When it comes to transparency in artificial intelligence (AI) there are different aspects to consider. One is understanding the reasoning process behind AI decisions and what data they use in that decision-making. Given that artificial neural networks are black boxes, that is a topic in itself, and will be the subject of another blog.
This blog talks about something more basic – how can you tell whether a document or image that you see is written by a human or is actually AI-generated? Large language models (LLMs), the underlying technology behind chatbots like ChatGPT, Claude, Grok and Perplexity are quite fluent in their communication. Their writing is certainly coherent and human-like. So how do you know whether that essay submitted as homework was written by Carol, as she claims, or was actually knocked off in a few seconds by an LLM? There is a whole industry of tools claiming to be able to detect AI writing, an industry valued at about $1.5 billion in 2024. There is an obvious market for such tools in schools and universities, assuming that they actually work as claimed. An article in Forbes reckoned that 17% of all academic content was generated by AI, while a UK report found a much lower figure, albeit of proven cases only. However, such figures assume that AI plagiarism technology actually works – does it?
The US Federal Trade Commission ordered a company called Workado to back up a claim that its tool was “98% accurate” in detecting AI writing. Independent testing showed that the true detection rate was 53%, so essentially no better than tossing a coin. This is not the only such tool with issues. An academic study that tested plagiarism tools on papers submitted by students in 2013 and compared them to ones submitted in 2023 found that the tools flagged up essentially the same proportion as being “potentially AI written”. Since ChatGPT only appeared in late 2022, it seems unlikely that the 2013 student papers were written using an AI, at least unless the students had access to a time machine. This lack of reliability amongst these tools is a serious issue, since accusations of plagiarism can seriously affect the career of a student or scientist submitting an academic paper. Moreover, there are further tools available (“AI humanisers” like HIXX Byass and BypassGPT) that claim to be able to edit a document to remove from it any tell-tale AI traces. A human student wishing to disguise an AI-written essay could do so without computer assistance, just by adding personal anecdotes or varying sentence length.
Detecting AI-generated images is also getting harder. Generative AI images in early 2023 tended to have limitations that were easy to spot. The AIs had trouble with human hands, often adding an extra finger. In some absurd cases, they would even add extra limbs to human figures. They also struggled with text. If you asked for an image with a specific piece of text, like a company name on a logo, you would usually get gibberish instead of the requested text. This situation has improved a lot, and by September 2025 AI image generators can produce very realistic images that only an expert could distinguish from a human-taken photo or drawing. One indication can be that a person in an image has flawless skin, too perfect for a human.
The tools still have strange limitations. If you ask an AI tool to generate an image of an analogue clock face it will almost always generate a watch showing the time ten minutes past ten, whatever time you have asked to be displayed. This is because watch adverts favoured their watches being set to this time since it was symmetrical and aesthetically pleasing. Consequently, this time shows up a lot in AI training datasets. To test this, I generated these images on 9th September 2025 from a few AIs with the following prompt.
“an analog watch face showing the time twenty past two”.
If you try different AI image generators you will often encounter the same issue, though occasionally you just get a random time, as the last one from DeepAI. Remember that LLMs are probabilistic creatures, and will occasionally go down less common paths in their predictions.
The issue of detecting AI-generated images has a serious side, such as the use of AI for deepfake videos. It is also an issue for artists and graphic designers, since AIs can be asked to produce paintings in the style of a particular artist. A simple example of this is shown below.
The quality may be debatable at this point, but if you are an artist then you can see that having made up images in your style could be a concern, especially if they are plastered all over the internet.
One country that has decided to do something about the issue of fake AI material is China, which is a serious force in AI. In September 2025, China passed a law requiring that all AI-generated content, whether text, image or video, must carry a label or watermark. Chinese AI companies WeChat, Douyin, Weibo and RedNote have already complied. Elsewhere in the world, such authentication is purely optional. Watermarks of this kind seem to be easily tampered with, but this initiative seems to be at least a good start.
As AI penetrates more and more aspects of society, the inability to distinguish fake and real content is becoming an increasingly important issue. Already, over half of new internet content is probably AI-generated. Since LLMs do badly when trained on AI-generated data such as synthetic data, this is a problem for the AI vendors, and not just humans trying to navigate content. AI content is poisoning the well of training data for other AIs. The scale of this problem is increasing and will only get worse. Identifying AI content seems like a fairly fundamental requirement in many areas, even if the technology to do so is not yet perfect.
China should get credit for trying to do something about this mess.
