MAZE
Solve the World's Most Challenging Puzzle
by Christopher Manson
A maze in the shape of a book, now in the format of a tumblr blog.
I invite you to enter my Maze. I say it is mine, because despite who else I might be, I am the architect as well as your guide. Your first goal is to find the shortest route through the Maze—a simple task, I assure you, if you know what to look for. I have planted clues throughout for your interpretation—or misinterpretation. Indeed, you will be fascinated by the Maze's ambiguity, stimulated by its mystery, stymied by its riddle. But fear not! I will be with you all the way. Fear not, that is, if
you truly believe that my clues or I can be trusted.
Enter room 1. Which door should you take from here? Someone in the narrative uses the word “story,” and the same word appears above the door to room 20. Is that the connection? Is there a connection? Give it a try and go to room 20, which is peculiar in its own way. Just inside the door to room 27 you see what looks like the bottom half of an archer’s arrow—an arrow pointing the way perhaps? I will not tell. Perhaps it wouldn't help if I did. It is up to you to decide, as you move from room to room, hoping that fact is not illusion and that your best judgment has not led you astray.
Tempted? Test your wits against mine. I guarantee that my Maze will challenge you to think in ways you've never thought before. But beware…one wrong turn and you may never escape.
DIRECTIONS
CONTEST RULES
PROLOGUE
Transcriber's note: it is my belief that the translocation of the Maze into an online, interactive Tumblr format constitutes a transformative work and is as such a fair use of Mr Manson's intellectual property, especially as the original book has long since gone out of print. However, admirers of the Maze and respecters of Mr Manson's artistry are strongly encouraged to seek out and purchase one of the remaining copies of the original book through online or brickspace retailers.
“Transformative work is about repairing the deficiencies of canon” no, transformative works is about using canon as raw material. Sometimes that means fixing the canon. Sometimes it means making the canon worse. Sometimes it means slapping a pair of Groucho glasses and a silly hat on the canon and hoping nobody notices what’s really going on.
Let’s talk about the oft-repeated accusation that AI is “stealing” or “plagiarising” content. It’s a loaded claim, designed to provoke — and like most loaded claims, it deserves careful unpacking.
🧠 When a Human Learns from Art…
Let’s say you read a novel. You love it. It influences how you think about pacing, dialogue, maybe even inspires a character or two in your own writing. That’s how humans learn: by absorbing, remixing, iterating. If you quote the novel directly without attribution, that’s plagiarism. But if you write something inspired by it — that’s transformative. That’s how all creative culture works.
We do not accuse someone of “stealing” when they say they were influenced by Dickens or Morrison or Kurosawa.
🤖 But When AI Learns the Same Way…
It doesn’t act like a pirate library or a photocopier, and it doesn’t normally regurgitate books word-for-word. What it does is learn patterns — structure, syntax, rhythm, pacing. In some edge cases models can reproduce memorised snippets, but that’s a technical and governance problem, not their primary behaviour.
Yet somehow, when an AI synthesises a new sentence influenced by its training data, we call it theft. Why?
⚖️ Where’s the Actual Plagiarism?
Plagiarism is about presenting someone else’s language or ideas as your own without proper credit — substantial uncredited copying of specific content. It’s an ethical and academic violation, not a technical process.
AI doesn’t intend anything. It has no authorship or ego; it just outputs patterns. Any ethical responsibility sits with the humans using it.
When a model reproduces verbatim text from training data, that’s a technical and governance problem — a risk to privacy and copyright that needs to be mitigated — not proof that every single output is theft.
🧂 A Dash of Hypocrisy?
You’ll often hear:
“AI is just scraping artists’ work and remixing it!”
But then we ask:
“Have you ever written fanfic? Used a prompt list? Played with visual references? Quoted a line in your fic title? Watched a tutorial? Written like your fave author for fun?”
Because if the answer is yes… congrats. You’re already engaging in transformative work, the same fundamental mechanism AI relies on. The only real difference? You're squishier.
🧾 “But AI Was Trained on Copyrighted Material Without Permission!”
This is the big one, isn’t it?
It’s true that many AI models were trained on large datasets scraped from the public web — which include copyrighted works that were publicly accessible. That’s not the same thing as deliberately raiding pirate libraries, but it’s also not ethically trivial, and some lawsuits argue that certain systems scraped paywalled material without consent.
Legally, this is still an active fight. In places like the U.S., regulators and courts haven’t given a single, final answer. Some legal scholars and early court decisions say that using copyrighted works as training data can count as fair use when it’s about learning patterns rather than reproducing the originals, especially if the outputs don’t compete with or substitute the source material. Others disagree, or are still deciding.
But “it touched copyrighted material” is not, by itself, proof of theft. If reading a copyrighted book teaches you how to write your own — did you “steal” it?
If “learning from content” equals “theft,” then your memory is a crime scene.
Does that mean there are no concerns? Of course not. Transparency, consent, opt-outs — all of those matter. But shouting “it was trained on IP without permission!!” isn’t a moral mic drop. It’s a simplification that falls apart under scrutiny.
💬 So, is AI really stealing?
Only if you are.
🔗 Further Reading / Sources:
U.S. Copyright Office – Copyright and Artificial Intelligence, Part 3: Generative AI Training (2025)
Overview of how U.S. law currently thinks about training on copyrighted works; concludes legality depends on context and fair-use analysis.
https://www.copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-3-Generative-AI-Training-Report-Pre-Publication-Version.pdf
Carlini et al. – Extracting Training Data from Large Language Models (USENIX Security, 2021)
Shows that verbatim memorisation can happen in edge cases — a real risk, but not the default behaviour of these models.
https://www.usenix.org/system/files/sec21-carlini-extracting.pdf
Micaela Mantegna – ARTificial: Why Copyright Is Not the Right Policy Tool to Deal with Generative AI (Yale Law Journal Forum, 2024)
Argues that stretching copyright to “solve” AI problems is a bad fit and risks harming creativity and the public.
https://www.yalelawjournal.org/forum/artificial-why-copyright-is-not-the-right-policy-tool-to-deal-with-generative-ai
Electronic Frontier Foundation – AI and Copyright: Expanding Copyright Hurts Everyone—Here’s What to Do Instead (2025)
Explains how using AI panic to expand copyright would undercut fair use, research, and small creators.
https://www.eff.org/deeplinks/2025/02/ai-and-copyright-expanding-copyright-hurts-everyone-heres-what-do-instead
Cory Doctorow – Copyright Won’t Solve Creators’ Generative AI Problem (Pluralistic, 2023)
A creator-centred critique of copyright maximalism as a fake solution to AI and labour issues.
https://pluralistic.net/2023/02/09/ai-monkeys-paw/
Every now and then I think about how subtitles (or dubs), and thus translation choices, shape our perception of the media we consume. It's so interesting. I'd wager anyone who speaks two (or more) languages knows the feeling of "yeah, that's what it literally translates to, but that's not what it means" or has answered a question like "how do you say _____ in (language)?" with "you don't, it's just … not a thing, we don't say that."
I've had my fair share of "[SHIP] are [married/soulmates/fated/FANCY TERM], it's text!" "[CHARACTER A] calls [CHARACTER B] [ENDEARMENT/NICKNAME], it's text!" and every time. Every time I'm just like. Do they though. Is it though. And a lot of the time, this means seeking out alternative translations, or translation meta from fluent or native speakers, or sometimes from language learners of the language the piece of media is originally in.
Why does it matter? Maybe it doesn't. To lots of people, it doesn't. People have different interests and priorities in fiction and the way they interact with it. It's great. It matters to me because back in the early 2000s, I had dial-up internet. Video or audio media that wasn't available through my local library very much wasn't available, but fanfiction was. So I started to read English language Gundam Wing fanfic before I ever had a chance to watch the show.
When I did get around to watching Gundam Wing, it was the original Japanese dub. Some of the characters were almost unrecognisable to me, and first I doubted my Japanese language ability, then, after checking some bits with friends, I wondered why even my favourite writers, writers I knew to be consistent in other things, had made these characters seem so different … until I had the chance to watch the US-English dub a few years later. Going by that adaptation, the characterisation from all those stories suddenly made a lot more sense. And the thing is, that interpretation is also valid! They just took it a direction that was a larger leap for me to make.
Loose adaptations and very free translations have become less frequent since, or maybe my taste just hasn't led me their way, but the issue at the core is still a thing: Supernatural fandom got different nuances of endings for their show depending on the language they watched it in. CQL and MDZS fandom and the never-ending discussions about 知己 vs soulmate vs Other Options. A subset of VLD fans looking at a specific clip in all the different languages to see what was being said/implied in which dub, and how different translators interpreted the same English original line. The list is pretty much endless.
And that's … idk if it's fine, but it's what happens! A lot of the time, concepts -- expressed in language -- don't translate 1:1. The larger the cultural gap, the larger the gaps between the way concepts are expressed or understood also tend to be. Other times, there is a literal translation that works but isn't very idiomatic because there's a register mismatch or worse.
And that's even before cultural assumptions come in.
It's normal to have those. It's also important to remember that things like "thanks I hate it" as a sentiment of praise/affection, while the words translate literally quite easily, emphatically isn't easy to translate in the sense anglophone internet users the phrase.
Every translation is, at some level, a transformative work. Sometimes expressions or concepts or even single words simply don't have an exact equivalent in the target language and need to be interpreted at the translator's discretion, especially when going from a high-context/listener-responsible source language to a low-context/speaker-responsible target language (where high-context/listener responsible roughly means a large amount of contextual information can be omitted by the speaker because it's the listener's responsibility to infer it and ask for clarification if needed, and low-context/speaker-responsible roughly means a lot of information needs to be codified in speech, i.e. the speaker is responsible for providing sufficiently explicit context and will be blamed if it's lacking).
Is this a mouse or a rat? Guess based on context clues! High-context languages can and frequently do omit entire parts of speech that lower-context/speaker-responsible languages like English regard as essential, such as the grammatical subject of a sentence: the equivalent of "Go?" - "Go." does largely the same amount of heavy lifting as "is he/she/it/are you/they/we going?" - "yes, I am/he/she/it is/we/you/they are" in several listener-responsible languages, but tends to seem clumsy or incomplete in more speaker-responsible ones. This does NOT mean the listener-responsible language is clumsy. It's arguably more efficient! And reversely, saying "Are you going?" - "I am (going)" might seem unnecessarily convoluted and clumsy in a listener-responsible language. All depending on context.
This gets tricky both when the ambiguity of the missing subject of the sentence is clearly important (is speaker A asking "are you going" or "is she going"? wait until next chapter and find out!) AND when it's important that the translator assign an explicit subject in order for the sentence to make sense in the target language. For our example, depending on context, something like "are we all going?" - "yes" or "they going, too?" might work. Context!
As a consequence of this, sometimes, translation adds things – we gain things in translation, so to speak. Sometimes, it's because the target language needs the extra information (like the subject in the examples above), sometimes it's because the target language actually differentiates between mouse and rat even though the source language doesn't. However, because in most cases translators don't have access to the original authors, or even the original authors' agencies to ask for clarification (and in most cases wouldn't get paid for the time to put in this extra work even if they did), this kind of addition is almost always an interpretation. Sometimes made with a lot of certainty, sometimes it's more of a "fuck it, I've got to put something and hope it doesn't get proven wrong next episode/chapter/ten seasons down" (especially fun when you're working on a series that's in progress).
For the vast majority of cases, several translations are valid. Some may be more far-fetched than others, and there'll always be subjectivity to whether something was translated effectively, what "effectively" even means …
ANYWAY. I think my point is … how interesting, how cool is it that engaging with media in multiple languages will always yield multiple, often equally valid but just sliiiiightly different versions of that piece of media? And that I'd love more conversations about how, the second we (as folks who don't speak the material's original language) start picking the subtitle or dub wording apart for meta, we're basically working from a secondary source, and if we're doing due diligence, to which extent do we need to check there's nothing substantial being (literally) lost -- or added! -- in translation?
On one hand, when I make "X is literally fanfic" comments, they're like 60% jokes because 'fanfic' is a specific mode of storytelling specific to fan culture and relationship to copyright and all that jazz.
On the other hand, given the percentage of modern genre fiction that is:
A) official remakes, reboots, or spin off of existing franchises
B) unofficial explicit or coded retellings of existing works
Or
C) written by authors who cut their teeth in modern fanfic communities
I do think it's fair to say "a grounding in transformative work is helpful in navigating modern spec fic"
Sometimes I forget how much work went into this and then I look at old comparisons like this.
I really like this side by side because it shows the amount of mesh modding and texture editing that went into cleaning up the base model's defective right eye. This was not a quick fix. It took a lot of trial, error, and patience.
For clarity, the image on the left is Stefano Valentini, the original base model. The image on the right is Sunder, an OC I created using that model as a starting point.
Just like some roleplayers use celebrity face claims or AI these days, I chose to work with existing video game models and transform them through mesh mods and texture edits to build original characters.
I do not own the models, as these are purely transformative and recreational works, but I do own the stories, and over time Sunder and others have developed deep backstories and a lot of personal meaning for me.
If you saw the OC post I shared back in October on @sadisticlens, that shows the current version of him. This image is a bit outdated, but it still highlights the contrast and all the work that went into transforming the base model.
