Classroom Approaches for Teaching Students How Generative AI Affects Truth

Put a photorealistic image of a person on the projector and ask the class, “Is this real?” Hands go up. Someone says yes, someone says no, someone says it depends. That last answer is the one you want. You are not asking them to be suspicious of every pixel. You are asking them to build a habit of checking. The S.U.R.E. steps from the National Library Board are the backbone because they do not care which tool created the media. Source, Understand, Research, Evaluate works whether the material came from a chatbot, a video generator, or a forwarded WhatsApp message from a family member. The failure mode is tool over-trust: a student who runs an image through one detector and takes the verdict as gospel has learned nothing durable. Start with the image, then teach the method.

The National Library Board's four-step sequence for information literacy is embedded in Singapore's school curriculum and public libraries. It survives any technology shift because it is a habit, not a tool. Source means asking who published this and why. Understand means checking whether you actually grasp what the material claims. Research means opening new tabs to see if the claim holds up. That is lateral reading. Evaluate means weighing the evidence fairly, including evidence that contradicts what you wanted to believe. The Singapore classroom context matters here because many students already know the steps through library programmes. The 2024 launch of the “S.U.R.E. for Schools: Generative AI” module made the connection explicit: the sequence now has a version built for chatbot outputs and synthetic images. The lesson is not about memorising four words as a mantra. It is about applying them in order until the order becomes automatic. When a student sees a deepfake of a public figure, the reflex should fire: verify the source, understand the context, research the claim, evaluate the evidence. The steps are the durable skill. The AI tool is just today's occasion to practise them.

The National Library Board's S.U.R.E. for Schools programme has a generative AI module that gives educators a ready-made structure. It walks students through the four steps with examples drawn from chatbot conversations and AI-generated images. It does not assume the students are afraid of the technology or that they should be. It assumes they will encounter AI-generated material and need a method to decide what to do with it. The module is free and designed for classroom use. That makes it a lower-effort starting point than building a lesson from scratch. The approach is deliberately calm: the goal is not to scare students into distrusting everything they see online, but to give them a verification habit that scales. The module also covers the limits of the method. It cannot tell you whether a specific image is fake. It can tell you what questions to ask before you share it. That distinction is the difference between media literacy and paranoia. For the educator, the module is a scaffold. It provides the vocabulary and the sequence. The teacher provides the conversation about why this matters.

Hands-on, not theoretical. Show a side-by-side comparison: a real photograph of a public figure and an AI-generated image of the same person. Ask the class to list everything that looks off. The hands. The ears. The lighting on the eyes. The text in the background. Then do the part that teaches the skill: run the AI-generated image through a reverse image search. The search will show where else the image has appeared online. If it is a synthetic image, it may not appear anywhere except AI galleries. That is the lesson. The activity works because it combines visual analysis with a concrete verification step. The reverse image search is the tool, but the principle is the durable one: check whether an image has appeared in a different context before you believe what it shows. The failure case is the AI realism trap, where a student sees a photorealistic image of a person who does not exist and assumes it must be real because it looks real. The activity does not require expensive software. It requires a browser, an image, and a curiosity about where the image came from. The goal is not to make students expert forensic analysts. It is to trigger the question: how would I check this?

Here is the uncomfortable truth: AI detection tools are not reliable enough to judge student work. Independent testing has shown false positive rates high enough to falsely flag human student writing as AI-generated. The problem is worse for non-native English writers. If you use a detector to accuse a student of cheating, you will eventually accuse an innocent student. The tools claim high accuracy. Those claims are marketing, not evidence. The classroom consequence is that you cannot outsource judgment to a checker. Design assignments that make AI-generated work visible. Require drafts. Require in-class writing. Require students to explain their process. The S.U.R.E. steps offer a better path: teach students to evaluate their own work and the work of others by asking where the information came from and whether the reasoning holds up. The failure mode here is tool over-trust, where both teacher and student believe a detector's verdict without understanding its error rate. The honest position is to tell students that no tool can reliably distinguish their writing from AI output. Then assess what they produce through the process, not just the final product.

Start with a live demo. Open a chatbot and ask it a question about a recent event in Singapore, something specific like a new policy or a recent statistic. The chatbot will answer with confidence. The answer may be completely wrong. Then show the lateral reading move: open a new tab, search for the claim, and check whether any credible source confirms it. The lesson is not that the chatbot lied. The lesson is that the chatbot is not designed to tell the truth. It is designed to generate plausible text. Hallucination is not a bug that occasionally happens. It is a property of the system. The four-step sequence gives students a script: Source (who is the chatbot citing?), Understand (does the answer make sense in context?), Research (can I find a second source?), Evaluate (is the evidence strong enough to share with a friend?). The demonstration works because the failure is visible and the remedy is concrete. Students leave knowing that a confident-sounding AI is the one that needs the most checking, not the least.

When the discussion moves to provenance, the C2PA standard is the technical anchor. C2PA is a set of specifications that lets creators and publishers attach metadata to digital files, showing when it was created, by what device or software, and whether it has been edited. Major tech companies support the standard. It is designed to give viewers a way to verify a file's history. But the classroom needs to teach what it cannot do. A C2PA credential proves that a piece of media was created or edited on a particular device at a particular time. It does not prove that the media is true. A politician could give a speech that is factually wrong and attach a credential to it. The credential proves the file is authentic, not the claim it contains. Provenance is a starting point, not a verdict. The four-step method adds the rest: check the source, understand the claim, research the facts, evaluate the evidence. For the Singapore classroom, this matters because deepfakes of public figures are a real risk. The response is not better tools. It is better habits.

Singapore's approach to deepfakes is sectoral, not a single omnibus law. The Protection from Online Falsehoods and Manipulation Act, amended in 2024, extends to deepfake material that presents false statements of fact. The election law specifically bans realistic AI-generated audio or visual material of a candidate saying or doing something they did not during the election period. The Online Criminal Harms Act covers other harmful material. The administering ministry is the Ministry of Digital Development and Information. The lesson for students is not the specifics of each statute. It is that Singapore draws a line between material that is merely unpleasant and material that is illegal. POFMA has issued over 800 correction directions since 2019. That number signals the government's willingness to act. The Media Literacy Council, disbanded in 2020, has been replaced by other efforts. The government's Factually portal and independent bodies like Black Dot Research continue to publish verifications. For the classroom, the legal angle is contextual. It explains why a deepfake of a politician during an election is not just deceptive. It is against the law. The verification steps still apply. The consequences of sharing are higher than a social media ban.

Layering the Skills Across Weeks

A media literacy curriculum for generative AI cannot be a single lesson. It has to be a sequence that runs across weeks, reinforcing the same skills from different angles. The four-step method provides the spine. The first lesson introduces the sequence and the idea that truth is not about what feels real but about what can be verified. The second lesson applies it to images, with a reverse image search as the verification tool. The third applies it to text, with a chatbot hallucination demo and lateral reading. The fourth applies it to video, where deepfakes are the hardest to spot and the highest stakes. The fifth ties it together with a discussion of algorithmic curation and echo chambers, explaining why students see the material they see and why their friends see different things.

Designing Assignments That Resist AI Cheating

The curriculum should be explicit that AI detection tools are not reliable enough for assessment. Design the assignments accordingly. The through-line is that verification is a habit, not a fact to be memorised. The curriculum works because it is not about any specific tool. It is about the question the student asks when they see something surprising online: how do I know this is true?

Singapore's fact-checking ecosystem gives the classroom a set of living examples. Black Dot Research, founded in 2021, is an independent fact-checker that publishes local verifications. The government-run Factually portal clarifies false claims circulating in Singapore. The Singapore Police Force tracks ten categories of scams, including those using AI-generated deepfake audio or video of public figures. These organisations are the tools a lateral reader opens in a new tab. Use them the way a professional would: check the source, see what the fact-checker found, and evaluate the evidence chain. The fact-checkers also model the verification steps in action. Their methodology is published. Their sources are named. Their verdicts are reasoned. The Singapore classroom can compare a claim to a fact-check and see the whole process. The failure case is single-source settlement: accepting the first fact-check you find without checking whether the fact-checker itself is credible. Verification is a layered process. The fact-checkers are one layer, not the whole answer.

Ready-Made Materials and Worksheet Adaptations

For the educator who wants ready-made material, the National Library Board's S.U.R.E. for Schools site has lesson plans and slide decks. Teachers can also adapt the Factually portal and Black Dot Research outputs into a worksheet. Give students a claim that is currently being circulated. Ask them to apply the four steps. Then compare their findings to the published fact-check. The exercise teaches the method and the ecosystem at the same time.

Structuring a Lecture Around Failure Cases

For a lecture, structure it around the failure cases: source confusion, recency bias, tool over-trust, and the AI realism trap. Each failure case gets a slide with a real example. The lecture ends with the single most important question a student can ask: how do I check this? The answer is always the same. The method is the four-step sequence. The specifics change with every platform and every tool, but the question is permanent. The lecture should be honest about the limits. No tool catches everything. No method is perfect. The goal is to reduce harm, not eliminate it.

Students cannot understand why they see the material they see without understanding algorithmic curation. Platforms like TikTok and YouTube optimise for engagement, not accuracy. The algorithm shows material that keeps the user scrolling. That is often material that confirms what the user already believes. Echo chambers form because people seek information that reinforces their existing views. The algorithm amplifies that tendency. Confirmation bias is the cognitive mechanism: people remember the evidence that supports their position and forget the evidence that contradicts it. The Singapore classroom adds a local dimension. Forwarded-as-received messages on WhatsApp and Telegram are the primary vector for misinformation in the country. A message that arrives from a trusted family member or a church group chat feels true because the sender is trusted, even though the sender did not verify it. The four-step sequence is the countermeasure: Source (who forwarded this and why are they forwarding it?), Understand (what is the claim actually saying?), Research (what do other sources say?), Evaluate (is the evidence good enough to share?). Algorithms are not neutral distributors. They are engagement optimisers. The user's job is to break the spell.

Media literacy is incomplete without privacy literacy. A student who shares a deepfake or forwards a scam message leaves a digital footprint. Digital footprints last. The data a person actively leaves, posts, uploads, and shares is a footprint. The data collected about them passively, browsing history, location tracking, data broker profiles, is a digital shadow. Both matter. Singapore's Personal Data Protection Act protects data about an individual who can be identified from that data or from that data and other information. The law does not protect a student from the consequences of their own shares. Once something is posted, it can be archived, screenshotted, or indexed even after deletion. The failure mode is privacy resignation: believing that because data breaches happen, there is no point in managing privacy settings. The verification steps do not cover privacy directly, but the evaluation step includes asking what the material reveals about the sharer. The Singapore context adds the ScamShield app, which has blocked over 200,000 scam calls, and the 1799 anti-scam helpline. Privacy is a skill. The skill involves thinking before posting because the post is permanent.

The legal structure around deepfakes in Singapore is less dramatic than the headlines suggest. There is no standalone deepfake law. Instead, there is a patchwork: the Protection from Online Falsehoods and Manipulation Act covers false statements of fact, the election law bans realistically generated audio or visual material of a candidate during the election period, and the Online Criminal Harms Act covers other harmful material. The Ministry of Digital Development and Information administers POFMA. The Singapore Police Force tracks impersonation scams using AI-generated deepfake audio or video of public figures as a distinct scam type. The classroom lesson is that the law can punish the distribution of a deepfake. It cannot undo the reputational damage or the emotional harm. The law is not the primary defence. Media literacy is. The verification steps are the preventive measure. The law is the backstop. The failure case is the student who assumes that because something is illegal, it will not happen, or that because it is not illegal, it must be true. Neither assumption survives contact with the actual information ecosystem. Teach the method first. Mention the law as context, not as the main event.

The forwarded-as-received message is the workhorse of misinformation in Singapore. Someone shares a message on WhatsApp or Telegram. It carries the label “forwarded many times,” which is almost a badge of honour rather than a warning. The message claims a public figure said something, a scam is operating in a particular neighbourhood, or a government agency issued a warning. The verification steps handle the forwarded message the same way they handle an AI-generated image: check the source. Who originally wrote this? Is there any evidence the public figure actually said it? What does the official government site say? The Singapore Police Force publishes scam advisories. The Factually portal corrects falsehoods. The failure case is source confusion, where a reader believes a claim because the person who forwarded it is trusted, without checking whether that person is the original source. The classroom activity is simple. Take a real forwarded message. Ask students to check whether the claim is true using the four steps. Then reveal the fact-check. The format of a message, a screenshot, a voice note, a forwarded chain, does not make it true. The verification habit does.

This subject serves the parent who is tired of fighting with a teenager about whether a celebrity really said that. It serves the educator who needs a concrete lesson and does not have time to build one from scratch. It serves the community leader who runs digital literacy talks at a community centre and needs accurate, sourced Singapore examples. It serves the genuinely curious student who wants to understand how the technology works so they can explain it to friends. It does not serve the reader who wants a checklist of prompts to make AI-generated art, because this subject is about what to do when you encounter the output, not how to create it. It does not serve the reader who wants a comprehensive guide to Singapore's privacy laws, because the material only covers what is directly relevant to verification. It does not serve the reader who wants a list of tools to detect AI material, because the research says those tools are not reliable enough to trust. The subject is narrow by design. It is about the single most important skill a student can learn in the age of generative AI: how to ask whether something is true.