Last month, a marketing agency I know sat in a review meeting staring at three freelancer submissions. All three read smoothly. All three sounded slightly, indefinably off. Nobody in the room could say for certain which, if any, had been written by ChatGPT.
That scene is playing out everywhere in Singapore right now. HR teams screening a flood of polished cover letters, lecturers marking student essays, content leads vetting outsourced blog posts: everyone is quietly asking the same question, and nobody has a confident answer.
So they paste the text into an AI detector, get a scary “98% AI-generated” score, and act on it. The problem is those tools give confident answers that are frequently wrong, and a false accusation against a real writer or a real applicant is a genuinely serious mistake.
This article is about how to detect AI writing the honest way. I will not hand you a magic tool that “just knows”. Instead, I want to teach you the actual tells of machine-written text, show you why detector scores deserve your suspicion, and give you an editorial process that holds up whether or not AI was involved.
Let me set the scene. A freelance writer submits an SEO blog for a local nasi lemak brand in Tampines, and something feels off before I have even finished the first paragraph.
The words are technically correct, the grammar is clean, but the whole thing reads like it was assembled rather than written. That gut feeling is usually where how to detect AI writing begins, and in this section I want to turn that instinct into a checklist you can actually use.
Before we start, one honest caveat: no single marker below proves anything. It is the cluster of signs appearing together that tips you off, and I will come back to why that matters.
The first place I look is word choice, because large language models have favourite words they reach for again and again.
You will see “delve”, “tapestry”, “landscape”, “realm”, and “testament to” popping up far more than any human writer would naturally use. Openers like “in today’s fast-paced world” or “in the ever-evolving digital landscape” are almost a signature.
For that nasi lemak blog, a real writer might say the sambal has a stubborn kick that lingers. The AI version said the dish “offers a delightful tapestry of flavours that tantalise the palate”, which tells you nothing and everything at once.
Watch for hedging too. AI loves to cover itself with phrases like “it is important to note”, “generally speaking”, and “can vary depending on several factors”, which softens every claim into mush.
The sentences also tend to be suspiciously symmetrical. You get pairs of clauses balanced against each other, “not only this, but also that”, over and over until the rhythm feels engineered rather than spoken.
Once the vocabulary raises my eyebrow, I zoom out and look at the shape of the piece.
Human writing breathes unevenly. Some paragraphs run long because the idea needed room; others are one sharp line. AI output often marches along in near-identical blocks, three or four sentences each, paragraph after paragraph.
Then there are the tidy listicles. AI adores a clean numbered list where every item has the same length and the same “bold label, then explanation” format, because that pattern is easy to generate and hard to get wrong.
Transitions are another giveaway. Real writers jump, backtrack, and interrupt themselves. AI glides through “firstly”, “moreover”, “furthermore”, and “in conclusion” like a student who memorised the connectives sheet.
If a whole article feels frictionless, with no digressions and no messy human texture, that smoothness is itself a signal.
The deepest tells are not about how the text sounds but about what it actually knows.
AI-generated content tends to stay on the surface. It will tell you that content marketing “builds brand awareness and drives engagement” without ever explaining how, for whom, or what happened when someone tried it.
Watch closely for invented statistics. A line like “studies show 73 percent of Singaporean consumers prefer local F&B brands” often has no source, no year, and no link, because the model produced a plausible-sounding number rather than a real one. Always try to trace a figure back to a named body such as IMDA or Enterprise Singapore, and be suspicious when you cannot.
The biggest tell of all is missing lived experience. A genuine writer for that Tampines stall might mention the queue that snakes past the coffee shop on weekends, or a specific SkillsFuture course a staff member took. AI cannot fake that local specificity because it was never there.
So it defaults to generic claims that could apply to any brand in any country. If a piece about a Singapore business never names a real place, scheme, tool, or person, treat that absence as a warning. The ChatGPT user guide for the Singapore workforce is a good companion read here, because understanding how the tools generate text helps you spot where they fall short.
Here is the part too many people skip. Any one of these signs on its own is not proof of AI, and acting like it is will get you into trouble.
Plenty of skilled human writers use “moreover” or the occasional balanced sentence. What flips my judgement is seeing the vocabulary tells, the uniform rhythm, and the hollow, sourceless claims all stacked together in the same piece.
I also want to flag a real fairness risk. Non-native English writers, including many talented professionals in Singapore, often write in cleaner, more formal, more structured English precisely because they learned it carefully.
That polish can look robotic to a careless reviewer, so never accuse someone based on style alone. Use these signs to raise questions and start a conversation, not to hand down a verdict.
Before you paste a suspicious paragraph into a detector and treat the result as gospel, it helps to know what the tool is really doing under the hood.
When I first ran a colleague’s genuinely human blog post through a popular detector, and it came back flagged as “likely AI”, I realised most people misunderstand these scores entirely. So let me demystify how to detect AI writing using detectors, and where they quietly fall apart.
Most detectors lean on two statistical ideas: perplexity and burstiness.
Perplexity measures how surprised a language model is by your next word. If the text uses predictable, common word choices, perplexity is low, and the tool leans towards “AI-generated” because that is exactly how a model writes by default.
Burstiness measures variation across sentences. Humans tend to mix a long, winding sentence with a short, punchy one, while raw AI output often keeps a steady, even rhythm.
Here is a worked mini-example. Imagine a Singapore government officer drafting a clear, plain-English notice for citizens: short sentences, common vocabulary, no fancy metaphors, consistent structure because that is good public communication.
That writing is genuinely human, but it looks statistically similar to AI: low perplexity, low burstiness. A detector may flag it. The person did nothing wrong; they just wrote in a clean, controlled style that the maths cannot distinguish from a model.
The second common approach is a classifier. Tools like GPTZero and Originality.ai are trained on large piles of text labelled as either human-written or AI-written, then they learn the patterns that separate the two.
Turnitin’s AI indicator, which students at NUS and NTU encounter when they submit assignments, works on a similar principle applied to academic writing.
This is the part I wish more people in Singapore understood before making decisions based on a number on a screen.
A detector does not tell you the truth. It estimates a probability based on statistical patterns, and a probability is not proof.
When a tool says “85% AI”, that percentage is easy to misread. It is not a claim that 85% of the text was written by AI, and it is not an 85% guarantee that the passage is AI-generated.
Depending on the tool, that figure usually reflects the model’s confidence that the text resembles the AI-written samples it was trained on. Different tools calculate and phrase it differently, which is why the same paragraph can score wildly across detectors. Understanding how that number is calculated is one thing. Understanding why it so often gets the answer wrong in practice is another question entirely, and that is where we turn next.
Before you paste anything into an AI detector and treat the score as gospel, I need you to understand something: those tools get it wrong far more often than their marketing suggests.
When people ask me how to detect AI writing, the first thing I say is that no detector gives you a verdict. It gives you a probability, and that probability is built on shaky ground.
Let me paint a scenario I have seen play out in Singapore. A polytechnic student submits a well-structured essay, the lecturer runs it through a detector, and it flags 80 percent AI. The student swears they wrote every word, but now they are defending themselves against a machine that cannot explain its reasoning.
That is not a fringe case. It is the predictable result of how these tools actually work.
Detectors work by measuring how predictable your word choices are. Simple, clean, grammatically careful writing looks statistically similar to AI output, because both tend to avoid quirky, unexpected phrasing.
Here is the problem. That is exactly how many strong non-native English speakers write in Singapore; a 2023 Stanford study put this to the test directly. Researchers ran seven popular AI detectors against TOEFL essays written by non-native English speakers and against essays written by US eighth-graders. The detectors were nearly flawless on the native-speaker essays, but misclassified more than half of the non-native speakers’ essays as AI-generated, with nearly every essay flagged by at least one tool. James Zou, the study’s senior author at Stanford, has said the bias traces back to how detectors measure “perplexity,” a score that non-native writers naturally trend lower on simply because of how they structure sentences. That is a real risk in Singapore, where English is often a second or third language alongside Mandarin, Malay, or Tamil.
Picture a hiring manager reviewing a cover letter from a talented applicant who learned formal English in school. The writing is measured and correct, so a detector flags it, and a genuinely good candidate gets binned on a false positive.
Neurodivergent writers can trigger the same result. Someone who writes in structured, methodical sentences, perhaps to keep their thoughts organised, produces text a detector reads as machine-generated.
In both cases, the tool is not catching cheating. It is penalising a particular writing style, and that should worry anyone using these scores to judge real people.
Even if false positives did not exist, detectors would still fail for the opposite reason: anyone actually using AI can slip straight past them.
The workarounds are trivial. A few minutes of light editing, running the text through a paraphrasing tool, or feeding it into a “humaniser” service will collapse the detector’s confidence.
So think about what this means in practice. The careful, honest writer gets falsely accused, while the person deliberately gaming the system walks through untouched.
A tool that punishes the innocent and rewards the evader is not a detection tool in any useful sense.
The third failure mode is time itself. Detectors are trained on the output of older AI models, but the models keep getting better at sounding human.
Newer systems like GPT-4o and Claude produce text with far more natural rhythm and variation than earlier versions. Every time a model improves, the detectors trained on yesterday’s output fall further behind.
This is not a bug you can patch. It is a permanent gap, because the writing keeps evolving faster than the tools chasing it.
The most telling admission came from OpenAI itself. The company quietly retired its own AI text classifier, citing its low rate of accuracy (OpenAI, 2023), and that was OpenAI judging its own product. When the maker of the models cannot reliably detect its own output, you should be deeply sceptical of any third party claiming they can.
So where does this leave you? Use detectors as one weak signal among many, never as a judge.
If you are assessing content for trust and quality, your energy is far better spent on the human signals that actually matter. Understanding E-E-A-T and how Google evaluates trust will teach you more about content quality than any detector score ever will.
The rule I hold to is simple. Never make a high-stakes decision, an academic penalty, a hiring rejection, or a publishing ban, on a detector score alone.
If you suspect AI, talk to the writer, ask about their process, and look at the reasoning behind the words. A number from a tool that OpenAI itself could not trust is not evidence. It is a prompt to look closer, nothing more.
Detection software gives you a probability score, not proof. When I train content teams here in Singapore, I tell them the same thing: the AI checkers are a first filter, but your own judgement is the real test.
Picture a small content team at a Singapore agency vetting a batch of outsourced articles before they go live on a client site, say a fintech or a clinic. The freelancer swears everything is original. Your job is to decide whether it earns a byline on your client’s domain.
That decision is where how to detect AI writing stops being a tool question and becomes an editorial one. Here is the repeatable method I use.
The fastest way to expose AI writing is to chase every claim to its source. Generative models are fluent, but they hallucinate citations, invent study findings, and attribute quotes to people who never said them.
So take each statistic in the draft and try to find the primary source. If a paragraph claims that a certain percentage of Singapore SMEs adopted a particular technology, look for the actual IMDA or Enterprise Singapore report it supposedly came from.
When the trail goes cold, you have your answer. A confident number with no verifiable source is one of the strongest tells that a machine, not a researcher, wrote the sentence.
Watch especially for links that look plausible but return a 404, or quotes attributed to a named official that appear nowhere in any real speech or press release. Human writers make mistakes too, but they rarely fabricate a source from scratch and then link to a page that does not exist.
Run through these checks on every draft:
If the fact check leaves you unsure, go to the writer directly. This is the single technique that no detector can replicate, and it is brutally effective.
Ask them to explain their choices. Why did they structure the piece this way, why did they pick that angle for a Singapore audience, and where did that particular claim come from?
Then set a live challenge. Ask them to add a local anecdote from their own experience, or to expand one section on the spot without touching a chatbot.
Someone who genuinely wrote the article can talk you through it and add colour easily. Someone who prompted a model tends to stall, repeat generic points, or produce a second block of the same polished but hollow prose.
I have seen a writer confidently submit a piece on PDPA compliance, then go completely blank when asked to name one real scenario where a local business tripped up on consent. That silence tells you more than any percentage score.
Finally, read the draft the way a discerning reader would, not the way a scanner does. You are hunting for authenticity, because verifiable specifics are far harder to fake than fluent prose.
Ask whether the piece references real Singapore context: named people, actual places, specific dates, the name of a real grant like the Productivity Solutions Grant, or a recognisable local case. AI drafts love to stay safely general, praising “the vibrant Singapore business landscape” without ever committing to a concrete detail.
First-hand experience leaves fingerprints. A human who has actually filed a SkillsFuture claim or sat through a WSQ course will drop in small, lived details that a model has no way to know.
This is exactly what Google’s helpful-content system and E-E-A-T guidance reward: genuine experience, expertise, and trustworthiness rather than surface fluency. If you want the wider picture on why this matters for ranking, I cover it in our guide on how to optimise for Google AI Overviews and AI Mode.
The reassuring part is that these three checks reinforce each other. When the sources hold up, the writer can defend their choices live, and the voice carries real local specificity, you are looking at content worth publishing under your client’s name.
Let me flip the perspective, because most people reading this are not trying to catch cheaters. You are a writer or a business owner who uses AI to get a first draft down faster, and you want the final piece to sound like a person actually wrote it.
That is a completely legitimate use. The goal of editing is not to fool a detector. It is to add the value, judgement, and lived experience the model simply does not have.
I want you to picture a real scenario. A small business owner in Singapore selling handmade candles asks ChatGPT to draft a Shopee product page and a monthly newsletter. What comes back is grammatically perfect and completely forgettable.
Here is how I would turn that draft into something worth reading.
The first pass is a deletion pass. AI loves to warm up with phrases like “in today’s fast-paced world” or “when it comes to scented candles.”
Strip those out. Your candle owner should open with the actual thing a customer cares about: “This soy candle burns for around 40 hours and does not leave soot on your ceiling.”
Also hunt for hedging words like “can help,” “may provide,” and “is designed to.” If the candle clears cooking smells from an HDB kitchen, say so directly instead of tiptoeing around it.
I explained earlier that AI writes in a flat, even rhythm where every sentence is roughly the same length. Editing fixes this fast.
Read the draft aloud. Where you hear a metronome, chop a long sentence in two, or fuse two short ones. Add a fragment for emphasis. Ask a real question the reader is thinking.
That uneven cadence is one of the biggest tells of human writing, and it is the easiest to restore.
This is where the real value comes in, and it is the part no model can fake for you.
The AI does not know that your candle owner tested twelve wax blends in her Tampines flat, or that customers kept asking for a pandan scent for Chinese New Year gifting. Those specifics are gold.
AI drafts often make confident claims with nothing behind them. Replace vague statements with facts you can stand behind.
If you cite a delivery timeline, use your actual courier’s stated range. If you reference a safety standard, name the real body, such as guidance from local consumer safety authorities, rather than a made-up figure.
Never let the model invent a statistic, and never attribute a quote it produced to a real person. If you are not sure a number is accurate, describe it qualitatively instead.
The final layer is voice. Your newsletter and your Shopee page should sound like the same person, with the same warmth, humour or bluntness throughout.
Keep a short style sheet: three words that describe your tone, a few phrases you always use, and a few you never use. Run every AI draft through that filter so the machine’s neutral register gets replaced by yours.
The best AI-assisted content is a human framework filled in by a model, not a model output lightly touched by a human. That distinction is everything.
This is the same principle we teach when we help people build a content marketing strategy that boosts your SEO: use AI to draft and speed things up, then invest your human effort where it compounds, in genuine expertise and experience.
Mass-producing thin AI pages will not build authority. A smaller number of edited, expert pieces that carry your real voice will, and that is the workflow worth adopting.
Once you know how to detect AI writing, the harder question is what you actually owe the people reading your work. In my experience, the anxiety readers bring to this topic is rarely about the mechanics of detection. It is about doing the right thing without getting burned.
Let me be direct: transparency and sound judgement matter far more than beating any detector. A piece that quietly used AI but is accurate, well cited, and clearly disclosed is on firmer ground than one that scraped through a checker but misled its audience.
The honest answer is that it depends on context, but the safe default is to ask before you assume silence is fine.
If you are a freelancer writing for a Singapore SME, I would raise it during the scoping conversation rather than after delivery. A simple line in your proposal, something like “I use AI tools to draft and research, then edit and fact-check every piece myself”, sets expectations and protects the relationship.
For students, the stakes are sharper. If you are submitting an assignment, disclosure is not an optional courtesy; it is often an academic integrity requirement, and getting it wrong can mean a formal misconduct case.
Here is a rough hierarchy I use when deciding how much to disclose:
There is no single national rule, and that trips people up. Our local universities have taken visibly different stances.
NUS, NTU and SMU each publish their own AI-use guidance, and these vary by faculty and even by individual course. One lecturer might permit AI for brainstorming but ban it for the final write-up, while another allows AI drafting as long as you cite it. I always tell students to check the specific course outline, not a general university statement they found online.
Workplaces are just as fragmented. Some Singapore agencies actively encourage AI to speed up drafting, while a bank or law firm may restrict it heavily on confidentiality grounds.
Blanket bans, in my view, tend to backfire. When an organisation forbids all AI use without offering an approved alternative, people simply use it quietly on personal devices, which removes any oversight and makes the risk worse. A clear “here is what is allowed, here is what is not, and here is how to disclose it” policy protects everyone better than a flat no.
This is the part I see professionals overlook most often, and it is the one with real legal teeth in Singapore.
The moment you paste a client’s customer list, an unpublished report, or personal particulars into a public AI tool, you may be disclosing that data to a third party. Under the Personal Data Protection Act (PDPA), if that information includes personal data, you could be breaching your obligations without ever intending to.
Picture a marketing executive who drops a spreadsheet of enquiry emails and phone numbers into a free chatbot to “clean it up”. That is a genuine PDPA risk, because those individuals never consented to their data being processed by an external model.
My rule of thumb is simple: never paste anything into a public model that you would not be comfortable seeing screenshotted. If you must use AI on sensitive material, anonymise it first, or use an enterprise tool with a proper data processing agreement.
Singapore’s broader posture actually supports this measured approach. IMDA’s Model AI Governance Framework frames AI adoption around transparency, accountability and human oversight rather than fear, which is exactly the mindset I would encourage you to carry into your own work.
So the takeaway is not “avoid AI” or “hide AI”. It is to use it openly, protect the data you touch, and let your own judgement, not a detector’s score, be the thing you stand behind.
Here is the uncomfortable truth I have arrived at after years of reviewing content: the obsession with how to detect AI writing is largely a distraction.
If you are running a lean agency in Tanjong Pagar or an in-house marketing team producing forty blog posts, fifty social captions, and a monthly newsletter, you cannot police every draft through a detector. You will burn hours arguing over false positives while the real problem, thin and unoriginal content, sails through untouched.
The better question is not “was this written by AI?” It is “is this good, accurate, and worth publishing?” A strong editorial process answers that reliably. A detector never will.
Most weak content is not weak because a machine touched it. It is weak because nobody defined what good looked like before the writing started.
When I brief a piece, I spell out three things up front: the voice we want, the sources the writer must actually cite, and the level of expertise the reader expects. For a fintech client explaining PDPA obligations, “sounds professional” is useless as a standard. “Cites the PDPC advisory guidelines, uses a real onboarding scenario, and reads like a compliance manager wrote it” is something you can hold work against.
Write these standards into a template every contributor sees. Cover:
Once that brief exists, it does not matter whether the first draft came from a person or a co-pilot. The bar is the same for both.
Detectors invite a lazy habit: you paste text, get a score, and treat the tool as the gatekeeper. That outsources judgement to software that is frequently wrong.
I would rather assign a named human as the accountable reviewer for every piece. Someone owns the fact-check. Someone owns the final sign-off. When a claim is unverified, or a section reads like generic filler, that person is responsible for fixing it, not a percentage on a dashboard.
In a small Singapore team, this can be simple. A junior marketer drafts with AI assistance, a senior reviewer verifies every statistic and every claim about a scheme or regulation, and a manager approves before it goes live. The chain matters more than the tooling.
This is also how you protect your brand. If a post misstates a SkillsFuture eligibility rule, no detector would have caught that. A human who knows the subject would.
The teams that produce the best content are not the ones that ban AI, nor the ones that let it write unsupervised. They are the ones whose people know how to direct it.
That is a skills gap, and it is one you can close deliberately. In Singapore, SkillsFuture-funded upskilling makes this genuinely affordable, so cost is rarely a real barrier.
At Equinet Academy, we built our AI-assisted content and content marketing training around exactly this idea: use AI to accelerate research, outlining, and first drafts, then apply human strategy, expertise, and editing to make the output original and trustworthy. When a whole team shares that baseline, your average quality rises far more than any detector could deliver.
Skilled teams also plan content around outcomes rather than volume. If you want the wider picture of how this connects to results, we set out a blueprint for sustainable digital marketing ROI that treats content quality as an investment, not a checkbox.
So stop asking software to tell your good content from your filler. Process, expertise, and genuine originality do that job better every single time, and they scale with your team in a way that no detection tool ever will.
If you take one thing from this guide on how to detect AI writing, let it be this: your judgement matters more than any tool’s verdict.
You can absolutely learn to spot the tells. The hollow specifics, the tidy tricolon structures, the fondness for words like “delve” and “landscape”, the confident claims with nothing real behind them.
But those signals are clues, not proof. I have seen genuinely human writing flagged as AI, and polished AI text sail through with a clean score.
That is why I would never let a detector percentage make a high-stakes call on its own, whether you are marking a student’s essay, vetting a freelancer, or approving a case study for your company blog.
Trust does not come from a number. It comes from verifiable specifics you can check, honest disclosure about how something was made, and a strong editorial process where a real person owns the final draft.
Ask yourself the questions we covered. Does this cite named schemes, real cases, or figures you can verify? Would a subject expert nod along, or wince? Can the author explain their sources?
Here is the reassuring part. AI is not something to fear. Used well, it is a drafting partner that frees you to focus on the thinking, the examples, and the voice that only you can bring.
The professionals who thrive are the ones who can create and edit AI-assisted content that holds up to scrutiny, not the ones hiding from the tools.
If that is the skill you want to build, our WSQ AI in Digital Marketing Course is a natural next step. It is SkillsFuture-supported and designed for working professionals in Singapore who want to produce content that is credible, strategic, and unmistakably their own.
Learn to write and edit with confidence, and the detector score stops being something you chase.
Article Written By
Micah is a passionate content marketing strategist at Equinet Academy who loves turning keyword research into clear, purposeful content plans built around what people are actually searching for. She focuses on creating people-driven blogs and resources that help the company grow while making sure readers genuinely learn something useful and feel more confident applying it.
Article Written By
Micah is a passionate content marketing strategist at Equinet Academy who loves turning keyword research into clear, purposeful content plans built around what people are actually searching for. She focuses on creating people-driven blogs and resources that help the company grow while making sure readers genuinely learn something useful and feel more confident applying it.
Receive the latest blog articles right into your inbox.
Reader Interactions