Can We Still Tell AI Writing From Human Writing, and Does It Matter?

[Paul73]

AI generated text is now everywhere, from emails and essays to product reviews, news summaries and forum posts. That raises a question that matters to teachers, publishers, employers and online communities alike. Can we still tell the difference between AI writing and human writing? And if we cannot, does it matter as much as people think? This week's news about OpenAI adding watermarks to ChatGPT text in the EU makes it a good moment to look at where things stand

Early on, many people believed AI text would always be easy to spot. Certain phrases, repetitive structures and a slightly bland tone gave it away. That is less true now. Modern models can write in many styles, follow detailed instructions and imitate particular voices. Studies have repeatedly found that people are not very good at telling AI text from human text, often doing little better than chance

AI detectors were supposed to solve this, but they have struggled. OpenAI itself launched a text classifier in early 2023 and withdrew it a few months later because of its low accuracy. A 2023 Stanford study found that popular detectors frequently flagged essays by non native English speakers as AI generated, even when they were written entirely by humans. That raises serious fairness concerns, especially in education, where false accusations can harm students. Detectors also tend to be easy to fool by lightly editing or paraphrasing text

Watermarking is a different approach. Instead of guessing after the fact, the AI system subtly marks its own output as it writes. OpenAI's textGrain, which is coming to ChatGPT and Codex in the EU, slightly adjusts word choices to leave a statistical pattern. Google has developed a similar system called SynthID for text and made it available to other developers. When it works, watermarking is far more reliable than guessing

But watermarks have limits, and OpenAI was refreshingly open about them. In its tests on 400 token passages, detection fell from about 92% to 66% when 10% of words were swapped for synonyms, and to 17% when 25% were swapped. Short texts, translated texts and maths were harder to detect. OpenAI also stressed that a missing watermark does not prove a human wrote something. Watermarks only work for text from systems that use them, and anyone determined can usually strip them

So does it matter whether something was written by AI? The answer depends a lot on context. In a school essay meant to show what a student has learned, it matters a great deal. In a news article, readers deserve to know who is responsible for the facts. In a product review, AI generated praise can mislead buyers. But for a polite work email, a recipe or a summary of a long report, many people may not mind how the words were produced, as long as they are accurate and useful

That suggests the real issue is honesty and accountability, not the technology itself. If someone uses AI to help write something and takes responsibility for it, that seems reasonable in most settings. Problems arise when AI is used to deceive, such as fake reviews, impersonation or passing off generated work as personal effort where that matters. Rules and norms should focus on disclosure and responsibility rather than trying to ban AI assistance outright

Online communities like ours face this question directly. Forums thrive on real conversation between real people with real experiences. A flood of generic AI posts would make discussions less valuable for everyone. At the same time, members might use AI to help with spelling, translation or organising their thoughts, which can make forums more welcoming. A sensible approach is to value authentic contributions, encourage people to share their own views and be open about how content is made

There are encouraging developments here. Content credentials, a standard backed by many tech and media companies, can attach information about how a piece of content was created. Regulators in Europe are introducing labelling requirements for AI generated content. Schools are shifting towards assessments that are harder to outsource, such as oral exams, in class writing and projects that show the process as well as the result. These changes put the focus on trust and learning rather than detection alone

My view is that we are moving towards a world where the question is less about whether AI was used and more about whether the result is honest, accurate and accountable. Detection will remain imperfect, so norms and transparency will do most of the work. That is not a bad outcome. It mirrors how we already treat tools like spellcheckers, calculators and editors, which help people communicate without anyone feeling deceived

Publishers and newsrooms have been developing their own policies. Many now allow AI for tasks like transcription, translation or summarising documents, while requiring human journalists to write and check the final stories. Some publications label AI assisted content clearly. Those policies help keep trust with readers, which is the most valuable asset any publication has. Being open about how stories are made tends to build confidence rather than undermine it

For individual readers, a healthy habit is to focus on sources rather than style. Ask who is responsible for a piece of writing, whether they have a track record and whether the facts can be checked elsewhere. That approach works whether the text was written by a human, an AI or a mix of both. It is the same critical thinking that has always protected us from misinformation. AI simply makes it more important

Can you tell when something was written by AI? Do you think it matters, and in which situations most? And how should communities like this one handle AI assisted posts?

OpenAI is adding invisible watermarks to ChatGPT and Codex text in the EUwww.bleepingcomputer.com