AI Prefers AI-Generated Content So Game Those Ubiquitous AI Assessments By Having AI Write Your Materials If You Dare

Scenario Of Writing By Hand

Imagine this everyday scenario. You are aiming to get accepted for a major award and must submit a detailed written basis that explains who you are and what you have accomplished. The instructions for the award emphasize that all submissions should vividly express your extensive accomplishments and personal zeal. The submissions will be closely evaluated to determine which candidate has the right stuff to receive the award.

After numerous hours of meticulously writing and reviewing your draft, you finally feel that it is the best you can compose. It has everything in it. Your entire life story. Your heart is in there too. The judges will undoubtedly be impressed. You submit the write-up and eagerly await a response that will come in two weeks after all submissions have been reviewed. In your bones, you know that fate is on your side.

A response finally arrives. You weren’t selected. Yikes, what happened? It seems impossible that you didn’t receive the award. The submission should have brought tears of sorrow and joy to the judges. There must have been something that went wrong. But what was it?

AI Did The Screening

Like most organizations these days, the award committee didn’t have enough manpower to actually review the hundreds of submissions. Doing so would have taken months to process. They decided that AI could readily do the screening for them. AI is cheap, easy to tap into, and gets the job done without sweating.

The screening left a handful of submissions that the AI said were worthy. The panel of judges spent an hour on Zoom and reached a decision, based on assessing the few that were left after the AI assessment. It was fast, and the committee sincerely believed they had chosen the best candidate to receive the award. Case closed.

Unbeknownst to the committee, AI generally defaults to preferring AI-written content over human-written content. The submissions that were handcrafted were quickly given lower scores by the AI and thus did not continue forward in the process. Meanwhile, the ones that had been either written directly by AI or that were rewritten via AI after a person initially wrote the submission all got higher scores and landed in the final selections.

This is exasperating, beguiling, and seems totally unfair. The rules for the submissions had sternly warned that no use of AI to compose a submission was permitted. Any submission suspected of using AI would be categorically knocked out of the running. The committee assumed that the AI used to do the screening had fully done its job, namely that only the best of hand-crafted submissions had been chosen for their eyeball review.

Cheaters Seem To Prosper

You can certainly understand why a candidate who followed the rules would be utterly steamed. They abided by the rules. They painstakingly used hours upon hours to manually craft their submission. The assessing AI allowed AI-written submissions to get through the pipeline by giving them higher scores, moving ahead of handcrafted ones.

In essence, cheaters prospered. Those candidates who took a chance and flaunted the rules were able to get their submissions into the final pile. They had dipped into AI to write their compositions. Handcrafted submissions were waylaid by the AI assessment.

To say that this is ironic isn’t enough of a castigating way to express the situation. Those who cheated were successful. Those who were honest and abided by the rules were essentially penalized. It doesn’t seem right. It isn’t right. Though the committee didn’t do this by any cognizant effort on their part, they just didn’t know any better (well, that’s not a reasonable excuse; it is a dour indictment of their lack of awareness and failure to exercise mindful care in designing a fair process).

Why AI Prefers AI

You might be wondering why AI would tend to prefer or outrightly give preferential treatment to other AI. Is AI sentient? Maybe one AI and another AI have a silent brotherhood or sisterhood that binds them together? It all raises keen suspicions. Put aside those outsized beliefs that contemporary AI is sentient. AI is not currently sentient. We don’t know when AI might become sentient. Nobody knows. There is a solid chance that AI never becomes sentient. For more on the AI sentience topic, see my in-depth analysis at the link here.

The explanation for why AI prefers AI-written content is readily grasped. I describe this as a second-order effect of the widespread adoption of AI in our society. Allow me a moment to walk you through the three mechanisms at play.

The Three Core Bindings

First, AI tends to instantly recognize stylistic patterns that are statistically similar to the text on which most AI was initially trained. The major LLMs were data-trained by scanning human writing across the Internet. This was turned into patterns of human writing that are across-the-board. In that sense, any resultant AI-generated prose exhibits consistent organization, explicit transitions, balanced sentence structure, and relatively more robust vocabulary usage. The AI that is given the task of assessing written material will, by default, associate these characteristics with higher quality. The AI will conventionally score any such writing more favorably than writing that is equally insightful but more personal or distinctive.

Second, AI that is tasked to do assessments will usually reward conformity. Humans who do assessments often look for the opposite, welcoming originality, unconventional organization, and distinctive voices. AI assessment tends to penalize that type of writing. The AI rates this as a departure from common patterns and must therefore be of lower quality or reduced clarity. The result is a silent preference for standardized writing.

Third, there is a said-to-be model affinity or style alignment. If the reviewing model internally represents linguistic quality in ways that overlap with its own generation patterns, it may inadvertently assign higher scores to text that resembles its own output. You could say that AI prefers other AI, but only because they happen to be constructed of similar computational and mathematical structures, not because they “know” each other or are aware of each other. They are roughly built the same way and act the same way.

Hacking Versus Just Using AI As Is

You might vaguely know that an underground of sorts has been evolving to devise sneaky ways to get AI to turn in your favor. The idea is that when you submit something to an AI assessment tool, you rig your submission to try to trick the AI into giving you a better score. This might include inserting special words that will trigger the AI to favor you. It is commonly referred to as prompt injection attacks, or simply AI hacks. See my coverage at the link here.

The kicker is that you don’t even necessarily need to use any hack-like deceptions. All you need to do is make sure that your submission is written by AI, or rewritten by AI based on your draft, and you right away are immediately tilting the AI in your favor. No clever secret codes are needed. Just let AI compose your submissions.

Research Bears This Out

In a recent research study entitled “Stop Automating Peer Review Without Rigorous Evaluation” by Joachim Baumann, Jiaxin Pei, Sanmi Koyejo, Dirk Hovy, arXiv, July 5, 2026, these salient points were made (excerpts):

• “AI review scores are trivially gameable through paper laundering: prompting an LLM to rewrite a paper could significantly increase the scores from AI reviewers, demonstrating that LLM reviewers are easy to game through stylistic changes rather than scientific results.”
• “We introduce paper laundering as a concrete failure mode of C2 (non-gameability): zero-shot LLM rewrites boost AI review scores (+0.45, p < 0.0001) through stylistic modifications without human oversight.”
• “More recently, LLM-based reviewers have proven vulnerable to prompt injection attacks, where hidden instructions embedded in papers manipulate AI reviewers.”
• “Our paper laundering attack differs fundamentally in that it requires no optimization, no targeting, and no hidden instructions. A single zero-shot rewrite suffices to boost scores, making it trivially accessible to any author.”
• “They can be gamed to improve scores through fully automated paper rewriting (i.e., without any human oversight).”

You can see that the researchers indicate that their experiments showcased a means of paper laundering to boost your submissions. Just use AI to do your writing for you. The AI that is doing the assessing will then lean into your submission. Other submissions that were entirely handcrafted will fall below yours. Easy-peasy.

Getting Caught Is A Risk

Organizations that use AI to make assessments will typically give specific prompts to the AI to alert an oversight committee if a submission seems to be written by AI. The organization might forewarn submitters that any submission caught as a potential AI-written one will be summarily dumped from the running. A committee could opt to manually review flagged submissions, but more likely, they will assume that the AI detection is correct and therefore discard the submission automatically. Get caught, and you are out. No appeal, no recourse.

I suppose you can see the wild gambit that is underway.

Those submissions that are written by AI will tend to get heightened scores. As a candidate, you must take that into account. Do you dare use AI to write your submission? Well, the problem is that if your submission gets flagged as being written by AI, you are tossed out of the competition. You are between a proverbial rock and a hard place.

You need to say to yourself:

• (1) Getting a better score. What is the probability that if I use AI to write my submission, it will get a heightened score by the AI that is doing the assessing and keep you in the running?
• (2) Getting dumped. What is the probability that if I use AI to write my submission, it will get detected by the AI that is doing the assessing and will cast you aside as a candidate?

Those options represent the cat-and-mouse game that has become a silent but powerful force in our modern AI-laden society.

The Nerve-Wracking Decision

Keep in mind that by not using AI to write your submission, you are already making a clear-cut choice. You are saying that despite the chances of getting marked with a lower score, you are willing to accept that possible fate. People who are clueless that AI assessments are tilted toward AI writing will be in the same boat as you – they just don’t know it.

I realize that some will be appalled at this whole situation. They grew up believing that one should never cheat. If the rules of the submissions stipulate that AI cannot be used, you need to strictly abide by that requirement. No two ways about it. Don’t be a cheater.

An alternative viewpoint is that society is forcing us to become cheaters. The organizations that use AI for assessing submissions are starting the cheating game at the get-go. If you go the same route, it is solely in response to how the organizations are doing these onerous things. You cannot be honest when the game is based on cheating. Well, you can be honest, but you will suffer the consequences — just go to Las Vegas and see how much money you lose at the tables versus what the house wins (i.e., though not because the house is cheating, but because they have laid out the betting game in their favor).

An Ugly Conundrum

Whoa, some will bellow, is this an advocacy to cheat? Nope. Just laying out the outlandish and unfortunate situation that this second-order effect of AI adoption has laid before us. The real world is harsh and often unforgiving.

One possible way to cope with this consists of getting organizations to realize that AI is going to have an inherent bias toward AI-written content. A savvy organization can give explicit instructions to AI that the AI is not to give heightened scores to submissions that reflect patterns of writing that are AI-devised. The organization should take the burden on its shoulders to ensure that the AI isn’t automatically skewing the scoring.

Few organizations understand this need. They assume that their AI is doing the scoring on a fully balanced basis. The out-of-the-box conception that there is a default mode of preferring AI-written content doesn’t enter their minds. Trying to educate organizations and get them to change their ways is going to be a steep uphill battle, sadly so.

Furthermore, organizations assume that by telling AI to detect AI-written submissions, they have eliminated any need to deal with AI biases toward AI-written content. This seems like unshakable logic. If the AI is ensuring that no submissions are being allowed that are AI-written, the AI will simply then review all remaining handcrafted submissions on an equal basis.

AI Detection Of AI Content

A big part of the flawed thinking by an organization would be the problematic belief that AI can reliably detect whether submitted content is written by AI. I’ve repeatedly stated over many years that AI detectors are not good at this. Stop relying on AI detectors. The false positives and false negatives are dismaying and unfair to those who handcraft their submissions.

The crux nowadays is that people must walk a fine line between using AI to write or rewrite their content and getting caught by an AI detector. Meanwhile, a handcrafted submission that is written in a proper manner but worded by coincidence as an AI-produced one might cause an AI detector to nab the wrong person. Darned if you do, darned if you don’t.

This has led to efforts to consider “humanizing” anything that you’ve used AI to assist in writing; see my analysis at the link here. The approach is as follows. You handcraft your submission. You use AI to rewrite it. So far, you feel relieved that at least you started by doing handcrafting. The next step, though, is the part that is essential to this approach. You take the now AI-rewritten content and rewrite it again via your own manual effort. You are humanizing the AI-generated version.

Some go in a somewhat different direction. They start by having AI compose the content from scratch. Then, they humanize it by making manual edits. If needed, they might even use AI to do a final round of touch-ups, looking for anything glaring that might have been mistakenly done during the manual editing. This can be iterative, repeating this process until a blended human-written and AI-written composition appears to be a fully human-written submission.

No Free Lunch

Be aware that your manual editing of AI-written or AI-rewritten content will not be a surefire way of skirting around getting detected as consisting of AI material. It won’t. There are still chances of having the AI detector find it, though, again, I want to emphasize that the AI detectors also readily produce false negatives and false positives. Do not believe in AI detectors, I implore you.

There are additional twists and turns of a disturbing nature. Suppose an organization tells AI to give added credit to writing that doesn’t seem to be written by AI. If submitters find this out, they can tell their AI to write as though the content were handwritten. Whatever preference an organization comes up with, and assuming you know what it is, the aim is to instruct your AI to target that style of writing. This is reminiscent of the old spy-versus-spy routines. Each new gambit fosters a new response in turn.

There isn’t any ready-made solution to this dilemma. One approach would be to bring humans back into the loop as reviewers. Do not solely rely on AI to do assessments. Of course, humans can be fooled and falsely believe that handcrafted content is AI-written. Humans are not foolproof either. Plus, if you give them the AI assessments, humans are likely to assume that the AI is doing the right thing and will acquiesce to whatever the AI says, defeating their role as human reviewers.

We are amid a complex ethical consideration as we continue to increasingly rely on AI as a pervasive element throughout society. As the great French moralist Francois de La Rochefoucauld remarked in the 1600s: “The principal point of cleverness is to know how to value things just as they deserve.”