How to Avoid AI Detection in Your Research Paper the Ethical Way
A lot of the advice out there for avoiding AI detection will tell you to do things like vary the length of your sentences, avoid the rule of three, and avoid words that LLMs tend to use (e.g., don’t “delve into” anything).
This advice isn’t worthless, and it may help your paper bypass AI-detection tools. However, it won’t fool humans who are looking for more than the absence of AI “tells” because reviewers are primarily interested in whether a paper shows signs of critical thinking.
So what does this mean for you if you’ve written your paper yourself but are still being accused of using AI?
It could mean that the problem isn’t that your paper contains too many AI signals but rather that it doesn’t contain enough human ones.
And if this is the case for you, then the standard workarounds like restructuring a paragraph here and replacing a “delve” there aren’t going to solve the problem. And that’s because those workarounds merely change how the writing looks without adding more actual human thought, which is what reviewers are looking for.
In other words, if you want your writing to sound like you wrote it instead of a machine, you need to put your own ideas, knowledge, and insights into your paper.
This article explains some ways you can do this. They’re not quick or easy, but they work.
Favour specific phrases over vague ones.
Because LLMs lack critical thinking abilities, they often fail to identify which details about a subject or specific study are worth noting.
Instead of guessing (and potentially guessing wrong), they often play it safe and provide a broad statement that sounds true but doesn’t really say anything.
To give you an example of what I mean, here’s an AI-generated paragraph that you might see in an introduction section of a research paper:
Although obesity in England is influenced by multiple interacting factors, the food environment plays an important role in shaping dietary behaviour. Policies that improve the nutritional quality of food purchases and encourage healthier consumption patterns may therefore contribute to reducing the population burden of obesity. However, interventions that rely solely on individuals making informed dietary choices may be insufficient to address the wider environmental factors influencing food consumption. Consequently, there is growing interest in policy approaches that modify the information available to consumers and incentivize improvements in the nutritional composition of commercially available foods.
That paragraph contains a lot of vague phrases, including classic AI favourites such as “is influenced by” and “plays an important role in shaping.”
As a result, while the paragraph contains some meaning, it could take multiple readings to extract that meaning.
And that’s because the LLM that produced this text wasn’t thinking, “Will readers be able to easily understand what I’m trying to say?” which is something humans naturally do when speaking or writing.
So, when a reviewer sees a lot of phrases like “plays an important role in shaping,” they’ll suspect AI was used.
If your paper contains many such phrases and you want to make it sound more human, revise the text while asking yourself one question about each sentence:
Would a human reader with basic knowledge of my topic be able to make sense of this relatively easily?
Here’s what the AI-generated paragraph above looks like after I applied this rule:
The food environment (e.g., the convenience and relatively low cost of unhealthy fast food) has particularly troublesome effects on what people choose to consume and is partly responsible for the increasing obesity rate in England. While policies have aimed to reduce obesity by encouraging people to purchase healthier foods, interventions that assume people are in full control of their dietary choices may be ineffective since they fail to address the environmental factors that influence what and how much people decide to eat. Two such environmental factors that have recently gained interest among policymakers are the nutritional content of commercially available foods and the nutritional information provided on these foods’ packaging.
Among the changes I made to this paragraph, the ones most closely related to specificity vs. vagueness are:
The broad term “food environment” includes an example to improve clarity.
The direction of the food environment’s effect is clarified (it has “troublesome effects” rather than “playing an important role”).
The generic verb phrase “encourage healthier consumption patterns” is replaced with more specific, concrete language (“purchase healthier foods”).
The generic term “informed dietary choices” is replaced with a more specific alternative (“what and how much people decide to eat”).
Instead of only saying that interest in the matter has increased, the paragraph now says who in particular is interested (policymakers).
The two specific environmental factors of interest are named (“the nutritional composition of commercially available foods” and “the nutritional information provided on these foods’ packaging”).
Avoid inflated claims.
LLMs sometimes like to use grandiose language to make things sound more impressive or interesting than they are.
To be fair, this kind of language is also commonly used by humans, and there’s nothing inherently "AI" about referring to an event as representing “a critical turning point” or “a fundamental shift.”
However, a reviewer’s AI detectors will go off if a paper overuses these kinds of phrases.
Again, this is because inflated claims show an absence of critical thinking. They merely say that something is meaningful without specifying what is meaningful about it.
Here’s an example of an AI-generated sentence that uses an inflated claim:
Identifying the factors that influence whether firms adopt sustainability labels could bring about a critical transformation in the implementation of policies that promote sustainability labelling on the packaging of food products.
This sentence highlights one of the main issues with inflated claims: they conceal information.
In this case, the phrase “bring about a critical transformation in the implementation of policies” doesn’t let the reader know three things they probably want to know: (1) the nature of the transformation, (2) who would instigate the transformation, and (3) what people or situations the transformation could affect.
To fix the sentence, we need to plug these details in:
Identifying the factors that influence whether firms adopt sustainability labels could help policymakers introduce changes that would promote these labels’ usage by making it less costly for producers to apply them to their food products’ packaging.
Avoid introductory phrases that add words without adding context.
Academic writing commonly includes sentences that provide some context before the main claim is made. For example:
In today’s increasingly globalized world, consumer goods are often imported from countries other than where the primary product originated.
Here, the author used the transitional phrase “In today’s increasingly globalized world” because they didn’t just want to note a recent change in imports. They also wanted to mention the reason for the change (the world has become more globalized).
Because these phrases are common in academic writing, LLMs have been trained to use them.
The trouble is that LLMs don’t always know when an introductory phrase is actually helpful and when it’s just adding words for no reason.
So, one way to make your writing more human is to look for these introductory phrases in your own paper and ask, “Is this providing necessary context for the reader?”
If yes, then you can leave it in, perhaps rewording it if you’re particularly concerned about sounding like AI. For example, the above sentence could be written as:
Owing to globalization, consumer goods are often imported from countries other than where the primary product originated.
If no, then you can either reword or delete the introductory phrase, depending on what you feel is more appropriate.
Here’s an example where the introductory phrase is misused:
In today's increasingly globalized world, language barriers remain a significant obstacle to communication across diverse linguistic communities.
Since globalization isn’t causing, worsening, or in any other way affecting language barriers, the introductory phrase is inappropriate and should be changed:
Even as the world becomes more globalized, language barriers remain a significant obstacle to communication across diverse linguistic communities.
Add your own personal viewpoints.
One thing AI can’t do is reach into your mind, extract your thoughts, and put them in your research paper. At least not yet.
And therefore, one of the best ways to avoid AI detection is to put your personal viewpoints into it.
In general, the discussion is the most suitable section for this since it’s where you’re expected to give your interpretation of your results and the insights into your topic you’ve gleaned from conducting your study.
If an author doesn’t make any interpretations or present any new thoughts in light of their results, their paper is more likely to be flagged as AI-generated.
Below is the kind of passage that could make reviewers think a paper was AI-generated:
One possible explanation for the difference between the two policies is that nutrient warning labels explicitly identify products exceeding specified thresholds for nutrients of concern. This approach may allow consumers to recognize less healthy products quickly without requiring them to interpret several separate nutrient indicators. Traffic light labels provide more detailed information about individual nutrients, but interpreting the overall nutritional quality of a product may require consumers to consider several indicators simultaneously. The more direct presentation of nutrient warnings may therefore influence purchasing decisions more strongly.
Aside from the typical language-level tells, this passage comes across as AI-generated because the author has added nothing new to the discussion.
If this were my paper, I would be curious why less detailed information led to healthier purchase decisions, and because my background is in psychology, I would want to consider what psychological mechanisms could have influenced the results.
Another author with a different background and different knowledge might interpret the finding through a different lens.
Meanwhile, AI has no lens, so it often ignores the parts of the discussion section where the author is supposed to give their own perspective on a finding.
With that in mind, one of the best ways to avoid having your paper flagged as AI-generated is to add your own thoughts about your findings, like this:
The difference between the two policies could be explained by the fact that, unlike traffic light labels—which provide detailed information about individual nutrients—nutrient warning labels clearly show consumers when a product exceeds a specified threshold for sugar, saturated fat, and other nutrients of concern. This allows the consumer to quickly and easily identify unhealthy foods. In this way, nutrient warning labels’ explicit presentation of nutrient warnings may influence purchasing decisions. This could be because of people’s tendency to weigh negative information more heavily than positive information when making purchase decisions (e.g., Hall et al., 2023). However, this supposition conflicts with previous findings indicating that combining positive and negative nutritional information—as is done with traffic light labels—promotes healthy product selections more strongly than negative information alone—as is done with nutrient warning labels (Grummon et al., 2024). Perhaps traffic light labels were less effective than nutrient warning labels in the current study because green traffic light labels indicate low levels of unhealthy nutrients rather than high levels of healthy ones. Further research is required to resolve this inconsistency.
LLMs are currently unable to write paragraphs like this because, as with the other humanized examples throughout this article, the revised paragraph requires critical thinking.
It requires a human to look at the current study’s findings and perform some logical reasoning to propose why the findings occurred.
The downside is that coming up with and evaluating possible reasons for a finding takes time and effort.
But that’s precisely why this is one of the best ways to prevent your paper from being suspected of being written by an LLM. Because it shows that you’ve carefully thought about your results.
There are many other ways to show that you wrote your paper besides those presented in this article. If you need help identifying where your own paper may need to be revised to reduce content that could be flagged as AI-generated, I do this as part of my Full Language Editing Service.
Click here to learn more.
