Discussion The numerous false positives of artificial intelligence detectors

[TheVanishingPoint]

In recent days, and in the past as well, I've noticed something curious: when I write a text in Icelandic (the language I use most) and then pass it to a translator, the AI detector completely changes its assessment. A text that I wrote as "100% human" drops to maybe 27% or 60% "AI" after translation, and sometimes it's even detected as completely artificial. Has something similar ever happened to you with your own language?

This below is an example, yet nothing has been distorted because only the translator was used.

Líknardráp er ekki afneitun lífsins heldur viðurkenning á rétti hvers einstaklings til að setja mörk eigin þjáningar. Lífið er ekki skylda og að breyta tilvist í ævilanga refsingu fyrir þann sem er fastur í líkamlegum eða andlegum sársauka er birtingarmynd grimmdar, dulbúin sem siðferði. Það er engin reisn í því að halda lífi í líkama með gerviaðferðum þegar hann hefur þegar sagt allt sem hann gat sagt, og það er ekkert kærleiksríkt við að neyða einhvern til að þrauka þegar eina óskin er að hætta. Sönn frelsi felur í sér möguleikann á að velja hvenær og hvernig kveðja á, án ótta við dóm eða refsingu. Að hafna líknardrápi þýðir að líta á manneskjuna sem eign laga eða trúarbragða en ekki sem eiganda sjálfs síns.



Euthanasia is not the denial of life but the recognition of the right of each individual to set the limits of their own suffering. Life is not a duty, and turning existence into a lifelong punishment for someone trapped in physical or mental pain is a manifestation of cruelty, disguised as morality. There is no dignity in keeping a body alive by artificial means when it has already said all it could say, and there is nothing loving in forcing someone to endure when their only wish is to end it. True freedom includes the ability to choose when and how to say goodbye, without fear of judgment or punishment. To reject euthanasia means to see the person as the property of law or religion, not as the owner of himself.

This text also obviously appears 100% human because it was written by me.



Hún sat við gluggann, horfði út á hafið þar sem öldurnar léku sér í kvöldlogninu. Hann hafði lofað að koma áður en sólinn færi til hvílu. Þegar hún heyrði létt skref nálgast, vissi hún að loforðið stæði. Hann kom inn, með saltbragð sjávarins á vörunum og bros sem ylti upp hjarta hennar. Þau sögðu fá orð; augun þeirra sögðu allt. Í þögninni milli þeirra var tónlist, í snertingu handanna var eilífð. Þau stóðu þar saman, á meðan sólin seig niður og himinninn málaðist með litum sem hvorki tími né fjarlægð gátu strokið út.

However, when the English translator is applied, it turns out to be 100% false, according to chat gbt-0 (the most reliable AI detector). You can try it.



She sat by the window, looking out at the ocean where the waves played in the evening calm. He had promised to come before the sun went down to rest. When she heard light footsteps approaching, she knew the promise was kept. He came in, with the salty taste of the sea on his lips and a smile that warmed her heart. They said few words; their eyes said everything. In the silence between them there was music, in the touch of their hands there was eternity. They stood there together, while the sun sank and the sky was painted with colors that neither time nor distance could erase.


---------

This graph demonstrates how reality can be distorted from the initial document, revealing the truth: even the so-called "most reliable" AI-based screening tool, GPTZero, can continually produce sensational fakes. It highlights the unreliability of these tools, which are increasingly used for smear and denigrating campaigns rather than constructive criticism. In the hands of some, they are machines that destroy reputations, sow doubt, and distrust without any real cause. GPTZero and other similar detectors are inadmissible as evidence in any court of law, and it's unclear why they should be admissible in other environments.

 

[EternalShore]

I've encountered the same issue with resumes, likely because all good resumes sound nearly identical, which sounds like AI~ only bad resumes sound unique~
In regards to this, it's likely because Google Translate uses AI in its translating tools~ additionally, it will not translate all the intricacies of the way your speaking style in Icelandic and merely spit back out the literal meaning of what you said in a common sense English way~ You'll likely see that that same English text does not translate exactly the same way as it was translated originally back into Icelandic~

 

[Pluto]

 

[TheVanishingPoint]

I've encountered the same issue with resumes, likely because all good resumes sound nearly identical, which sounds like AI~ only bad resumes sound unique~
In regards to this, it's likely because Google Translate uses AI in its translating tools~ additionally, it will not translate all the intricacies of the way your speaking style in Icelandic and merely spit back out the literal meaning of what you said in a common sense English way~ You'll likely see that that same English text does not translate exactly the same way as it was translated originally back into Icelandic~

In 2016, Google Translate went through a major shift, moving away from the old phrase-based statistical approach to embrace neural translation. Before that change, sentences were broken into smaller pieces and then stitched back together, often resulting in stiff or awkward output. Now the system processes the entire sentence in one go, making it sound far more natural. This improvement comes from the use of deep neural networks and attention-based models that focus on the key elements of the original text, as well as the Transformer architecture, which was added later. Over time, the platform has also learned to handle many languages within a single model and even translate language pairs it was never directly trained on by leveraging intermediate languages.
Although it is a form of artificial intelligence, a translation produced by this system often slips past AI detectors because it isn't created from scratch. The meaning, structure, and human logic of the source material are still there. On top of that, the translation often preserves quirks, ambiguities, or stylistic choices from the original features that purely generative models tend to smooth out. Detectors typically work by looking for patterns such as overly uniform sentence lengths, repeated grammatical structures, evenly spaced punctuation, or a lack of the small imperfections that appear in spontaneous writing. When those patterns aren't present or are diluted by the variability of the source text, the detector can make the wrong call.
That's where the biggest weaknesses show up: entirely human writing can be marked as "AI-generated" if it fits certain patterns, while machine-generated text can pass unnoticed if it's been carefully edited. Academic researchers, professional translators, journalists, and technical writers are all at risk of being misclassified. A person writing in a second language might produce flawless grammar but miss the natural variation of a native speaker; technical or scientific professionals often repeat formulas and structures for the sake of clarity; legal or manual translators keep strict consistency in terminology; and some people revise obsessively until their writing is almost unnaturally tidy. All of these are examples of genuine human work that, to a detector, might look like it came from a machine.
When a detection system can't tell the difference between a machine translation or a well-structured human text and something a model has generated, it shows just how wide the margin of error really is and how unreliable these tools become if someone tries to use them as absolute proof of authorship.