Malaysia Day came and went, and with it came a flood of Jalur Gemilang graphics. Some were beautiful. Some were AI-generated. And a handful were straight-up fake, showing events that never happened and statistics nobody ever published.
Officials have now urged content creators to verify AI-made visuals before posting them. Good advice. But here is the part that matters for you: the checklist they are recommending takes about ninety seconds, and it works on everything, not just national day content.
If you run a brand page, a community WhatsApp group, a surau announcement channel or just a personal account that people actually read, this is a habit worth building now. Because the alternative is posting a correction later, and corrections never travel as far as the original.
What actually happened
In the run-up to 16 September, several fabricated or misleading images circulated across Malaysian social media. Some were celebratory graphics with subtly wrong details. Others were more serious: images presented as news photos, or visuals tied to claims about events and figures that did not check out.
None of this is unique to Malaysia. It is happening everywhere, for every national holiday, in every language. What makes it newsworthy here is the speed of the shift. Two years ago, producing a convincing fake celebratory graphic took skill, software and time. Today it takes a text prompt and eight seconds.
The tools got good. The guardrails did not keep pace. And the sharing habits of ordinary users have not adjusted at all.
Why national holidays are the perfect storm
Misinformation researchers have a name for the content that spreads fastest: emotionally activating material. Things that make you feel proud, angry, or protective. Not things that make you think.
Malaysia Day content sits right in that zone. It touches national identity, history, the formation of the federation, Sabah and Sarawak, and often religion too. People share it fast because sharing it feels like an act of belonging, not an act of publishing.
That is exactly the problem. Nobody reverse-image-searches something that makes them feel patriotic. You see the flag, you see the caption, you hit forward. The verification instinct switches off precisely when it is needed most.
Bad actors know this. So do engagement-farming pages that do not care whether a graphic is real, only whether it performs. And a page with 200,000 followers posting one fabricated statistic can put that number into circulation permanently, because screenshots outlive deletions.
The ninety-second checklist
Here is the practical version of what officials are recommending, expanded into something you can actually run before you hit post.
- Reverse-image search anything unfamiliar. Google Lens, TinEye, or Google Images search-by-image. If a dramatic photo has no history anywhere, that is a signal.
- Read the text in the image. Generative models still mangle lettering, especially in Bahasa Melayu and Jawi script. Look for warped characters, nonsense words, or letters that dissolve at the edges.
- Check the flag. The Jalur Gemilang has fourteen stripes and a fourteen-point star. AI models get this wrong constantly. Count them.
- Verify the claim, not just the picture. If the graphic states an event, a date or a number, search for that claim independently. Did it happen? Does the statistic exist anywhere official?
- Raise your guard on sensitive content. Anything touching race, religion, royalty or national sentiment deserves an extra minute, because those are the posts that spread fastest and cause the most damage when wrong.
The tells that AI still leaves behind
Detection is getting harder, but current-generation image models still have consistent weak spots. Worth knowing them.
Text remains the biggest giveaway. Banners, signboards, name tags, shop fronts. Models approximate the shape of writing without understanding the words. Zoom in on any text in the frame and it often falls apart.
Hands and crowds are the second tell. Count fingers. Look at the back rows of any crowd scene, where faces tend to blur into repeated or melted forms. Look at where limbs meet bodies.
Then there is the lighting. AI images frequently have a uniform, slightly airbrushed glow with no consistent light source. Shadows fall in directions that do not agree with each other. Reflections in glass or water do not match what is in front of them.
Architectural detail also breaks down. Tiles that change pattern mid-wall, railings with inconsistent spacing, buildings whose windows do not line up. For Malaysian content specifically, watch for songkok and baju kurung details rendered slightly wrong, or landmarks like KLCC and the Sabah State Mosque with geometry that is almost right but not quite.
Be honest with yourself though, lah. These tells are disappearing. Treat them as useful hints, not proof. A clean image is not a verified image.
What this means if you run a page
The stakes scale with your reach. A personal post to 300 friends is embarrassing. A brand page post to 50,000 followers is a reputational problem your marketing team will be handling all week.
If you manage any account with meaningful audience, build a rule now. One person verifies before anything with a factual claim goes out. It costs you two minutes per post and saves you from the scenario where a competitor screenshots your fake statistic and it becomes the thing people remember about your brand.
Community group admins have it hardest. You are not producing content, you are relaying it, often from five other groups, often from uncles who forwarded it from somewhere else entirely. The volume makes verification feel impossible. It is not, but you have to be selective: verify the posts making claims, let the pure well-wishing graphics pass.
And if you use AI to make your own Malaysia Day or Hari Merdeka graphics, that is completely fine. Label it. A small line saying the visual was AI-generated costs you nothing and builds trust with an audience that is getting more sceptical by the month.
The correction problem
Here is the thing people underestimate. Corrections do not work well.
Research on misinformation has been consistent on this for years: a correction reaches a fraction of the audience that saw the original, and even among those who see it, the false claim often sticks. People remember the statistic and forget that it was debunked. The retraction post gets a tenth of the shares.
So the calculation is simple. Ninety seconds of checking beforehand is cheaper than any amount of cleanup afterwards. Sit with your teh tarik, run the checklist, then post.
Where the tools are heading
Detection tooling exists but it is genuinely unreliable. AI-image detectors produce false positives on real photographs and miss well-made fakes. We would not build a workflow around them.
Content provenance standards are more promising. C2PA is an industry effort to embed cryptographic metadata into images at the point of capture or generation, creating a verifiable record of where a file came from and what edited it. Several camera makers and AI companies have adopted it. The catch is that metadata gets stripped when images move through social platforms and messaging apps, which is exactly where the problem lives.
Platform labelling is improving too, slowly. But labels are applied inconsistently and can be worked around. None of this is a solution you can rely on today.
Which brings us back to the boring answer: the human check still matters most. Your own two minutes of scepticism outperforms every automated tool currently available to you.
Beyond Malaysia Day
The official advisory was framed around one holiday. The risk is not.
Think about the next twelve months. Hari Raya graphics. Deepavali greetings. Chinese New Year campaigns. Election season content. Disaster footage during monsoon flooding, where fake images of damage circulate within hours and complicate actual relief efforts. Fabricated quotes attributed to politicians, business leaders and religious figures.
Every one of those follows the same pattern: high emotion, fast sharing, low verification. The checklist does not change. Reverse search, read the text, verify the claim, slow down on anything sensitive.
We would add one more habit. When you catch yourself feeling a strong emotional pull toward sharing something, treat that feeling as the trigger to check rather than the reason to post. The stronger the reaction, the more likely something was engineered to produce it.
What we would actually do
If you take one thing from this, make it a rule rather than an intention. Intentions fail under time pressure.
Our suggestion: anything with a number, a date, a named person or a claimed event gets checked. Everything else, use judgement. That single filter catches most of the damaging material without turning your feed management into a full-time job.
Bookmark Google Lens on your phone. Learn to long-press an image and reverse search it in three taps. Once that is muscle memory, the ninety seconds becomes fifteen.
Generative AI is not going away, and honestly, most of what it produces for celebrations is harmless and rather nice. The problem is not the technology. It is that our sharing habits were built for an era when making a convincing fake was hard work. That era ended. Our habits need to catch up, and the good news is that catching up takes less effort than you think.