It’s a question I’m asked constantly: by clients, by colleagues, by friends, at parties, in almost any conversation where I mention what I do for work.
How can you tell if something was written by AI?
The question is usually asked as though there must be a trick to spotting AI text, as if there’s a word we always look out for, a rhythm or recurrent phrase that sets off alarm bells, or a red flag of a punctuation mark (I’ll never give up my beloved em dash, though I’ve embraced the unspaced en dash as an American in London). Everyone’s looking for some visible seam in the text where the machine gives itself away.
Sure, sometimes there is. AI-generated content has recognisable habits: clean but generic language, overly tidy structure, repetitive phrasing, cautious claims, abstract vocabulary, inflated corporate adjectives, an annoying fondness for explaining things the reader already understood three sentences ago.
However, as I’ve been thinking through my answer to this question more, I’ve found that the real difference between AI-generated text and human writing is rarely found in a single word, but rather in the thinking behind the words.
Generative AI can produce fluent, grammatical, plausible text at extraordinary speed. That’s what makes it attractive to so many users, but it’s also what makes it risky. Because, at a glance, the result can look finished when it’s only half-baked. Paragraphs are in order, tone may be broadly appropriate, grammar certainly clean, and the piece may even have those hallmarks of basic writing: a beginning, middle and end.
However, clean sentences aren’t the same thing as good writing.
Good writing isn’t just language arranged correctly. It’s the result of judgement: what to include, what to leave out, what to emphasise, what to challenge, what to prove, and what the reader should understand differently by the end of it.
For me, it boils down to this: AI generates text. Humans write.
Why this question matters now
AI content is getting harder to avoid. It’s turning up in reports, funding applications, proposals, academic papers, newsletters, LinkedIn posts, website copy, personal statements, donor communications, job applications, marketing materials, and thought leadership.
AI is taking up more space, but not building more trust
Analysis from Graphite.io suggests that primarily AI-generated content now makes up roughly half of online articles published in the datasets they studied. Their 2026 update showed that this amount reached 48% within 24 months of ChatGPT’s launch and has since plateaued around 50%. That doesn’t mean half of everything you read online is AI-generated, and the methodology depends on detection tools that have their own limits, but it does suggest something worth sitting with: AI-generated content isn’t marginal anymore. It’s part of the reading environment now, whether we like it or not.
And readers aren’t as indifferent to it as some people assume.
A 2025 cross-country study from the Reuters Institute at the University of Oxford found that people were far more comfortable with AI being used for behind-the-scenes tasks, like spelling and grammar edits or translation, than with AI assistance being used to write whole articles or rewrite content for different audiences. People might accept AI as a mechanical tool used for some elements of content creation, but it is hard to overcome the perception that natural language processing cannot grasp the emotional experiences or thought processes which lead humans to put pen to paper.
The closer AI gets to authorship, the more uneasy audiences become.
For anyone who uses writing to explain or represent themselves, this isn’t an abstract debate about how advanced AI affects the future of work. A proposal has to convincingly persuade another human and stand out from the crowd. A charity appeal has to build trust and inspire action. A student essay has to show depth of understanding. A personal statement has to sound like an actual person wrote it. A report has to establish authority. A website has to make an organisation feel credible and distinct. A leadership article has to offer a point of view, ideally an original one that could only come from them.
When writing has a job to do, the hollowness of generic writing is a liability.
companies (Image from Pixabay).
What are the signs of AI-generated text?
A warning first: no single word, punctuation mark or stylistic habit proves something is an AI-assisted draft or an AI-generated text.
So the better questions don’t ask ‘can I (or someone else) prove this was AI?’ They ask:
- Does this writing show evidence of someone having actually decided something?
- Does it make a claim?
- Does it know what matters and what doesn’t?
- Does it use evidence properly?
- Does it sound like someone thought the argument through, or like someone assembled the language they expected to need?
- Does it have the texture of real expertise, or just a passable impression of it?
The following is not a checklist you can run through to make AI-generated text sound more human, but rather is a collection of observations. These set off alarm bells for me that a piece of writing might actually be AI-generated text.
1. Generic language, in clusters
AI-generated writing gravitates towards words that sound professional, safe and broadly applicable. Its aims for a natural, human-like tone. The problem is never one word. It’s the cluster.
- ‘Landscape’, ‘realm’, ‘tapestry’, ‘framework’, ‘roadmap’, ‘journey’, ‘transformation’, ‘nuance’ – fine in the right context.
- ‘Leverage’, ‘foster’, ‘harness’, ‘navigate’, ‘elevate’, ‘underscore’, ‘empower’, ‘resonate’ – same.
- ‘Robust’, ‘seamless’, ‘dynamic’, ‘meaningful’, ‘comprehensive’, ‘groundbreaking’, ‘pivotal’ – also fine, on their own.
Put several of them in the same short passage, though, and the writing starts to feel inflated without getting any more precise:
- ‘By leveraging a robust framework, organisations can navigate today’s complex landscape and foster meaningful engagement across diverse stakeholders.’
The problem
That sentence is grammatically sound and sounds like business writing, but it says almost nothing.
What framework? Which organisations? What kind of engagement, with which stakeholders? What’s actually complex here, and what’s changed?
A human editor asks the questions that sentence is dodging. AI-generated content leads to the use of vocabulary that signals importance without providing tangible meaning. Writing has to earn that level of importance, with detail.
2. Vague preamble
AI loves to set the scene before it gets anywhere near the point.
- ‘In today’s rapidly evolving digital landscape…’
- ‘As organisations navigate an increasingly complex environment…’
- ‘In the modern world, effective communication has never been more important…’
The problem
None of these is technically wrong. They just delay the point, gesturing at relevance without saying anything specific, doing the work of an introduction before the writer has worked out what the piece is actually about (and wasting a reader’s time getting to the point).
A stronger opening tends to do something more deliberate: start with the question the reader already has, introduce a curiosity gap to hook the audience, offer a surprising fact, offer the problem outright, dig into a concrete example or scenario.
‘If your team has started receiving documents that look polished but still take an hour to make sense of, you may be dealing with AI-generated workslop.’
That opener gives the reader something to recognise immediately. It implies a problem without needing to remind anyone that the world is changing fast. They already know that.
3. Fluency that doesn’t move
This is the one I think of as fluency without force, and to me it’s often the strongest signal of the lot.
The prose glides. Paragraphs are balanced. Each sentence follows naturally from the last, but the argument ultimately leads nowhere. A point gets introduced, lightly expanded, restated, wrapped in a tidy closing line, and then the next paragraph does roughly the same thing. This may sound like a lot of writing on the internet already – due to years of search engine optimization squeezing and stretching certain forms of writing into recognisable patterns – but I promise there’s a meaningful difference for those of us who professionally read, write and edit this type of writing.
The problem
It’s a treadmill effect: motion, no progress.
That’s why AI-generated content can be exhausting to read even when nothing in it is technically wrong. It keeps going, but your understanding doesn’t deepen with it. You get to the end of a section having been carried along by competence rather than convinced by anything.
Human writing has pressure behind it. It selects and cuts. It decides this point matters more than that one, and isn’t shy about it. AI produces a sequence, while writers build an argument.
4. Tidy for the sake of being tidy
AI-generated text is often highly organised, which is useful for an outline, but becomes a problem the moment every idea gets forced into the same shape regardless of whether it fits.
- Numbered lists where prose would actually be more persuasive and compelling.
- Every section split into exactly three points – three benefits, three challenges, three recommendations – whether or not the material breaks down that neatly.
- Colon-heavy headings: ‘Content Strategy: A Guide to Building Trust’.
- A recap sentence at the end of every section, telling you what the paragraph above already made obvious.
None of this by itself proves that AI ‘writers’ have been at work. Lists can help, and headings can help, of course. A conclusion sentence can be necessary to clarify a point or transition to another.
The problem
The red flag is the mechanical regularity of it – structure replacing thought, rather than serving it. A list can dress up weak content as organised without making it mean more. Three points can create an impression of balance even when one of the three obviously matters more than the other two. A tidy recap can end up managing a reader and filling space on a page instead of engaging them and deepening an argument.
Human writing trusts the reader more than that. It doesn’t need to narrate every move or fit every idea into a bland, unnecessary formation.
5. Relentlessly even rhythms that flatten out the odd
AI prose tends to settle into a single rhythm: medium-length sentences, similarly sized paragraphs, a register that never really shifts. That can read as professional at first, but it flattens fast.
Human writing varies more. A short sentence lands a point. A longer one carries a complication all the way through to its natural end. A paragraph might be one line, if that’s what the idea needs.
A writer reaches for a contraction, a sharper verb, a less expected word or some other nuance because the voice of the piece demands it in that moment, not because a rule says so. The oddness and idiosyncrasy of human writing gets erased in the mechanical evenness and flattened prose of AI-generated text.
The problem
AI tends to smooth all of that original meaning away. You notice it most in copy that’s meant to sound conversational and somehow doesn’t.
Examples include awkward phrasing when there are no contractions where a person would naturally use one, a semicolon where nobody talks like that, a tone so even it’s empty and verging on hypnotic.
There’s nothing technically wrong with it, which, as an editor who loves grammatical correctness and clean rhythms, is not my issue with it. The issue is that it’s often been polished past the point of being interesting or compelling to a reader.
6. Hedging instead of deciding
One of the clearest tells of AI ‘writing’ is a reluctance to land anywhere or take a real stand. Depending on how they’re prompted, AI tools can flatten the boldness of ideas, or inflate and exaggerate claims, but generally the text tends to land somewhere in a balanced medium. It’ll acknowledge complexity, balance advantages against challenges, gesture at nuance and ultimately conclude that the answer depends on context. Sometimes that’s genuinely true. More often, it’s a way of avoiding the work of deciding.
Measured isn’t the same as non-committal. For example, the best business writing is usually careful about not alienating potential clients, sure, but it still says something to make their company stand out in the market and seem worth engaging with.
AI-generated text tends to perform balance instead: both sides offered, nothing decided and core claims softened or uninterrogated.
The problem
AI is trained to be helpful, balanced and broadly acceptable, which is useful for some jobs and useless for this one. When writing has a job to do, it needs to say something, not empty words wrapped up in a clean package.
Readers usually don’t just want information. They want to know what someone thinks and believe the credibility of their claims. Without that, the writing might still be readable, but it won’t be memorable and it won’t persuade anyone of anything.
7. Examples that feel like placeholders
AI can imitate specificity, but what it reaches for is usually a stand-in:
- a ‘busy professional,’
- a ‘modern organisation,’
- a ‘diverse range of stakeholders,’
- a ‘customer journey.’
Again, these are all technically plausible, but still never quite observed or grounded in the detail that grants writing real credibility. They apply to almost anyone, which means they land with no one in particular: the classic ‘jack of all trades, master of none’ trap.
Human writing carries the residue of someone having paid attention: a detail from a real client report, a feeling from lived experience, a phrase overheard in a meeting, a problem that came up three times in one week or a statistic that set the trajectory of the argument.
The problem
Although it’s tempting to speed through and skip through these elements, this specificity seriously matters. A charity team can’t build donor trust on generic warmth, a consultant can’t demonstrate expertise through abstraction, a student can’t show understanding by summarising around a topic instead of engaging with it.
AI speaks in hollow categories, but human writing has the author’s experience to ground the words in experience that resonates with readers.
8. Confident before it’s true
The factual risks here aren’t separate from the writing risks, and they’re often the same problem. The MIT AI Risk Repository now catalogues more than 1,700 risks associated with AI systems, including the generation of false or misleading information that people call ‘hallucinations’.
Generative AI can produce fabricated citations, invented statistics, misattributed quotes and claims that are plausible right up until you check them. It can present outdated information with total confidence, blur sources, get the sequence of events wrong, and hand you numbers that sound natural because they’re precise rather than because they’re true.
I’ve lost track of how often this has come up in work I’ve edited – it’s shocking to see how many instances of ‘J.’ initials in reference lists have been erroneously spelled out as ‘John’ because an AI tool generated the most likely name it could be.
A Scientific Reports study looked at ChatGPT-generated literature reviews in 2023 across 42 topics, analysing 636 bibliographic citations, and found that 55% of GPT-3.5 citations were fabricated outright, against 18% for the later model GPT-4. Even among those who did not specifically cite AI, a significant share still contained substantive errors. That should matter to anyone producing evidence-led content: students, researchers, report writers, etc.
There’s a workplace cost, too. Research covered by Harvard Business Review on AI-generated ‘workslop’ found that 40% of surveyed US full-time employees had received low-effort AI-generated work in the previous month, and that dealing with each instance took them an average of one hour and 56 minutes. The same research found people who sent that workslop were rated afterwards as less creative, capable, reliable, trustworthy and intelligent by the colleagues who received it.
That’s the part people miss when they’re tempted to rely on these tools: AI-generated text doesn’t always save time. It often just moves the work downstream, to whoever has to fact-check it or rewrite it before it goes anywhere near a client and risks damaging reputation.
The problem
When AI hands you a draft built on weak or invented support, the risk isn’t just that a source needs checking. It’s that the piece may be built around a claim that should never have been made at all. That’s why a light edit of AI-generated text can be more dangerous than it looks: the grammar probably improves, the obviously machine-like phrasing might even be weeded out by the right prompts, the tone can smooth to a client-friendly register, but the underlying weakness stays exactly where it was.
As I said at the start, the problem was never really the wording. It’s the thinking.
What AI detectors can and can’t tell you
Detection tools – such as Pangram, Grammarly, ZeroGPT or GPTZero – are evolving rapidly, just like AI models are. There are even AI text ‘humanisers’ – tools working across multiple languages that are designed to help people avoid AI detection systems. With or without these types of tools, it’s true that heavily edited AI text can bypass AI detectors, which are still far from perfect.
How AI detection tools work
AI detection tools generally look for statistical patterns: how predictable the text is, how much sentence length varies, whether certain vocabulary clusters appear, whether the prose has the kind of regularity that machine-generated language tends to have.
While a lot of these patterns line up with the observations I’ve gathered and shared here, AI detections tools tend to focus their assessments on two core elements of evaluation: perplexity and burstiness.
- Perplexity measures how predictable a text is to a language model – human writing tends to be less predictable, AI tends to choose the most statistically likely next word, then the next, then the next based on patterns it has learned from vast amounts of training data.
- Burstiness measures variation in sentence length and complexity across a piece. Human writing moves more between short and long, simple and complex. AI tends to sit in the middle and stay there.
However, AI detections systems get it wrong, too. The truth is that it is possible for a human to write mechanically. A human can overuse corporate language, open a blog with ‘in today’s fast-paced world’, or use ‘leverage’ three times in one paragraph without an algorithm in sight. Who knows, a corporate writer might even want to use the words ‘tapestry’ or ‘landscape’ because it speaks to them and expresses their idea best – I won’t make assumptions, it’s possible!
A heavily edited AI draft might sail through detection tools undetected, while a piece that a writer painstakingly crafted over months can score a high percentage. A formulaic human draft can get wrongly flagged, or someone writing in multiple languages can produce prose that reads as ‘more regular’ than a detector expects. A literary style built on deliberate repetition can confuse the whole system (and send a writer into an existential crisis about her own humanity).
How to interrogate a text beyond detection tools
While these AI detection tools raise useful questions, given their limitations and the ongoing evolution of all sorts of related products on the market, they don’t settle anything on their own. For now, they’re evidence to weigh alongside the writing itself, not a verdict.
In my view, the more useful test is editorial:
- Does the piece actually say something?
- Does it show real expertise and experience?
- Is the evidence used properly?
- Does the structure serve the point, or just dress it up?
- Does the tone fit who it’s for?
- Does it feel authored and thought through, or merely assembled?
What human writing does differently
Human writers do more than produce sentences. They decide what a piece is really saying. They notice when an argument is overclaimed, under-evidenced or structurally wobbly. They cut the throat-clearing or flag an exaggeration. They catch three sections saying the same thing in slightly different words. They know when a conclusion hasn’t been earned or isn’t achieving what the author set out to do, and they can hear when a paragraph is technically fluent but strategically empty.
That’s why a professional writer or editor matters most when content has to persuade, differentiate or carry real reputational weight.
- Executive communications need tone control and a sense of consequence.
- Reports need argument architecture, not just information laid out in order.
- Website copy needs to be distinctive and feel authentic to your business.
- Charity communications need sincerity and some emotional depth without slipping into sentimentality.
- Student and academic work needs evidence, clarity and a thought that’s genuinely the writer’s own.
- Thought leadership needs an actual point of view, not just the trappings of one.
AI tends to average language out and lacks the judgement and authority needed to achieve these goals. AI algorithms churn out pieces which have the same meaning, or, in other words, no meaning at all. Human writers and editors make writing sound like someone, not everyone.
That distinction matters for search, too. Google’s own guidance is clear that it doesn’t reward content for being human-made or punish it for being AI-assisted. What it asks is whether the work is genuinely helpful, reliable and people-first. Not ‘does this look polished?’, but ‘does this deserve to be trusted?’ That’s a fair standard, and a useful one to hold your own writing to, whoever or whatever drafted the first pass.
Where AI might still be useful
However you feel about AI, none of what I’ve said here means I’m declaring that AI has absolutely no place in a process for people trying to write, as long as you’re aware of the risks and consequences and making an informed decision about what you’re comfortable with. Used carefully and with rigorous validation, you might find that AI tools enhance efficiency on the back end of the operations that lead to written content. For some, it might even prove to be a decent assistant: summarising material, generating rough options, organising scattered notes, producing variations, helping someone get past the blank page when the blank page is the actual obstacle.
That’s not necessarily authorship, though. The moment a piece needs to persuade, differentiate, represent expertise or carry institutional credibility, the work has to sit firmly in human hands. Someone has to decide what it’s really saying, test the claims, shape the structure, validate the sources, protect the tone and challenge the easy phrase that doesn’t hold up under pressure.
That’s the difference between AI-generated text and human writing. AI tools can produce sentences that are clean, plausible and fast. However, human-written content has never just been the production of sentences. It is the expression of voice and opinion and expertise, with serious consequences for reputation. It’s the result of attention, experience, judgement and decisions during the writing process. Most of all, it comes from someone who knows what they think and what they mean to say.
For now, I think that part stays stubbornly human.
If you have a piece of writing that needs to sound credible, clear and genuinely human, get in touch by emailing [email protected]. Whether it’s a report, a proposal, a website, an article, academic paper, social media post or a piece of thought leadership, we can help make sure the writing does the job it’s there to do, with your voice intact and your judgement on the page.
