How to spot AI-generated writing

According to a 2026 analysis by Graphite, half of articles published online are primarily AI-generated. Since its launch, generative AI has been used to spread hallucinations, discriminatory biases, and ‘slop’ across everything from celebrity interviews to scientific papers.

How can we trust what we read? We need to know when to scrutinise text more closely and resist authority bias. We need to be aware of the signs that point towards an article having been written by AI. This article serves as a guide.

AI’s shortcomings

As discussed in more detail in my previous article, AI is inaccurate not merely through mistakes or distortions of the truth, but through indifference to it. It generates patterns of words without any regard for how truthful or accurate those sequences are – and how could it? AI doesn’t know the meaning of the words it generates, only their order in a pattern. Whenever AI is correct, it is so without understanding whether that information is true. AI isn’t optimised for accuracy, but rather to appear helpful and harmless.

Although some may argue that AI is helpful, it is far from harmless. Besides generating errors, AI also reproduces biases, including regressive gender stereotypes, homophobia, and racial profiling.

Furthermore, AI uses confident language, sounding knowledgeable and authoritative. This triggers authority bias (to which every human is susceptible), making readers more trusting of AI’s statements. The very language AI generates makes us less able to question whether its ‘facts’ are accurate. This is especially dangerous because AI generates patterns, not accurate facts, and reproduces discriminatory biases. This is what we are trusting and believing without question.

We need to recognise AI writing in order to doubt its statements. We can’t let authority bias lead us into blind trust. We need to critically evaluate AI writing, which begins with identifying it first.

Common signs of AI writing

Although several studies show that humans struggle to accurately identify AI writing, being aware of how it presents remains invaluable.

This Wikipedia page is perhaps one of the most comprehensive guides to spotting AI writing. It provides an in-depth breakdown of AI writing traits, along with an example of each trait in writing. The article is also kept up to date, ensuring accuracy as new models are released.

The traits most relevant to readers are covered below. Each trait is hyperlinked to the corresponding section of the Wikipedia page for access to more detail and example text.

Linguistic quirks and nonsensical formatting

These are perhaps the most well-known AI ‘tells’. These include:

There is also a collection of ‘AI vocabulary’ words that AI models overuse. This varies from model to model (e.g., ChatGPT versus Gemini), and changes over time as models are updated (e.g., GPT-4 to GPT-5). The Wikipedia article lists these in detail, ordered by AI ‘era’.

For example, ‘delve’ was heavily used by ChatGPT-4 (active between 2023 and mid-2024), but the AI vocabulary of ChatGPT-5 (during 2025) includes the words emphasizing, enhance, highlighting, and showcasing.

AI also tends to generate unnecessary, small tables for information that would be more sensibly presented in regular sentences, and overuses boldface text for emphasis.

Overly wordy

AI is prone to repetition. Models overcorrect for this with elegant variation, using elaborate synonyms instead of repeating the same word. For example, when writing about a book author, instead of using “the author” or their name multiple times, further references might use “the writer”, then “the award-winner”, then “the acclaimed author”, then “the innovator of fiction”, and so on.

However, many human writers also exhibit this, as evidenced by this 2010 article from The Guardian (well before the popular use of AI) which mocks carrots being described as “popular orange vegetables”, and other elegant variations.

This effect may also be masked if the AI-user has generated different paragraphs separately for a piece of writing.

Dramatic

AI writing often places undue emphasis on a subject’s legacy and importance, even when the topic is mundane or unimportant. Vague but important-sounding phrases will often link the subject to broader themes, such as “this marks a pivotal moment”, or “plays a key role”, or “contributing to the …”.

Similarly, AI tends to exaggerate a subject’s notability and media coverage, using expressions such as “featured in”, “[list of media] and other media”, or “active social media presence”. These may be accompanied by superficial analyses, often with vague attributions that do not actually support that opinion.

Vague and generic

AI models operate on patterns, tending to produce the most statistically likely phrases that apply to the broadest spectrum of circumstances. Specific and nuanced facts are much rarer – especially those about specific topics – so they are underrepresented in an AI’s training data. As a result, instead of providing a specific fact, the AI often generates generic (and usually positive) statements.

This tendency often couples with overstated importance, resulting in vague but exaggerated descriptions. For example, “inventor of the first train-coupling device” might become “a revolutionary titan of industry”. Wikipedia likens this to a painting of a figure that becomes blurrier and less distinct, while simultaneously proclaiming ever more loudly that the figure depicted is important.

Inhuman errors

The most obvious of these is when AI-generated text includes communication intended for the user. This can occur if the user copies the text wholesale by mistake, or fails to remove such communication during editing.

Communication for the user might include instructions or summaries of the generated text’s contents, such as “Here’s a template for your wiki user page. You can copy and paste this onto your user page and customize it further”, or a “[topic] explainer: …” paragraph. AI users might also inadvertently leave in an AI’s knowledge cut-off disclaimer, for example, “as of my last knowledge update”.

Another type of error is the use of fake quotes and references. Although this can be a pretty blatant tell, sometimes faulty citations are due to human error. A hyperlink might fail because the author accessed it through their university library (which is inaccessible to those outside the university), or because it was copied and pasted with the start or end missing (a uniquely human mistake).

Ineffective indicators

There are many popular indicators of AI that are ineffective and may even suggest the opposite. It is important to be aware of these to avoid making false accusations and to develop a more accurate understanding of how AI writing presents.

Perfect grammar does not indicate AI usage; authors and editors may be highly skilled, perhaps with a professional background. A mix of formal and casual language styles might indicate neurodivergence, experience in technical or academic writing, or youthfulness. Perhaps counterintuitively, AI writing often tends to be verbose and dramatic, rather than being bland or robotic.

Context is key

As noted in my previous article, AI is trained on human writing, so the traits it reproduces are human traits. The presence of any of the above characteristics is not definitive proof, but merely an indication to proceed cautiously. Text should always be examined in context.

Some questions to consider are:

• Is the author invested in their work, as would be expected from someone who has dedicated time and effort to writing it?
• Do they understand what they’ve written if asked to explain it?
• Are they producing an implausibly large volume of writing, more than a person could reasonably create in that time frame?
• Did they only start writing prolifically after the rise of AI (roughly 2024 onwards)?

Keeping up with AI

When a new AI model is released or updated, its traits change. It is important to seek recent and specific advice on what to look out for.

Even with all this, you’re still likely to misidentify AI writing quite frequently, so do not rush to judgement. Instead, ask the author questions about their writing’s content and their writing or research process. First drafts, plans, and file version histories are useful evidence, but be wary of reverse-engineered first drafts.

Regardless of AI traits, remember to always evaluate critically what you are reading. Use contextual clues and fact-check claims by consulting other sources on the same subject. All the world’s knowledge is at our fingertips – unfortunately, so is all the world’s misinformation.

Comments? We want to hear them. Please write to editor@westenglandbylines.co.uk

Friends of Bylines Network

There has never been a greater need for grassroots journalism that investigates the stories that really matter, holds power to account and champions the voices of everyday citizens. We are proudly powered by volunteers but what we do isn’t free.

STAND WITH US for independent, citizen-led journalism that makes democracy stronger, and you will even get some exclusive benefits.

BECOME A FRIEND