“AI use: no”, “Text written by a neural network is not paid for!” – such notes and others like them have started to appear on text exchange platforms in the comments to orders. What’s the deal? Neural network texts are highly praised, said to be almost indistinguishable from human writing, yet almost nobody wants to work with them. What are the advantages of “live” authors and is it possible to differentiate original text from generated ones?
What is ChatGPT?
ChatGPT is a chatbot based on artificial neural network technology, adapted for direct communication with humans. You don’t need programming skills to use it; just ask questions as if you were talking to another person, and the bot will search the internet for answers based on its previous experience in information searching and data collection.
The more queries the bot processes, the better it becomes at gathering information and generating accurate responses, making its communication more “humanlike”. Doesn’t it sound similar to teaching a child as they grow?
What tasks can ChatGPT currently perform instead of a human?
- Searching for information instead of using a search bar. If you need a summary on a certain topic, there’s no need to type a question into the search bars of Yandex or Google; direct it to the bot instead. It will autonomously find information on the internet, analyze it, and present the answer in an accessible human form. The only downside noted by users: the bot does not provide links to the websites from which the information is taken, so there’s no way to verify its accuracy.
- Composing meta tags, product descriptions, email newsletters. These are small texts containing keywords or phrases, but even here artificial intelligence can make mistakes or write outright nonsense. Therefore, do not forget to proofread the results before posting them on your site or sending them to customers.
- ChatGPT is also good at developing software code, which can be used as a base for creating software applications. Sure, it writes only code snippets, not full applications, but for programmers, this is a time-saver; the received code is easier to refine than writing from scratch. In addition, the neural network can be asked to decipher what a certain code will do, supplement it with missing elements, find errors in the program’s operation, and much more.
- Write a story, essay, or article on a given topic using keywords. If it turns out short initially, it can be asked to elaborate or find more information. You might start by composing a plan for the future article, and then elaborate on each point in more detail. The articles turn out unique and sometimes quite original, so many authors have started to profit from selling such works.
Is it fair to pass off text from a neural network as your work, and can you fill your site with such content if you need help but don’t want to pay extra for a copywriter?
How a Person Can Distinguish AI Text
Although neural networks become more sophisticated each year, there are still ways to recognize text written by a neural network through deep and detailed analysis:
- A neural network may write obvious nonsense or unreliable data if not directed to explore a topic more deeply or to confirm specific facts. Therefore, in short texts, it’s easier to find inconsistencies and false facts.
- The information provided may be dry and specific. If we compare a student’s essay on a particular subject and a neural network’s output, the difference will be noticeable: in a human essay, there will be clear logical connections between events, associations with other works or real-life situations, and subjective assessments of “like – dislike”. The neural network simply presents the material without trying to conduct a deep analysis of relationships, causes, and consequences.
- There is no personalization. It’s unlikely that a neural network will write “I thought”, “I had doubts”, “I was told”. But the phrase “according to millions of users” seems more like machine work from the network, although even a person is capable of including such a thesis in their work.
- There are no mistakes. As much as we wish otherwise, there are no perfectly literate people; everyone makes mistakes or typos. But not a machine brain.
- Neural networks cannot yet use slang and abbreviations.
- In the past, Zipf’s Law was used for statistical analysis to distinguish handwritten texts from machine ones. The idea is that we all have “favorite” words which we use in spoken and written language. Such words will be in the top three for frequency of use in the text, and the frequency of word #1 will be 2-3 times higher than the frequency of all subsequent words. However, AI is capable of maintaining this law if such a parameter was set by developers. After all, statistics is an exact science that is entirely understandable to a digital brain. Yet, this type of analysis could be used as one of the factors when identifying machine-generated content.
These factors will come in handy for “eye analysis” if an article is singular or small-scale. For examining a large data set or a whole block of texts on a given subject, better to utilize specialized services that can identify machine-generated text.
Services That Help Identify AI Text
Machine-written text will pass an anti-plagiarism check because it was generated from scratch and uniqueness parameters are inherently laid into the result by default. Therefore, special services have been developed that analyze different text parameters, identifying AI work with a high probability.
-
GLTR was developed by researchers from Harvard and IBM and guesses AI texts based on GPT-2 and GPT-3 with an accuracy of about 66% plus-minus. Unfortunately, it does not currently work with the Russian language.
-
GPT-2 Output Detector – a service from ChatGPT’s creator, OpenAI. The service assesses the likelihood that the work was written by a neural network. Consequently, some handwritten pieces might be incorrectly classified as neural-generated, especially if they contain dry facts or statistical data.
-
GPTZero Classic analyzes sentence structure and length, as well as article predictability. There are paid and free versions, and you could use either. The probability of detecting AI is quite high, but online data suggests the service can be outwitted. It also does not analyze Russian texts very well.
-
ChatGPT itself may analyze text or its excerpts for you. Start by asking if it was written by a human, then specify how the neural network determines authorship. Perhaps the neural network will suggest additional ways to distinguish “human” compositions from machine-generated ones.
Remember that these services cannot guarantee 100% reliability in authorship identification. Some of them were created and “trained” on the GPT-2 model, while most chats operate on GPT-3.5.
Whether to place AI-generated texts on your website or use them for work is a decision for everyone to make individually. If it’s not about verifying a student’s academic knowledge or a specialist’s skills, why not save time and delegate the routine work to a chatbot? And spend your time on something pleasant and human-like, such as spending time with loved ones, a task AI has not yet mastered.
