How to make your content unlike AI and why it matters even if you do not
Why does your article look like it was written by AI?
This question has become the number one trending topic this year. It combines all our 21st century suspicions and insecurities about fake news and misinformation.
Creators and consumers are playing a gambling game. With AI technology in hands, creators have more power than ever before, but they try best to hide the fact they used AI. Meanwhile, consumers have to discern and question who exactly generated this content.
Why you need to identify AI content yourself?
The reader should not have to rely on content detectors or expect Google to protect them from AI content, and AI-assisted writers should not have to worry that big tech is out to get them. There have long been rumours that AI will harm SEO, and that search engine algorithms will protect readers from content not written by humans. This is not the case, and it is not likely to happen. At the end of the day, language is language. SEO does prioritise relevance, engagement and authenticity, so there is some truth to it, but it cannot distinguish between AI and human text. And frankly, not all low-quality content is written by AI. In the right hands, AI-generated content can be excellent.
How I discern AI content:
I am good at reading AI content and I make a living out of spotting a good bluff. I don't believe the common misconception like "AI writing has no emotion". This perceiption is misleading and rules out many fine, objective human writers who create with precision and subtlety. In fact, AI is capable to generate the most moving pieces.
I look for patterns, repetition, or overuse of particular phrases or quotes. I analyse the fluency of the content, the choice of examples, the depth and breadth of topics covered in the article. And for those who are experienced and immersed in the field of generative AI, they are indeed better at spotting AI-generalted content.
How to use critical analysis to expose AI.
You can use this trick for good or for bad. I am not here to judge you. It can help you detect content yourself, or edit and cover up obvious signs. I think at least it's better to improve your writing, if you use this trick to hide AI content.
Ultimately, the quality and readability of the content is what matters most in the future, whether it works with AI or not.
Why not use an AI detector?
Why you don't use a lie detector? Both are known for being unreliable. In fact, AI content detectors are even more unreliable than lie detectors whose false positive rate between 10% and 30%, depending on the study.
In July 2023, OpenAI shut down its AI classifier. It had a failure rate up to 74%.
Ethan Mollick, a professor at the Wharton School of Business and technology expert, advised that “AI detectors have a high false positive rate and should not be used”. However, the detectors are still being used and provide false confidence, or false fear, depending on your attitutes toward AI.
The flaws of AI content detectors.
TurnItIn was used to assess 38 million student essays for AI writing in just three months (April to June 2023), but the company admits their algorithm incorrectly flagged 4% at the sentence level. This means that potentially 1.8 million essays could have been incorrectly flagged as ‘AI content’.
In an educator FAQ last week, OpenAI was blatantly stating:
"Are AI detectors effective? In short, no. While a number of organisations (including OpenAI) have released tools that claim to detect AI-generated content, none have been shown to reliably distinguish between AI-generated and human-authored content.’
Whether or not they are good at capturing AI, the fact remains that detectors are widely used, and if you write with AI, it is wise to avoid being flagged.
I don't use AI, why should I care?
Ironically, we may have to start using AI content detector to protect human-originated work, almost like a writing version of CAPTCHA, but integrated into the writing style and grammar. This may not be a bad thing, but rather encourage better rhetoric and writing skills.
This is why this article is useful for both purely human writers and AI-assisted writers. It's about avoiding clichés and adding value.
Use the AI content detection removal tool.
One way to avoid AI content detectors is to use a generator specifically trained to circumvent them. The leading tool in this space is Undetectable.ai.
One can run content generated by any other AI writing tool through Undetectable.ai. And you're free to use this information!
Undetectable.ai uses a federated multi-model architecture, combining custom and fine-tuned models to identify text that might trigger an alert, remove it, and then automatically reword it for you. Whether this rewriting is more human-like or just bypasses the AI's rada, it is up to you to decide.
It does allow you to see how your text is perceived by the major free and paid AI detectors. While we have established that content detectors are unsuccessful in most cases (potentially only identifying AI content correctly in 20-30% of cases), AI content detection removal tools like Undetectable.ai can help keep it from being suspected.
How AI content detectors work, or, why they don't!
AI detection kits measure two main metrics: perplexity and burstiness. Basically, perplexity refers to the predictability of the next word in a sentence, while burstiness refers to the evenness of sentences --- humans write with sentences of varying lengths, while AI is more consistent. These same metrics are also used to measure the performance level of LLMs.
AI content detectors use generative adversarial methods. The detector basically asks, Would I write this? Or more precisely, would I complete the next word in the sentence like this?
It's a game of adversarial AI trying to predict the next possible word. This is actually how AI writing works, programmatically, word after word, based on statistical results. If it guesses right, a logically similar AI model may have initially generated the text.
This works partly good, but it relies on both AIs using the same vocabulary.
You can get around, or confuse, any detector by combining different models. Many AI text generators utilise multiple LLMs and switch between the most appropriate models depending on the case, e.g. Jasper AI uses OpenAI, Neo X, T5, Bloom, and a proprietary model.
Moreover, the latest AI language models are becoming increasingly fluent and human-like, which exacerbates the need for more powerful means of AI detection, which are trained through an adversarial process. In turn, models like Undetectable.ai are trained to bypass that kind of technique.
Learn how to recognise AI art here.
Computer scientists warn that "eventually, there will be nothing special about texts written by AI that can consistently distinguish them from those written by humans, and detectors can be defeated". But in my experience: there is a factor that should not be overlooked. This is where our human intuition comes into play.
Human intuition: why rely on humans for AI detection?
When we deal with language-based AI, a strong method we could use is rhytoric. Critical reading is the antidote to AI content. While technology is developing at a rapid pace, the intrinsic value of the subtle nuances and imperfections of human communication remains.
In cases where AI models generate descriptive or connected content, they tend to rely on default mechanisms to link ideas or provide references, even if it is repetitive or does not fully fit the context. Overreliance on this technique makes the content seem inauthentic, more like a machine mimicking human speech patterns than actual human conversation.
It's not that AI text no longer resembles human writing; it's that AI mimics the structure and rhythm of human prose so well that it fails to capture the spontaneity and unpredictability of real human communication. While AI can replicate the framework, it sacrifices nuance and unique expression – those irregularities – that make our words come alive, replacing them with conformity and technical correctness.
Some people just happen to be sensitive enough to the complexities and idiosyncrasies of language, and thus detecting the clumsy clichés that algorithms inadvertently produce in their relentless pursuit of patterns, consistency and machine-driven precision.
18 obvious signs of AI content.
1. Interactive commands. This one is ridiculous, but it got past the editors of the scientific journal Nature, so it's worth repeating: look for the phrases ‘regenerate response’ or ‘as an AI language model, I’. It's a clear sign of copy-and-paste straight from the chat window.
2. Textbook-perfect spelling. Because it is a large-scale language model, the AI knows the spelling of almost every word. In addition, the AI does not make typos because, well: there are no fingers. And we all know what human did.
3. Always neutral perspective. AI-generated content is overly neutral. This tendency is due to the fact that AI models are designed to avoid displaying bias or causing offence and to appeal to a broad audience. The content usually adopts a ‘on the one hand, on the other’ approach. This is different from a balanced perspective, which weighs up the pros and cons but may still come to a conclusion.
4. Overuse of the "key". Once you recognise its ubiquity in AI-generated content, the word "key" sticks out like a sore thumb. A telltale sign is its repeated use in formulas such as "When you learn or start something, the key is to ..." While the word is indeed key in some cases, looking out for repetition patterns can help distinguish AI text from human writing.
5. "Dive into, explore, discover." AI has a preference for specific word patterns. These include the use of the verbs "dive into", "explore", and "discover" in CTAs. This is especially true when followed by ‘...and enter the exciting world of xx’. It's a copywriting technique, but I personally have avoided it after noticing how frequently it appears in AI-generated content.
6. "Unlock". Another word to avoid. It's especially common in AI marketing.
7. "ensuring". AI uses this word more than any reasonable human.
8. "A little xx". This is a colloquialism I now only see in AI.
9. Excessive use of demonstrative pronouns. High frequency use of words like ‘the, this, these, that, those’ may be an indicator of AI. Daphne Ippolito of Google Brain says "A very simple clue is the word 'the' appearing too many times". Demonstrative pronouns are among the most common words in the English language. As a result, AI models tend to use them frequently when generating content because they are statistically common in the training data.
10. Sentences beginning with "Note that...". When you come across content that constantly reminds the reader, it could be a sign of AI.
When the author is about to discuss an unpopular concept, they will pause a little bit to draw attention. This is a pattern AI models love. Of course, it is worth noting that human writers use the same technique as well. The difference is the frequency with which they are used. AI may overuse or apply these techniques where humans would not. Remember, the key is to spot the repetition!
11. Metaphor + three-part structure. This pattern is usually [Topic] is more than [basic description or literal explanation]; it is a [deeper or metaphorical explanation] [verb] [broader context or meaning expressed as a three-part structure].
Example: Reading is not just a hobby; it is a bridge connecting our social history, imagination and collective consciousness.
Once you start noticing this technique – a metaphor immediately followed by a three-part structure to reinforce its meaning – you'll start seeing it everywhere in AI-generated content.
12. Cannot construct an inverted sentence. This is a bit harder to explain. An inverted sentence is a way in which the expected order of words in a sentence is reversed. It is used to create a specific emphasis or poetic effect. The most famous example is Yoda quote, "Do or do not. There is no try".
It's very difficult to get ChatGPT to reliably invert sentences, and it shapes the way AI processes language. This highlights the directness of AI in getting from point A to point B. AI-generated text is procedural, word after word, so it's not good at the pre-thinking required to form inverted sentences. AI tends to avoid backflips and verbal gymnastics. So, "It's this sentence that an AI is unlikely to write."