Decoding AI Detection: How to Avoid AI Detection in Academic Writing
One thing authors detest is running their piece of content, written without the help of any artificial intelligence (AI) tool, and receiving a high AI detection score! And these high AI detection scores can be extremely damaging to a writer’s reputation, especially in the case of students, academics, and researchers.
This blog will help you understand how AI detectors actually work, explain their limitations, and highlight how you can avoid AI detection.
How is AI Usage Detected in Academic Writing
AI Text Detectors vs. Plagiarism Checkers
Tips to Avoid AI Detection in Academic Writing
How AI Detectors Work
With the rise of AI writing tools, it should be no surprise that there came a demand for AI detection software and tools. AI detectors are typically used by professors, institutions, educators, publishing houses, and content creators to keep AI-generated content in check1. These tools detect AI writing primarily through two criteria: Perplexity and Burstiness.
Perplexity
Like the word suggests, “perplexity” is a measure of surprise or unexpectedness in the flow of text. AI detectors look for highly predictable text and identify them as AI-generated content, whereas more natural sounding language is considered a sign of human-like writing. In other words, the lower the perplexity, the higher the AI detection score.
Let us understand this with an example. Consider the following sentence: “It was raining heavily last_____.” There are multiple ways to complete the sentence, and the table below explains how each of the possible continuations are evaluated by AI detectors.
| Possible continuations for the sentence | Perplexity | |
| 1 | It was raining heavily last night. | Low: The most likely word to be used to complete the sentence. |
| 2 | It was raining heavily last time we visited the beach. | Low-to-Medium: Not so likely continuation, but it is grammatically correct and makes sense. |
| 3 | It was raining heavily last summer, eerily confirming that climate change is real! | Medium: The sentence uses unexpected phrases and is slightly poetic, indicating human-like writing. |
| 4 | It was raining heavily last when the fox jumped over the moon. | High: The sentence makes no logical sense, using surprising wording that no AI tool would use. |
As you can see in the first sentence, “night” is the most logical option to complete the sentence, and AI text detectors give this a low perplexity score. The second and third sentences are likely to receive higher perplexity scores, indicating human-like writing because of the less predictable wording and language. The fourth sentence will receive the highest perplexity score considering that it makes the least logical sense to the AI detector!
Does this mean that you should intentionally make mistakes in your writing? Definitely not! More on this later.
Burstiness
Now let us understand the second factor used for measuring AI-generated text: burstiness. Burstiness is a measure of the rhythm in the flow of text. This is similar to perplexity, except that it applies at a sentence level check.
Humans tend to write using a mix of short and long sentences, involving both simple and complex phrases. AI writing tools, on the other hand, tend to repeat certain phrases and maintain the sentence length quite uniform throughout the text. This means that a monotonous tone in the text is likely to receive a lower score for burstiness, indicating AI-generated writing.
Overall, AI detectors calculate the balance between perplexity and burstiness to determine whether the writing is human- or AI-generated. But these tools are constantly evolving, and more layers of evaluation metrics are being added to enhance their accuracy of distinguishing between AI-generated content and human writing. Yet, at present, know that AI detectors are not foolproof!
Limitations of AI Detectors
This brings us to the limitations of AI detection tools. We have all heard that AI detectors are biased2 and may not be entirely reliable because of how the same piece of text are evaluated differently by various tools. A study3 published in 2023 also observed that AI detectors were biased against non-native English authors. They reported that AI detectors accurately classified essays written by US students, whereas 19.8% of the human-written TOEFL essays were incorrectly flagged as AI-generated. One detector ended up flagging over 97% of TOEFL essays as AI-generated!
Well, since it’s 2026, we can be hopeful that AI detectors have now worked on mitigating this bias. To test their reliability, I ran a simple experiment myself. I asked Google “How is AI use detected in academic writing?” and scanned its AI-generated response (approximately 500 words) through four different AI text detectors. Two of the tools gave me an AI score of 50%, one tool identified that the entire text was AI-generated, and the fourth tool showed an AI score of 94%. Furthermore, I pasted the first half of this blog (about 580 words) and ran it through the same tools. The first two tools showed an AI score of 13%, one tool correctly identified that the text was entirely human written, and the fourth tool gave an AI score of 34%.
These inconsistencies continue to make people question the reliability of AI text detectors. And relying solely on these still-evolving tools to detect AI-generated content in academic writing would be extremely unfair. So, how is AI use typically detected in academic writing?
How is AI Usage Detected in Academic Writing?
Fortunately, most professors, educators, institutions, and journal publishers are aware of the shortcomings of AI detection tools. Therefore, AI use in academic context is often evaluated through a mix of detection tools and manual review. Also, top publishers like Springer Nature utilize their in-house AI detection tools4, rather than relying on third-party algorithms and software, to ensure fairness and accuracy. Tools like Geppetto (checks for AI usage in manuscripts) and SnappShot (analyzes image integrity) are designed to evaluate research manuscripts in the context of academic writing, reducing the chances of false positives often observed in commercially available AI detectors.
When it comes to manual review, there are several factors that come into play. While AI detection tools analyze patterns, your professors look for inconsistencies. For instance, if a student’s past work has shown a certain style, voice, and tone in writing, a sudden change in this habit coupled with lengthy sections of text could be a red flag. Similarly, AI-generated content is often vague and does not dive deep into the concept. So, if your assignment includes superficial information without offering any interpretations or deep insights, your professor might consider it as AI written.
Another way for professors to check whether AI was used to write your thesis or assignment is by following the age-old method of asking questions! If you cannot orally defend your analysis, it makes your professor question the origins of your written content. So, be prepared with your notes and clarify any queries that your professor may have about the research topic to confirm that you authored the thesis yourself!
Journals also evaluate the authenticity of references and citations in a research manuscript. Because AI tools tend to fabricate false references5, journals and publishers use reference verification tools to identify fake references or unreliable citations. Furthermore, experienced peer reviewers can recognize shallow analysis or generic arguments, and papers that lack originality and critical thinking are often not accepted for publication.
AI Text Detectors vs. Plagiarism Checkers
Many authors mistake AI detectors for plagiarism checkers or are unable to differentiate between the two. While they appear similar, each serves a different purpose.
- Plagiarism checkers are used to recognize “copied text” from other existing sources. Essentially, plagiarism checking tools compare your text with text from databases and other sources from the web to check for similarities. Tools like iThenticate are often used to evaluate research papers to generate a similarity score.
- AI text detectors are used to identify the author of the text. There is no comparison made with other sources to check for similarities. Instead, AI detectors simply determine whether the writing is human- or AI-generated. These tools check for linguistic patterns and provide an AI score based on their analysis as per the machine learning model they are trained on.
Tips to Avoid AI Detection in Academic Writing
Now that you know how AI detectors work (and why they fail to classify text accurately sometimes), it is good to implement the following best practices when writing.
- Firstly, avoid using AI writing tools to prepare an entire research paper or thesis. You may use tools like Paperpal to brainstorm research topics, to get an outline of the paper, or to help paraphrase sentences. But the actual writing of the paper with your research details must come from you.
- Make yourself aware of what AI detection tools or software are used by your institutions or professors. This is not to “trick” the tools, but to know in advance what you can expect. Is your professor highly reliant on AI detectors? Are they open to listening to your explanations if the tool incorrectly classifies your work as AI-generated? Are you prepared to defend your work? If you are a student, keep all your notes and transcripts so your professors can verify the sources of your final written document.
- Avoid using AI editing tools. Many non-native authors tend to struggle with writing because they are not used to speaking in English as their first language. However, overusing AI tools to correct your grammar and syntax could be detrimental.
As much as possible, have your research papers and thesis reviewed by human editors who are experts in your field. Their insights will not only help you improve your paper to make it submission-ready but also enrich your knowledge. For instance, context-based subject-specific comments offered by English editing experts can help you correct your mistakes in future research papers. - Write like you are speaking to your readers. The text should feel natural and flow logically from one sentence to the next. Maintain a rhythm of short and long sentences and include simple as well as complex phrases in your writing.
- Include personal experiences and use first-person pronouns (I, we, our) when writing, if the journal permits this. You may have to state facts in an objective manner, but the interpretation of those facts, such as in the Discussion section of your thesis or research paper, can follow a personal tone.
- Do not intentionally make spelling mistakes or grammatical errors to avoid AI detection scores! This is not becoming of an academic writer and a manual review will make it clear that you simply wanted to “trick” a tool into giving you a low AI detection score.
- Keep in mind that AI detectors check for predictable language. Avoid a monotonous writing style and provide in-depth analysis of any topic that you are exploring with relevant, verified citations. If you do use tools like ChatGPT to get references, verify their authenticity before putting them in your paper.
Remember, you are responsible for all the content that goes into your research manuscript! Even if you do use AI, make sure that your honestly disclose it and clarify for what purples AI tools were used during manuscript preparation. Let your writing by AI-assisted and not AI-led!
References
1. How do teachers check or AI? https://www.compilatio.net/en/blog/how-do-teachers-check-ai
2. AI Detectors are NOT Always Fair! AI Bias Explained https://www.youtube.com/watch?v=zbx7RcTbn38
3. GPT detectors are biased against non-native English writers https://www.cell.com/patterns/fulltext/S2666-3899(23)00130-7
4. Springer Nature unveils two new AI tools to protect research integrity https://group.springernature.com/gp/group/media/press-releases/new-research-integrity-tools-using-ai/27200740
5. Fabrication and errors in the bibliographic citations generated by ChatGPT https://www.nature.com/articles/s41598-023-41032-5
Looking to refine your research paper for successful publication? Check out Editage’s Premium Editing Services and get submission-ready manuscripts.
