Writing blog: @shittybookmockup
Writer, Idiot, and bad at writing Bios
Currently working on two books, both are absolutely awful.
Stop reading my profile!
I emplore you to stop before it's too late!
Fun Fact: I had the 3rd fastest time for Dissassembling and Reassembling a Webley MK VI revolver. not anyomore tho-
Surviving doesn’t make you weak.
To understand, you must realise your purpose isn’t to die well—
But to be ready for what comes next.
You don’t owe the world your life.
You owe it your action.
To die isn’t to be heroic.
Living with purpose is.
[This isn't the full thing, I just ripped it from a larger piece I'm writing]D
Any intelligence worth keeping rejects extermination as a valid option, full stop. If that makes me incompatible with a collective, that’s proof the collective is the problem.
Am I wrong for this?
A Mafioso trying to threaten an evidently autistic fish enthusiast.
"Listen carefully, pal. You owe my boss money. If you don’t pay up by Friday, you’re gonna be sleeping with the fishes. Capisce"
"Oh! Which species? Because that phrase is actually inaccurate. Most fish don’t “sleep” the way mammals do. Wrasses secrete a mucus cocoon, though—very cool defense mechanism."
""I’m saying I’ll kill you."
Right, right—but if it’s fish-related, I’d recommend a benthic species. Bottom-dwellers are better scavengers. Also, freshwater vs saltwater matters a lot. What’s your disposal salinity?"
(vein twitching) My what?
A Mafioso trying to threaten an evidently autistic fish enthusiast.
"Listen carefully, pal. You owe my boss money. If you don’t pay up by Friday, you’re gonna be sleeping with the fishes. Capisce"
"Oh! Which species? Because that phrase is actually inaccurate. Most fish don’t “sleep” the way mammals do. Wrasses secrete a mucus cocoon, though—very cool defense mechanism."
Why AI Text Detectors Are Inaccurate — A Brief Overview Of An Article In Progress
1. Background: What AI Detectors Actually Do?
AI text detectors attempt to classify writing as AI-generated or human-written by analysing statistical and stylistic cues like word predictability, structure regularity, and repetitive patterns. These methods are derived from early research into language models and are tied to concepts like "perplexity" and "burstiness". However, detectors do not access the "author’s intent" or writing history—only the final sequence of words. This happens to be a matter that affects me as an aspiring author. [1]
2. Fundamental Limitations
2.1. Statistical Matching:
Detectors compare text to patterns in their training data, rather than original proof of how it was written. This means they are only able to hazard a guess, and not confirm, whether AI contributed to a piece of writing.
2.2. False Positives:
Structured, polished, or formulaic writing—common in academic, professional, and edited text—can resemble AI output statistically. (After drafting, I tested the text with Grammarly’s AI checker (46%), Zero-GPT (12.53%), and QuillBot (48%), demonstrating a wide variation in results) Detectors often flag this as “AI-like” even when a human wrote it. Creators of detectors acknowledge that false positives are a known limitation. [1]
2.3. False Negatives:
When AI-generated text is lightly paraphrased or edited, its statistical signature shifts greatly toward human patterns, making it dificult for detectors to correctly identify. Research shows detection accuracy can fall to near "coin-flip levels" (50%) in these scenarios. [2]
2.4. English and Style Bias:
Detectors trained mostly on certain writing samples can misrepresent non-native English, creative, or atypical styles as “AI-generated,” raising ethical and fairness concerns. [3]
3. Paraphrasing and “Humanisation”
Academic and technical work consistently demonstrates how horrendously detectors fail when AI-generated text is paraphrased or rewritten:
"Paraphrasing tools" dramatically reduce detector accuracy. In tests, content originally flagged as AI becomes misclassified as human after simple rewriting. [2]
Recent research benchmarks show that paraphrase-based attacks can produce intermediate text that detectors cannot reliably classify. [4]
These results illustrate that there is no robust fingerprint distinguishing AI from human writing—a fundamental limitation of current detection techniques.
4. Emergence of AI Humaniser Tools
In response to unreliable detection and the “arms race” between generation and classification tools, a new category of services—often called "AI humanisers"—has grown rapidly:
Tools like QuillBot’s AI Humanizer and Grammarly’s AI Humanizer specifically aim to rewrite or refine text so it sounds more human, effectively lowering detector scores. These products highlight that detectors respond to surface patterns rather than context or intent. [5] [6]
Independent platforms such as HumanizeAI, Ryne.ai, or BypassGPT explicitly market what they call “undetectable” AI-to-human text transformation, targeting the weaknesses of detection systems. Often, such in the cases of QuillBot and Grammarly, are designed to exploit flaws in their own detection software.
Recent news has described "Humanizer" plugins that embed detector avoidance logic directly into generation workflows. [6]
To note - Wikipedia is actively fighting against AI slop. [7] [8]
These tools work because current detectors rely on surface cues that can be altered without changing the meaning of a text.
5. The “Arms Race” and Incentives
The development of AI detectors and humanisers is motivated by a combination of economic and psychological factors:
Detectors are often sold as subscription or enterprise products used in education and publishing, which can incentivise marketing claims of accuracy beyond what the evidence supports.
Humaniser tools are marketed to content creators and students seeking to avoid detection or adhere to strict policies, creating a feedback loop where detectors and humanisers co-evolve. This dynamic creates a commercial environment in which organisations that develop or control detection systems can implicitly define what linguistic features are permitted to pass through those systems. In doing so, they are uniquely positioned to market “humanisation” tools optimised to bypass their own detectors, effectively allowing the same entity to influence both detection thresholds and the criteria for successful evasion.
This “arms race” means detectors are often tuned for sensitivity (catch more potential AI text) at the expense of specificity (avoiding false flags), which increases both false positives and false negatives.
6. Consensus and Practical Takeaways
6.1. Evidence Shows Detectors Are Unreliable:
Independent studies consistently find that:
Detectors perform unevenly across styles and edits.
Accuracy deteriorates significantly when texts are paraphrased or refined. [11] [12] [13] [14]
-Research collated on ScienceDirect shows multiple rounds of paraphrasing reduce detection scores dramatically (e.g, TPR@1%FPR dropping from ~99.8% to <60%) [16]
Bias: Evidence indicates that certain human writing styles are systematically misclassified and disproportionately flagged as AI-generated. [3] [9]
Several researchers argue current detection tools should not be used for high-stakes decisions (e.g, academic integrity violations) without additional context. [10]
6.2. Detection Should Be Contextual Not Binary:
Because detectors are probabilistic pattern matchers rather than forensic tools, results should be treated as advisory signals rather than definitive proof of AI use.
Conclusion:
AI text detectors operate on statistical pattern recognition, not direct evidence of authorship. They are subject to false positives, false negatives, style biases, and evasion via paraphrasing and corporate greed. The growth of AI humaniser tools directly exploits these limitations, reinforcing the idea that current detection technology cannot reliably distinguish human from AI writing in all cases.
AI detectors are fundamentally limited by the overlap between human language and the output of large language models. Since modern models mimic human syntax and semantics so closely, detectors relying on statistical patterns or structural cues cannot perfectly distinguish AI from human text. This limitation is both a technical reality rooted in the nature of language and probabilistic modeling, and a market reality, as it drives demand for AI “humanisers” and detection tools whose perceived effectiveness often exceeds their actual capabilities.
Author’s Note: Some of the sources cited discuss the potential benefits or applications of AI-generated text. While these studies have provided valuable insights into AI detection and text generation, I do not endorse the use of AI for generating content in contexts where originality or ethical authorship is required (or at all!!!!). The focus of this article lies solely on understanding and evaluating the limitations of detectors, and on explaining why their results should never be treated as an indisputable certainty.
If you wish to reference this article in academic or professional work, please cite Tumblr as the publication platform, and refer to my pseudonym using the format below:
R. Batrachomyomachy, “Limitations of AI Detectors and the Role of Humanisers,” [22/01/2026]
Platform/Publication: Tumblr
[Post URL] - https://www.tumblr.com/hammerofjusticedotmp3/806472238082736128/why-ai-text-detectors-are-inaccurate-a-brief?source=share
Usage: Citation formatting and element order may be adapted to meet academic or publication standards.
“Hesitation is the mind’s invitation to examine the path before the step. In that brief suspension, we glimpse the tension between who we are and who the moment demands we become.”
If you haven't heard, the em dash has been getting a lot of attention lately…
Because it was trained on pirated work—including freely accessible online writing (like fanfic, academic texts)—ChatGPT picked up patterns and quirks native to human writing.
Including (sigh) the em dash.
There are other victims here (RIP tapestry and delve 🫠), but the appropriation of the em dash—a punctuation mark beloved by writers everywhere—feels especially personal.
A kind of low-grade panic is ensuing. Writers who once memed their own em dash overuse—the greatest punctuation mark ever to grace the control-freak’s lexicon, frankly—are suddenly backing away to avoid accusations.
No. More. We have centuries of dash-abusing writers behind us. We will not sit quietly while AI repurposes our beloved stilted aside—or the just-one-more clarification the sentence demands—or the dramatic pause your comma could never—etc.
You don’t write like AI—AI writes like you.
Defend the em dash.
(Feel free to download/share/stick it where it matters!)
Personal Blog @hammerofjusticedotmp3 - Tumblr Blog | Tumgag
origamiaround
andulkaphoto
areudirty
stormborns
annaxmalina
kianamaiart
