-checkers use, adapted for a single person at a keyboard. Step one is to isolate the claim. A chatbot answer is rarely one claim. It is usually several nested inside a paragraph. Extract the single factual statement you care about. Not 'What is the penalty for shoplifting?' but 'The penalty for shoplifting is a fine of up to $5,000 and/or imprisonment of up to 3 years for a first offence.' That is a claim with three parts: the fine amount, the imprisonment term, and the fact that it applies to a first offence. Each part is separately verifiable and separately fallible.

Step two is to open new tabs. This is the lateral reading technique that the Stanford History Education Group recommends. Leave the chatbot window entirely and search for the claim on its own. Do not ask the same chatbot to verify itself. That is like asking a compulsive liar if he is telling the truth. Search for the exact phrasing of the claim, then search for the component parts. For a local claim, your first stop should be Factually, the official fact-checking portal run by the Ministry of Communications and Information. Factually publishes clarifications on falsehoods circulating locally, and it is indexed by search engines. If the claim is about a statistic, go to the source: the Department of Statistics, the Ministry of Health, or the relevant statutory board. If the claim is about the law, go to Singapore Statutes Online, the authoritative source for enacted legislation.

Apply The S.U.R.E. Framework

Step three is to apply the S.U.R.E. framework. Source: Who published this information? Is it a primary source, an official website, a peer-reviewed journal, or a random blog? Understand: Do you actually understand what the source is saying, or are you interpreting it through the filter of the chatbot's framing? Research: Has this claim been fact-checked before? Search the claim on Google Fact Check Explorer, which aggregates fact-checks from verified signatories of the International Fact-Checking Network, including Black Dot Research and CheckMate, the Straits Times fact-checking column. Evaluate: Is the source credible for this specific claim? A retired police officer might be a good source for how police procedures work, but not for the text of the law. An official site is a good source for policy, but not for how a policy is enforced in practice.

Check Dates And Handle Failure

Step four is to check the date. AI chatbots are trained on a corpus that has a cutoff date, and they do not automatically update when new laws are passed or new statistics are published. A chatbot might confidently tell you the current rate of the GST, and be correct, for 2023. The rate is different now. Any claim that involves a number, a date, a name, or a title of a publication should be checked against the most recent primary source you can find. This is not a failing of the chatbot. It is a property of the technology. It cannot tell you what it does not know.

Step five is the failure case: what do you do when you cannot verify the claim quickly? You do not share it. You do not append a caveat like 'not sure if this is true' and hit send. That is verification bypass, and it is the single most common way misinformation spreads on WhatsApp and Telegram. You either do the work of verification, or you stay silent. There is no third option that involves forwarding a suspicious claim with a question mark and calling it due diligence.

Given how common hallucinations are, there is an obvious market for a tool that detects them automatically. The ScamShield app uses AI to filter scam messages and calls, but it does not currently detect AI-generated voice or video in real time, and it will not help you verify a factual claim in a chatbot's response. For AI text detection, tools like GPTZero, Originality.ai, and Copyleaks claim accuracy rates between 70% and 99%, depending on the length of the text and the model that generated it. That sounds reassuring until you read the fine print. The accuracy rate is the rate at which the tool correctly identifies text that is AI-generated. The false positive rate, the rate at which it flags human-written text as AI-generated, is a separate number, and it is far higher for non-native English text. Studies of these tools consistently find that they perform worse on text written by people for whom English is a second language, because the statistical patterns these tools use to distinguish AI from human text are trained on native-speaker patterns. The result is a tool that might be 90% accurate overall but is 70% accurate for a local writing in Singlish. A 10% to 30% error rate is not a verification tool. It is a coin flip with a bias. Never use a single AI-detection tool as the basis for a conclusion. The only honest use of these tools is as a trigger for further investigation. If the tool flags text as AI-generated, treat it as a signal, not a verdict, and apply the lateral reading workflow above.

Image And Video Verification

For images, the landscape is similar. Hive Moderation claims 99% accuracy for AI-generated image detection on standard benchmarks, but standard benchmarks are not real-world conditions. SynthID, Google's watermarking system, embeds an invisible watermark in images generated by Imagen and audio generated by Lyria, but it does not cover all AI generators, and it is trivially easy to defeat by screenshotting or compressing an image. The only image verification tool that works in practice is reverse image search, via Google Images, TinEye, or Yandex, which finds where an image has appeared before and whether it has been used in a different context. That is the tool that catches a deepfake of a public figure or a real photo from a disaster that has been recaptioned as a current event. For video, the InVID-WeVerify plugin, developed by Agence France-Presse and partners, allows frame-by-frame analysis and reverse image search of video keyframes. That is the only reliable way to check if a video is genuine or AI-generated. None of these tools is perfect, but they are all better than trusting your eyes, which the AI realism trap has already compromised.

The S.U.R.E. framework is the National Library Board's information literacy framework, embedded in the school curriculum and public libraries. It stands for Source, Understand, Research, Evaluate, and it is designed to be used by anyone, with any kind of information, on any subject. Applied to AI chatbot outputs, it looks like this.

Source: The chatbot is not a source. It is a medium. The actual source of any claim is the original document, dataset, or authority that the chatbot is drawing on, or fabricating. Your job is to find that original source and evaluate it. Is it an official website? A peer-reviewed study? A news article from a reputable outlet? Or is it nothing at all, because the chatbot invented the underlying source entirely? A key step is source transparency: does the organisation or person behind the claim name their sources? If a chatbot tells you 'according to a 2023 study by the National University of Singapore,' the verification step is to find that study. If you cannot find it, that is a red flag. If the chatbot cannot cite a source at all, that is not a failure of memory. It is a sign that the claim may be a hallucination.

Understand: Do you understand what the claim actually means? A common failure in AI communication is that people accept a claim because it sounds technical, without understanding the underlying meaning. If you do not understand what a 'portable alpha strategy' is, or what 'Good Class Bungalow' means, you cannot evaluate a claim that uses those terms. The Understand step is where you ask: what would it mean for this claim to be true, and what would it mean for it to be false? This is also where you check for the echo chamber effect. Are you accepting the claim because it reinforces what you already believed, or are you evaluating it on its merits? Algorithmic curation has made this harder, because your social media feed and your search results are optimised for engagement, not accuracy, which means you are more likely to see content that confirms your existing views.

Research And Evaluate

Research: This is the lateral reading step. Open new tabs. Search for the claim, the source, the author, and any specific numbers or names involved. Use Google Fact Check Explorer to see if the claim has already been fact-checked. For local claims, check Factually, Black Dot Research, CheckMate, and Sure Anot, CNA's fact-checking segment. For health claims, go to HealthHub and the Ministry of Health's website. For legal claims, go to Singapore Statutes Online. The research step is where you find out whether the claim has been verified, debunked, or is simply absent from the record. Absence from the record is its own finding.

Evaluate: The final step is to weigh the evidence. Decide whether the source is credible for this specific claim, whether the research supports the claim's specific wording, and whether there are other independent sources that agree. A single source that confirms the claim is single-source settlement, and it is a failure mode. You need at least two independent sources that agree, and if they disagree, you need to understand why. This is also where you consider the source's methodology disclosure: does the organisation behind the claim explain how it gathered its data? Does it publish its fact-checking process and funding, as required by the International Fact-Checking Network code? If not, treat the claim with suspicion.

You do not have to be a professional fact-checker to verify a chatbot's answer, but you do need a checklist, because your memory and your intuition are not reliable tools. This is the workflow, stripped to its essentials, designed for the moment when you are on a mobile phone, at 1am, with a forwarded message that contains a claim you suspect is false.

• Copy the exact claim. Not the paragraph, not the gist, the precise factual statement. If it is a number, a date, a name, or a title, isolate it.
• Open a new tab and search for the claim's exact wording in quotes. If a fact-check has already been done, this will surface it.
• Open a second new tab and search for the claim's component parts. Add 'site:gov.sg' to your search to restrict results to official sources, or check Factually directly.
• If the claim involves an image, run it through Google Images or TinEye. If it involves a video, use the InVID-WeVerify plugin to extract and reverse-search keyframes.
• Check the date on any source you find. If the source is older than the claim's context, it may be out of date.
• Apply the S.U.R.E. framework in your head: is the source credible? Do you understand the claim? Have you researched it beyond the first search result? Have you evaluated the evidence, or are you accepting the first confirmation you find?

This checklist takes under two minutes for a simple claim and under five minutes for a complex one. The burden is low. The payoff is the difference between being the person who spotted the falsehood and the person who spread it. The verification speed of fact-checking organisations is a useful benchmark: Factually, Black Dot Research, and CheckMate all publish corrections within hours to days of a claim circulating, which means that for any viral claim, the fact-check probably already exists. Your job is to find it before you share.

The market for AI text detection tools is a mess. The models that generate text are trained on vast corpora of human writing, and the detection tools are trained on text that the generators produced at a specific point in time. When a new model version is released, GPT-5, Claude 4, Gemini 2, the detectors' performance immediately degrades, because they have not seen the new model's output. The claimed accuracy rates of 70% to 99% are for the specific model versions the tool was tested on, not for all AI-generated text. The false positive rate is the real problem. A false positive is when a tool flags a piece of text as AI-generated when a human actually wrote it. For non-native English speakers, the false positive rate is dramatically higher, because these tools use statistical patterns like sentence length and word choice variety to make their determinations, and non-native writing is more consistent in its patterns than native writing. A tool that is 5% inaccurate on native English text can be 30% inaccurate on text written by a local whose first language is Chinese or Malay. That is not a verification tool. That is a defamation tool, because it will falsely accuse people of using AI to write their essays. Do not use these tools to make claims about authorship, and never treat their output as definitive. Their only legitimate use is as a trigger for further investigation: if a detector flags text as AI-generated, that is a signal to apply your own verification workflow.

Follow The Evidence, Not The Authorship

A better approach is to focus on the content, not the medium. Ask the chatbot for its sources, and then check those sources yourself. If a chatbot cites a specific study, find the study. If it cites a statistic, find the original dataset. If it cannot provide sources, or if the sources do not exist, you have your answer. This is the digital footprint approach: you are following the trail of evidence, not trying to identify the author of the text. The advantage is that it works even when the AI detection tools fail, and it works for any language, any subject, and any level of technicality.

Here is what the workflow looks like when applied to a real-world claim. Suppose a forwarded message on WhatsApp claims that the authorities have announced a new 'digital wellness' payout for citizens who spend less than two hours per day on social media, and that the payout is available for a limited time through a link in the message. The message includes a screenshot that appears to come from an official website, complete with a logo and a credible layout.

Your first reaction should be scepticism, because this claim has the hallmarks of a scam, but 'this is a scam' is a conclusion, not a method. Apply the workflow. First, isolate the claim: a payout is being offered, for a limited time, through a specific link. Second, open new tabs. Search for 'digital wellness payout' in quotes. If the claim is real, it would have been announced, and the announcement would be on gov.sg, Factually, or the relevant ministry's website. If the claim is false, you will find either nothing or a fact-check from Black Dot Research or CheckMate. Third, do not click the link in the message. The link is the scam. It will either phish for your personal data or install malware. Instead, hover over the link to see the actual URL. If the domain is not a domain ending in '.gov.sg', it is fake. Fourth, check the date: if the claim involves a new policy, it should be searchable; if no legitimate news source covers it, it does not exist. Fifth, apply S.U.R.E.: the source is an unsolicited forwarded message, the understanding is that official announcements are not made this way, the research step involves checking the primary sources, and the evaluation is that the claim fails because it appears nowhere on any official website.

This is the difference between a person who has been trained and a person who has been misled. The trained person does not need to know whether the specific claim is true or false. They need to know how to find out. The method is the same whether the claim is about a payout, a law, a statistic, or a piece of medical advice. Isolate, search, check the source, check the date, apply the framework. That is the entire skill, and it is learnable in an afternoon.

Tool	What It Claims	What You Should Actually Do
GPTZero, Originality.ai, Copyleaks	70-99% accuracy in detecting AI-generated text	Treat as a signal, not a verdict; high false positive rate on non-native English text
Hive Moderation	99% accuracy on standard benchmarks for AI image detection	Standard benchmarks are not real-world conditions; verify with reverse image search
SynthID (Google DeepMind)	Watermarks AI-generated images and audio	Only works for Google's Imagen and Lyria; defeats by screenshotting or compressing
Google Fact Check Explorer	Aggregates fact-checks from IFCN signatories	Search here first for any viral claim; if a fact-check exists, it will surface
Reverse Image Search (Google, TinEye, Yandex)	Finds where an image has appeared before	The most reliable tool for image verification; always use it before trusting an image
InVID-WeVerify	Frame-by-frame analysis of video keyframes	The only reliable way to check video; use for any video with a serious claim
Factually, Black Dot Research, CheckMate, Sure Anot	Singapore-specific fact-checking of claims in circulation	The first stop for any local claim; they publish within hours to days
ScamShield	Uses AI to filter scam messages and calls	Does not detect AI-generated voice or video in real time; use for phone scams

There is a failure mode specific to people who use verification tools. It is called tool over-trust. It works like this: you run a suspicious image through Hive Moderation, it tells you the image is 99% likely to be AI-generated, and you conclude that the image is fake. You have just made a decision based on a single tool that has a documented false positive rate and a benchmark that does not reflect real-world conditions. You have also outsourced your critical thinking to a black box. The tool might be right, or it might be wrong, and the difference between those two outcomes is the difference between you looking informed and you spreading a falsehood about someone else's content.

The correct approach is to treat detection tools as one input among many. Combine the tool's output with your own lateral reading. If a tool flags an image as AI-generated, your next step is not to share that verdict. It is to search for the image using reverse image search to see if the same image appears in a different context. If a tool flags text as AI-generated, your next step is to search for the text's claims, not the text's authorship. The tool tells you that something might be synthetic. The research tells you whether the claim is true. These are different questions, and confusing them is how intelligent people end up sharing debunked content.

The only verification method that survives contact with the real world is the one that does not depend on a single point of failure. That method is lateral reading, powered by the S.U.R.E. framework, checked against primary sources, and informed by the published fact-checking work of organisations like Black Dot Research and CheckMate. It is slower than clicking a button and accepting the result, but it is the only method that you, personally, can trust.

You have a legal responsibility when you share AI-generated content, and the authorities have been explicit about this. The Infocomm Media Development Authority (IMDA) has issued an Advisory on the Use of Artificial Intelligence Generated Content, which applies to specified social media services under the Code of Practice for Online Safety. The advisory requires platforms to identify and minimise AI-generated content that poses risks to public health, public security, and racial and religious harmony. What this means for you as an individual is that you cannot hide behind 'it was just a chatbot' if you share content that breaches the law. The law does not have a 'the AI made me do it' defence.

The Broader Regulatory Framework

The regulatory framework is broader than a single advisory. The Model AI Governance Framework, first published by the Personal Data Protection Commission in 2019 and updated in 2020, covers the principles of transparency, explainability, and human oversight for AI systems. AI Verify, the testing framework and software toolkit published by IMDA in 2022, provides a way to validate an AI system's performance against those governance principles. And the National AI Strategy 2.0, published by Smart Nation in 2023, includes a specific focus on building trust in AI through testing, assurance, and public education. You do not need to read any of these documents to be a responsible user. You need to know they exist, because they shape the environment in which you are making your own decisions about what to share.

Platform Rules Versus The Law

The real point of mentioning the legal framework is to remind you that platform terms of service are not the same as the law. Platforms can take down content that does not technically violate any law, and they can leave up content that does, at least until a court or the relevant authority acts. Content moderation is not legal action, and legal action does not require content moderation. If you are ever in doubt about whether something you are about to share is legal, the safest course is to not share it. That is not a legal opinion. It is a risk-management decision. The downside of being wrong is a criminal record, and the upside of being right is that you avoided looking gullible.

There is a privacy dimension to fact-checking that is easy to overlook, especially when you are in the middle of verifying a claim and want to move quickly. When you upload a suspicious image to a reverse image search tool, you are sending that image to a third-party server. If the image contains your face, your location, or your personal belongings, you have just shared that data with a company that has its own privacy policy and its own commercial interests. The tool might be free, but you are the product. This is the trade-off between tool utility and privacy preservation: the tools that are best at verification are also the tools that collect the most data about what you are verifying.

Footprint Versus Shadow

There is also a distinction between your digital footprint and your digital shadow. Your digital footprint is the data you actively leave behind: the posts you make, the images you upload, the comments you leave. Your digital shadow is the data collected about you passively: your browsing history, your location data, your data broker profiles. When you use a verification tool, you expand both. You are adding to your footprint by sending the tool your data, and you are adding to your shadow by giving data brokers more information about your interests and concerns. If you are verifying a claim about a sensitive topic, a health condition, a political issue, a personal matter, you need to be aware that your searches and uploads are not private. The best privacy-preserving approach is to use a private browsing window, a VPN, and a search engine that does not track you, and to avoid uploading images that contain more information than necessary to verify the underlying claim. If a reverse image search requires uploading an image that contains your face or your home, consider whether the verification is worth the privacy risk.

Here is the paradox: the more you care about verification, the more you need to worry about your own data. This is not a reason to stop fact-checking. It is a reason to fact-check intelligently. The same lateral reading that verifies a claim can also be used to verify the tools you are using. Search for the tool's privacy policy, check whether it shares data with third parties, and see if there have been any data breaches. If a tool's privacy practices do not meet your standards, use a different tool. There is always an alternative.

This page keeps returning to the vocabulary of misinformation and disinformation, and it is not academic pedantry. The distinction matters because it changes what you do after you encounter a false claim. If you believe someone shared a chatbot's hallucinated answer because they did not know it was false, you are dealing with misinformation, and the response is to correct the record with evidence, kindly. If you believe they shared it knowing it was false, you are dealing with disinformation, and the response is different: you do not engage, because engaging gives the disinformation a platform. In practice, you cannot always tell which is which, so default to the assumption of good faith until you have evidence to the contrary, but keep your own standards high.

The speed at which you verify a claim matters too, and it is one of the few metrics in this field that is both measurable and meaningful. Fact-checking organisations publish their verdicts within hours to days of a claim circulating. Black Dot Research and CheckMate both have track records of turning around verifications quickly, which means that for any claim that has been circulating for more than a day or two, the fact-check probably already exists. Your verification speed, the time from when you encounter a claim to when you either verify it or dismiss it, should be faster than the rate at which the claim is spreading. If you see a forwarded message that has already been shared widely, someone else has already fact-checked it. Your job is to find that fact-check before you share it onwards. This is the single most effective thing you can do to slow the spread of falsehoods, because you are not just protecting yourself. You are protecting everyone in your contact list.

Let us be direct about the most common failure mode, because it is the one that most people do not even realise they have. You receive a forwarded message on WhatsApp that claims something alarming about a new policy. You are busy, or you are not sure, so you forward the message to a group chat with the comment 'Is this true?' appended. You have just amplified the claim. You have not verified it, but you have increased its reach, because the people in that group chat will see the claim in its full detail, and many of them will not read your comment, and some of them will forward it onwards. Your 'just asking' is not neutral. It is participation. This is verification bypass, and it is how misinformation spreads in the age of algorithmic curation.

The correct behaviour when you are not sure is the opposite: do not share the claim at all. Do not share the claim and ask if it is true. Do not share the claim with a disclaimer. Do not share the claim and then debunk it in the same message, because the claim has already been seen by everyone who will ever see it. The only way to prevent the claim from spreading is to not share it. If you must share something, share a link to a fact-check that debunks the claim, because that is the only message that should be forwarded. Your digital footprint includes every message you send, and you cannot take it back. The confused outcome, the one where you think you have contributed to the conversation when you have actually contributed to the problem, is the worst outcome, because it is invisible to everyone except the algorithm that keeps showing people the false claim because it is getting engagement.

Not all fact-checks are created equal, and the difference is in the source transparency and methodology disclosure. A good fact-check names the original source of the claim, shows the evidence chain from the claim to the verdict, and explains how the fact-checker reached their conclusion. A bad fact-check just says 'false' and moves on. The International Fact-Checking Network (IFCN) requires its signatories to publish their methodology and funding, so you can evaluate whether the fact-checker has a conflict of interest. This matters because a fact-check without methodology disclosure is just an opinion, and an opinion without source transparency is a rumour with a byline.

When you read a fact-check from Black Dot Research, CheckMate, or Sure Anot, ask yourself these questions: What is the original claim being fact-checked? Who is making the claim? What evidence does the fact-checker provide to debunk or support it? Is the evidence primary, or is it based on a secondary source? Does the fact-checker name the original source, or do they only say 'according to an official'? Does the fact-checker explain their process, or do they just present a conclusion? The answers to these questions determine whether the fact-check is a reliable source for your own verification. A fact-check that does not name its sources is a source of the same problem it is trying to solve.

This is also why lateral reading is superior to a faith-based approach to fact-checking. When you trust a single fact-checker, you are relying on their authority. When you read laterally, you are checking the fact-checker's work, looking at their sources, and reaching your own conclusion. The IFCN code requires signatories to be transparent about their methodology, and you should hold them to that standard. If an organisation cannot tell you how it knows what it claims to know, it does not know.

Verification is work, and the amount of work it takes is the user burden. Every extra step in the verification workflow is a step that most people will not take, which is why the misinformation industry is so effective. The solution is not to shame people for being lazy. It is to reduce the burden by building verification into your existing habits. Have a small set of go-to sources that you trust, and check those sources every time, without exception. For local claims, your go-to sources should be Factually, Black Dot Research, CheckMate, and Sure Anot. For health claims, HealthHub and the Ministry of Health. For legal claims, Singapore Statutes Online. For statistics, the Department of Statistics. The more you use these sources, the faster you will get at verifying claims, because you will know exactly where to look.

The update frequency of your go-to sources matters as well. A fact-checking portal that stops being updated is worse than no portal at all, because it gives you false confidence. Check the date on every fact-check you read, and be suspicious of any source that does not publish dates. Similarly, be suspicious of any source that does not have a methodology page, because that means you cannot evaluate its claims. The goal is to build a personal verification workflow that takes less than a minute for routine claims. If it takes longer than a minute, you will not do it when it matters.

This page has focused on the practical skill of verifying AI chatbot outputs, and it has done so using the local context because that is the jurisdiction this site covers. It has not attempted to explain the theory of generative AI, or the history of the technology, or the philosophical question of whether an AI can lie. Those are interesting questions, but they are not actionable. What is actionable is the workflow, the tools, and the sources. What is also not covered is comparative international law: this site does not cover the laws of other jurisdictions except to note when a scam or falsehood originates overseas, which does not change the applicable legal framework for the person reading this page.

The choice to keep this page practical is deliberate. There are entire books on the ethics of artificial intelligence, and you will not be a better fact-checker for reading them. What you will be better at, if you are a regular user of AI chatbots, is distinguishing between what a chatbot says and what you can independently confirm. That is the only skill that matters, and it is the one this page has tried to teach.

The subject of fact-checking AI chatbot outputs suits a specific kind of person: someone who uses generative AI tools regularly, who cares about accuracy, and who is willing to spend two minutes checking a claim before sharing it. It suits the student who uses ChatGPT to help with research, the professional who uses it to draft emails, the journalist who uses it to generate story ideas, and the ordinary person who receives forwarded messages on WhatsApp and wants to know what to believe. It particularly suits the reader who is aware that the authorities have taken a proactive stance on AI governance and who wants to understand what that means in practice.

It does not suit someone who wants a tool that does the thinking for them. No such tool exists, and any page that claims it does is selling something. It also does not suit someone who believes that AI is always right, or always wrong, because that person has already made a decision based on faith, not evidence. And it does not suit someone who wants to verify every single claim they encounter, because that is a full-time job. The goal is not perfection. It is to catch the costly errors, the fabricated law, the invented statistic, the fake image, and to build the habit of checking the claims that are easy to check. If you are that second kind of person, this page will not help you, and you can stop reading now. If you are the first kind, you now have a method, a set of tools, and a list of sources. Go use them.

A tool that is 5% inaccurate on native English text can be 30% inaccurate on text written by a local whose first language is Chinese or Malay.