The sheer volume of AI output has outpaced human writing online, with the amount of AI-written articles outnumbering those written by humans back in November 2024. AI is being used everywhere with one magazine going so far as to ‘interview’ an AI fabrication of a celebrity instead of the real person.
But that’s just the AI writing which is disclosed upfront and publicly. Are writers being honest about their AI use?
A 2025 survey of 1,200 book authors found that nearly half of them use AI in their writing. Of those, not only do 81% use it to conduct research, but 74% do not disclose their AI use to readers. For journalists, 78% of respondents to a 2023 survey said their organisation had already used AI in some capacity that year. This use ranged from news-gathering to AI-generating the written content itself.
Despite the prevalence of AI use, however, several studies show that people struggle to identify AI writing, and are no better at it than random chance even after receiving training.
This article explores why accurate identification of AI writing is necessary, then discusses the methods of identifying AI content.
Why does this matter?
AI is notorious for its hallucinations – instances where it fabricates information like quotes or sources which don’t exist.
However, Townsen Hicks et al. argue that ‘hallucination’ is a misleading descriptor of this phenomenon. Instead, they suggest a word based on the philosopher Harry Frankfurt’s writing on truth, lies, and ‘bullshit’.
Frankfurt defines ‘bullshit’ as a statement made independently of whether or not it is true. A truth-teller knows the truth and seeks to present it accurately. A liar knows the truth and conceals it or leads people away from it. In contrast, the bullshitter is indifferent to the truth. The truth or falsehood of the content is irrelevant.
This is how AI operates. As Jordan Harrod writes:
“AI systems are fundamentally unconcerned with truth. The AI system is optimised to provide responses that appear helpful, harmless, and honest – but this appearance is in service of suitability for purpose, not accuracy.”
For an AI to be ‘hallucinating’ it would need to be faithfully describing what it ‘perceives’ to be true as an honest truth-teller which is merely mistaken. Instead, AI is indifferent to the truth. I will use Frankfurt’s term ‘bullshitting’ to refer to this feature of AI.
This isn’t a bug that can be fixed but rather a product of how AI operates. Every AI is a large language model (LLM), which works by statistically predicting which word comes next in a string of text. The LLM ‘learns’ the pattern for which words and phrases appear next to each other, then reproduces it. LLMs do this even if information is missing or vague, creating ‘bullshit’.
There are attempts to correct this by creating models which scour the internet for high quality sources of information, or use tools like an in-built calculator. But both of these fail: AI can’t always access high quality sources, nor judge how trustworthy said sources are (i.e. whether they are biased or exaggerated marketing claims). It also evades using tools.
Ultimately, AI remains indifferent to truth or fabrication. Another flaw is that it reproduces a range of biases, including regressive gender stereotypes, homophobia, and racial profiling.
Exploiting our biases
AI delivers this ‘bullshit’ in convincing ways, exploiting several biases in our brains which make us less able to critically judge the accuracy of an AI’s output.
Not only do users tend to trust AI output without checking it for errors, but those with a good understanding of how AI works and how to use it (high AI-literacy) are more likely to overestimate their ability to judge the AI’s accuracy and trust it without question.
Those errors then get published where they’re not only read by people but future AI models pull from this information too, creating a cycle of mistakes and nonsense information.
Ultimately, AI is far more likely to be inaccurate than human writing. Worse still, those using AI for research are likely to put too much trust in the facts an AI presents, even though AI output presents truth and falsity alike with equal confidence and importance.
AI writing is a threat to the quality and integrity of journalism, and our oversight for it needs to be better.
The folly of using AI-checkers
So, if AI writing is little more than compellingly written ‘bullshit’, how do we recognise it in order to place its statements and ‘facts’ under more scrutiny? AI-checkers seem like an attractive solution; they are convenient to use and offer a numerical value of AI-ness to humanness. However, using a machine to verify a machine is rife with error.
For starters, AI models are constantly being tweaked, updated, and upgraded to new models. Each time this happens, the pattern of their writing changes, and the outdated AI checker becomes even more inaccurate. It is a constant arms race between AI models and AI checkers.
Even worse, AI checkers are neither accurate nor reliable, often judging content as human when it is AI generated and vice versa. The latter leads to unfounded accusations of ‘AI writing’, creating distrust within organisations. This is especially detrimental where accusations of misconduct, plagiarism, and misattribution could harm a student’s academic record or stain a journalist’s reputation.
As with AI generators, AI checkers contain their own biases too. For example, they are more likely to flag writing by non-native english speakers as being AI. This is also true for neurodivergent and disabled authors.
AI checkers are also easily fooled. A 2023 study of 12 different AI checking tools found that accuracy fell from 74% to just 42% after AI writing had been edited in minor ways. This obfuscation doesn’t even need to be done by hand; many sites offer automatic ‘AI humanisers’ alongside their AI detectors.
AI humanisers and why they’re a problem
An AI humaniser is an online software tool which promises to take the hallmarks of AI writing and replace them with more ‘human-sounding’ traits in order to fool AI detectors. Similar to how an AI checker is used, a user need only paste their text into the AI humaniser, click a button and let the online tool edit it for them.
Beyond just online tools, nowadays many freelance editors also specifically advertise their services as humanising AI writing.
Predictably, AI humanisers aren’t reliable either, often returning mixed results. Some users report that their writing was rated more likely to be AI after using a humaniser. Because AI humanisers work by changing phrasing and structure, accurate details or nuance can be lost. This problem is additional to the ethical dilemmas of plagiarism and academic/journalistic integrity.
Key traits of AI writing to look out for
So if we can’t rely on AI checkers, how can we reliably spot AI writing? This is far from an exact science, and aside from a few very blatant tells it is unlikely anyone can identify an AI text with any certainty.
That said, AI models operate on patterns, which result in a trend of traits – these traits are covered extensively in Wikipedia’s own Signs of AI Writing article. This is perhaps the current most comprehensive guide to identifying AI writing. It includes a demonstration of each trait using example text. It also covers the nuance of specific models (e.g. ChatGPT versus Claude) and model versions (e.g. GPT-4 versus GPT-5).
Perhaps most importantly, the Wikipedia article is kept up to date. When an AI model is next updated, Wikipedia will quickly adjust their guide to reflect the new changes.
A more technical guide to spotting AI writing will follow in future; for now, some traits of AI writing are as follows:
- Heavy use of em-dashes (—), particularly where a comma would make more sense.
- Curly quotation marks, especially if mixed with straight quotation marks.
- Emojis to decorate headings or as bullet points.
- Phrases which follow a ‘It’s not X, it’s Y’ pattern, including ‘No X, no Y, just Z.’
- Rule of three: ‘adjective, adjective, adjective’ or ‘short phrase, short phrase, and short phrase’.
- Heavy use of AI vocabulary words
- Overstated importance, especially for mundane or minor subjects.
- Fake quotes and citations (i.e. ‘hallucinations’).
- Instructions or comments for the AI user which were accidentally copied along with the main text, e.g. ‘Here’s a template for you, copy and paste it to edit further.’ or ‘As of my last knowledge update’.
Caveats and vigilance
Ultimately, AI is trained on text written by humans so its writing traits are reproductions of human traits. The presence of any AI writing traits needs to be judged in context rather than taken as standalone proof.
Questions to consider are: Is the author invested in their work and do they understand what they’ve written? Are they producing an impossibly large amount of writing, more than a person could reasonably make in that time frame? Did they only start writing prolifically after the rise of AI (roughly 2024 onwards)?
As readers, we need to carry this vigilance with us when reading any text, both online and printed, across all fields of writing from fiction novels to scientific journals.
This vigilance is necessary because:
- AI checkers are inaccurate and prone to becoming out of date.
- AI writes in a way that sounds compelling, exploiting our biases to make us trust it without question.
- AI has no concept of truthful or factual information; its only purpose is to appear convincing.
Vigilance is needed because, if you’re like me, you don’t want to be fed or regurgitate ‘bullshit’.
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.
