Chapter 8

The workshop

"Phi is my language." The document that opens the project's account of itself, the colophon, begins with those three words and spends a paragraph earning the possessive: the fourteen consonants and five vowels chosen, the single rule set over the grammar, the refusals decided, every ruling signed in a canon file. Then, a few paragraphs on, the same document says the thing a reader in 2026 has learned to brace for, and marks it as a change rather than a constant. It did not always work this way. It does now: the language is built with AI assistance.

Most projects that use these tools handle that fact the way a restaurant handles the freezer: nothing on the menu mentions it. Phi's colophon puts the freezer on the menu. It names the arrangement, describes it in detail, states the ethical costs without discounting them, and stands somewhere specific to answer questions. That arrangement, and the analogy the document reaches for to explain it, are this chapter's subject, and the analogy repays close checking. Underneath it sits the oldest question in the decorative arts, wearing new clothes: what does made mean, when a designer works through other hands?

The blue hands

The colophon's analogy is William Morris, and the choice is not decoration. Morris is already in this book; his novel supplied the boatman of the opening pages, and the language is transmuting all eighty thousand words of it. The same man ran a workshop.

From 1881, Morris & Co. operated out of Merton Abbey, a seven-acre former silk-weaving works on the River Wandle, chosen for water of the special quality the dye vats needed (Mackail 1899). By the 1890s around a hundred people worked there (Grace's Guide n.d.). Dye vats sat below a stained-glass studio; chintzes were block-printed; carpets and tapestries were woven by hand. Morris designed the patterns. Employed printers and weavers executed them, to his standard, under his eye.

What keeps the arrangement from being a gentleman's fiction is that Morris had put his own body through the crafts he directed. From 1875 he made long visits to Thomas Wardle's dye works in Leek, writing that he was "working in Mr. Wardle's dye-house in sabots and blouse pretty much all day long." His first dyeings, his biographer records, were done with his own hands, which became "habitually and unwashably blue" (Mackail 1899). In 1879 he set up a tapestry loom in his bedroom and wove a piece he called "Acanthus and Vine," logging 516 hours in a work diary, often at the loom by first daylight before the day's business (Mackail 1899; Society of Antiquaries n.d.). When he later stood over the vats and looms at Merton Abbey, he was not supervising a mystery.

And yet at his last venture the analogy sharpens into its purest form. At the Kelmscott Press, founded in 1891, Morris designed three typefaces and, by a count handed down from the Press's own circle, 644 initials, borders, and decorative designs (Stirling 2020-21). He specified the paper and rejected inks until one satisfied him. But the punches were cut by Edward Prince, and the type was set by trade compositors, among them Thomas Binning, the shop's union representative. Morris could not do their work. A press worker remembered him trying to lay type and putting it into the wrong boxes: "There, bother it; in the wrong box again!" (Sparling 1924). The Press printed fifty-three books, by the closing count of its own secretary (Cockerell 1898). Nobody has ever suggested they are Prince's books, or Binning's. The designer who could not set a line of type made them, because making, in a workshop, follows the designing mind and the standard it enforces.

Whose pattern

Stated that way, the analogy sounds finished. The history has one more thing to say, and it is the uncomfortable one, so it belongs here.

The veto did not actually cover everything. Morris & Co. wallpapers were printed under contract by Jeffrey & Co. of Islington (Fairclough 1981), and six outside manufacturers are known to have woven textiles under contract to the firm (Parry 1983). More telling still, the designing itself became plural. From the late 1880s a former shop assistant named John Henry Dearle, trained up through the tapestry looms, began producing repeating patterns; after Morris died in 1896, Dearle became Art Director. For most of a century one celebrated pattern, Compton, was attributed to Morris, on his own daughter's word that it was his last design. Scholars now credit it to Dearle (Parry 2013). Museums relabeled it. Dearle's patterns still sell under the Morris & Co. name today.

Notice what the scholars' argument is about. Nobody disputes that the weavers and printers were executing rather than authoring; the whole controversy is over which mind designed a pattern, because that is where attribution lives. The Dearle case is not an argument against the workshop model. It is the workshop model's fine print: the credit follows the designer only while the designer is actually the one designing, and when that quietly stops being true, the label on the work becomes a lie that takes a hundred years to correct.

That fine print is the standard Phi's arrangement has to meet, so it is worth saying precisely what the colophon claims and how a reader could check it.

The record of the veto

The claim is that one person designs and decides, and that the AI, which the colophon calls an instrument rather than apprentices, drafts against those designs and keeps the books. The instrument proposes candidate words built to the designed phonology; it drafts translations against the designed grammar; it holds consistency across a lexicon of 1,275 words and every example sentence in every document, at zero tolerance, through the validator this book met in chapter 7. Nothing merges while any check fails. That is the accounting half, and it is real but unremarkable; a spreadsheet is diligent too.

The half that answers the Dearle problem is the decision record. Phi's repository logs every language decision in a canon file and a numbered decision register, and the reviewed history, over six hundred merged pull requests and still counting, is largely a record of an instrument being corrected. The examples are not hypothetical. A verb glossed hit entered the vocabulary as decision D030 and was retired by decision D033 when an audit found that a general impact verb made interpersonal hitting an unmarked extension; the lexicon now says forceful contact, displacement, damage, or injury, each as its own claim. The question word wia had its gloss corrected from which to how many to match what the grammar had designed it to do. A discourse marker was recoined outright, kewai becoming whekai, because a ruling settled that words of its class begin with a breath, and this one did not. Thrown out, re-glossed, recoined: those are the verbs of a veto in use.

The colophon's own longer answer reaches back before the instrument existed, and this one a reader can audit without leaving the repository. The project began, it says, as a folder of Markdown notes in Obsidian; the first commits are dated April 2021, and by the end of October 2022, 289 commits on, the files already carried the four fricative digraphs, the five pure vowels, the syllable that opens with a consonant and closes with a vowel, the rule admitting no cluster but a digraph, and the exceptionless refusal of diphthongs that still governs every word in the lexicon. wa was already the question word and lo already made a plural. There were Tengwar fonts in the manual directory. Every one of those claims is a git show away, and no model then in existence could have drafted a line of what they return.

The colophon is equally plain about what did not last: nine plain consonants then against ten now, and a particle system rebuilt from the ground up, most of its early forms since reassigned or retired. What survived is the sound of the language, which is both the first thing a listener meets and the hardest thing to change once anything has been written in it. That part was settled years early, by one person typing into a notes application. Whatever the arrangement described above is, it is not a language a machine proposed and a human waved through.

Whether that authorship arrangement deserves the word made is a question each reader settles for themselves. What the record establishes is narrower and checkable: the pattern drawing has stayed at one desk. If it ever migrates, the way Morris & Co.'s quietly migrated to Dearle, the evidence will be in a public ledger within days, not in a museum's corrected label a century on.

The notebook graveyard

The colophon also says why the arrangement exists, and its reason survives checking. A complete language with a real literature has essentially never been finished by one person. The standard case is the most gifted language maker who ever lived. Tolkien worked on his Elvish languages for roughly sixty years, and in his own words the languages were the point of everything else: "The invention of languages is the foundation. The 'stories' were made rather to provide a world for the languages than the reverse" (Carpenter 1981). Yet he revised them endlessly and finished nothing. The scholar closest to the manuscripts concludes that a fixed, complete form was "not even a goal," and that more connected text survives in Gothic, a language dead for a millennium, than in all of Tolkien's inventions combined (Hostetter 2006). What remains of sixty years is glorious fragments: laments, oaths, a hymn.

For Tolkien the endeavor itself was the reward, and on those terms he did not fail. Phi's terms are different. A language built as a practice has to be finishable, learnable, and large enough to live in, and the colophon's arithmetic is blunt: a full grammar, a living lexicon, and a novel's worth of glossed literature outruns one person's remaining decades. Done by hand alone, it would have gone into the notebook graveyard beside Tolkien's fragments. The designer decided the idea deserved to exist at full size more than the process deserved to be pure. That sentence, at least, asks nothing on faith: the shelf from chapter 7 either exists or it does not.

The bill, itemized

Purity was traded away, and the colophon refuses to pretend the price was zero. Two critiques attach to these tools, and both are in better focus now than when the colophon was first written.

The first is consent. The models were trained on vast bodies of human work whose makers were not asked. That is no longer only an accusation; it has case numbers. In Bartz v. Anthropic, a United States federal judge found in 2025 that training a model on lawfully acquired books was fair use, and in the same order found that the company's separate library of pirated copies was not; the piracy claims settled for 1.5 billion dollars, about three thousand dollars a book, with final approval in July 2026. The New York Times's suit against OpenAI survived dismissal and remains in discovery as this book is written. The law is splitting the question finely, acquisition from training, but the moral half needs no docket: the people whose sentences taught these machines were not asked, and a project that benefits should say so rather than wait to be told.

The second is energy. Here honesty means refusing both the panic and the shrug. A single text query costs a fraction of a watt-hour by the best current estimates, with independent figures putting a typical query somewhere between 0.3 and 3 (Ritchie 2025; You 2025). The aggregate is another matter: data centers drew about 415 terawatt-hours in 2024, around 1.5 percent of the world's electricity, and the International Energy Agency's central projection has that roughly doubling by 2030, with AI the main driver and the agency's own reports stressing how poor the underlying disclosure is (IEA 2025). Both facts are true at once. One person's language project is a rounding error inside a rounding error, and it still runs on the same grid as everything else, and the colophon's position is the defensible one: hold the tension with eyes open rather than argue it away with smallness or good intentions.

The norm that converged

When the colophon was written, plain disclosure was a personal scruple. It has since become the norm nearly everywhere the tools are permitted, and the convergence is striking because the domains had no reason to coordinate.

Amazon requires publishers to declare AI-generated text, images, and translations (Amazon Kindle Direct Publishing n.d.). The United States Copyright Office draws its registration line between AI as "a tool to assist" and AI as "a stand-in for human creativity," and requires disclosure of generated content (US Copyright Office 2025). Nature refused AI authorship in the first month of the boom with a one-sentence reason that has aged well: "any attribution of authorship carries with it accountability for the work, and AI tools cannot take such responsibility" (Nature 2023). Steam prints developers' AI disclosures on the store page (Valve 2024). The Linux kernel now accepts machine-assisted patches under a documented rule that an AI may be named as assistant but "only humans can legally certify" the origin of the work, and Fedora's policy says the same in different words (Linux kernel documentation 2026; Fedora Council 2025). And from 2 August 2026, European law requires disclosure when AI-generated text is published to inform the public, with an exception precisely where "the AI-generated content has undergone a process of human review or editorial control and where a natural or legal person holds editorial responsibility" (Regulation (EU) 2024/1689, Article 50).

The community that will ask

There is also a plainer reason to disclose, and it deserves naming without decoration: the first community this language will meet has strong feelings on the subject, and they are not welcoming ones. Solarpunk Magazine's submission rules refuse "any work created or altered by generative artificial intelligence programs," citing data scraping, plagiarism, and the displacement of human artists and authors (Solarpunk Magazine n.d.). The largest solarpunk forum requires AI-assisted posts to be labeled. One writer in the movement keeps a vouched list of editors and cover designers who have pledged to work without the tools (Quinn 2024). That community occupies the refusal end of the disclosure spectrum, and its reasons are the same two costs itemized above, weighed by people who concluded the trade was not worth making.

Phi cannot meet that community by hiding its process, and should not try to win the argument by omission, least of all here, since a refusal held on ethical grounds is the very design move this book spent chapter 2 admiring. Some solarpunk readers will read the colophon and decline the language on the spot. That outcome is built into the price, and the colophon was written so that the decline happens on true information.

Morris would have recognized every part of this position, including its discomfort. His firm sold beautiful things to the rich while he preached socialism, and he did not resolve the contradiction; he confessed it, in a remark that reached print second-hand and six decades late, about spending his life "ministering to the swinish luxury of the rich" (Lethaby 1935). He paid wages above the market rate, wrote openly that profit sharing under competition could not go far, and kept working (Lazarus 1886). Holding a compromise in the open, without either abandoning the work or acquitting yourself, turns out to be a craft tradition too.

The vats at Merton Abbey needed days, and a man's arms to the elbow, to strike a blue that would not fade. Phi's instrument drafts in seconds. Between those two speeds sits the one thing that did not accelerate and was never asked to: the deciding. Every word in the language waited at one desk for one person to hold it up to the light, and the fast tool's whole output exists so that a slow practice, a language you cannot hurry, could exist at full size. That is the trade. The designer is at peace with it, says so in writing, and remains at the desk, where questions are answered.

References

Amazon Kindle Direct Publishing. n.d. Content Guidelines. The AI-content disclosure requirement, in force since September 2023. https://kdp.amazon.com/en_US/help/topic/G200672390. Accessed 22 July 2026.

Bartz v. Anthropic PBC. 2025. Order on fair use, No. 3:24-cv-05417 (N.D. Cal., 23 June 2025). Settlement of the piracy claims granted final approval 20 July 2026.

Carpenter, Humphrey, ed. 1981. The Letters of J.R.R. Tolkien. Boston: Houghton Mifflin. Letter 165, to the Houghton Mifflin Co., 30 June 1955.

Cockerell, Sydney. 1898. A Note by William Morris on His Aims in Founding the Kelmscott Press, Together with a Short Description of the Press by S.C. Cockerell, and an Annotated List of the Books Printed Thereat. Hammersmith: Kelmscott Press. The annotated list runs to fifty-three books, one of them printed twice, in sixty-six volumes.

Fairclough, Oliver. 1981. Textiles by William Morris and Morris & Co. 1861-1940. London: Thames and Hudson. Page 49 for Jeffrey and Co.'s contract printing of the wallpapers.

Fedora Council. 2025. AI-Assisted Contributions Policy. Version 1.0, approved October 2025, announced 22 October. https://docs.fedoraproject.org/en-US/council/policy/ai-contribution-policy/.

Grace's Guide. n.d. Morris and Co. Grace's Guide to British Industrial History. https://www.gracesguide.co.uk/Morris_and_Co. Accessed 22 July 2026.

Hostetter, Carl F. 2006. Elvish as She Is Spoke. In The Lord of the Rings 1954-2004: Scholarship in Honor of Richard E. Blackwelder, edited by Wayne G. Hammond and Christina Scull, 231-255. Milwaukee: Marquette University Press.

International Energy Agency. 2025. Energy and AI. Paris: IEA. https://www.iea.org/reports/energy-and-ai. Updated projections in Key Questions on Energy and AI, 2026.

Lazarus, Emma. 1886. A Day in Surrey with William Morris. Century Magazine, July 1886. Quotes Morris's letter of 21 April 1884 on wages and profit sharing.

Lethaby, W.R. 1935. Philip Webb and His Work. London: Oxford University Press, Humphrey Milford. The "swinish luxury" remark, made to Lowthian Bell around 1877, reached print here via Alfred Powell.

Linux kernel documentation. 2026. AI Coding Assistants. Documentation/process/coding-assistants.rst, merged to mainline January 2026. https://www.kernel.org/doc/html/latest/process/coding-assistants.html.

Mackail, J.W. 1899. The Life of William Morris. Two volumes. London: Longmans, Green. The dye-house quotes are in volume 1, chapter X; the tapestry diary and dawn habit at volume 1, page 373; the Merton Abbey site in volume 2, pages 34 and 35.

Nature. 2023. Tools such as ChatGPT threaten transparent science; here are our ground rules for their use. Nature 613: 612. https://doi.org/10.1038/d41586-023-00191-1.

Parry, Linda. 1983. William Morris Textiles. London: Weidenfeld and Nicolson. Revised edition 2013, London: V&A Publishing; page 259 for the Compton reattribution.

Quinn, Susan Kaye. 2024. Solarpunk Don't Need No AI. Bright Green Futures, 18 October 2024. https://brightgreenfutures.substack.com/p/solarpunk-dont-need-no-ai.

Regulation (EU) 2024/1689 (the EU Artificial Intelligence Act). 2024. Official Journal of the European Union, 12 July 2024. Article 50 transparency obligations, applying from 2 August 2026.

Ritchie, Hannah. 2025. What's the carbon footprint of using ChatGPT or Gemini? [August 2025 update]. Sustainability by Numbers, 21 August 2025. https://hannahritchie.substack.com/p/ai-footprint-august-2025.

Society of Antiquaries of London. n.d. The 'Acanthus and Vine' Tapestry. Collections highlight, with the collections record for object KM032 identifying Morris's 1879 work diary (National Art Library MSL/1939/2637). https://www.sal.org.uk/collections/explore-our-collections/collections-highlights/the-acanthus-and-vine-tapestry/.

Solarpunk Magazine. n.d. Submissions. The policy on artificial intelligence, in force since January 2023. https://solarpunkmagazine.com/submissions/. Accessed 22 July 2026.

Sparling, H. Halliday. 1924. The Kelmscott Press and William Morris Master Craftsman. London: Macmillan. Page 74 for the wrong-box recollection.

Stirling, John. 2020-21. The People of the Kelmscott Press. Journal of William Morris Studies 24(1-2): 95-105.

US Copyright Office. 2025. Copyright and Artificial Intelligence, Part 2: Copyrightability. Washington: Report of the Register of Copyrights, January 2025. The registration disclosure duty is in "Copyright Registration Guidance: Works Containing Material Generated by Artificial Intelligence," 88 FR 16190, 16 March 2023.

Valve. 2024. AI Content on Steam. Steamworks Development announcement, 9 January 2024. https://steamcommunity.com/groups/steamworks/announcements/detail/3862463747997849619.

You, Josh. 2025. How much energy does ChatGPT use? Epoch AI, 7 February 2025. https://epoch.ai/gradient-updates/how-much-energy-does-chatgpt-use.

Phi sources

The colophon quoted throughout is colophon.md at the repository root. The rulings and their dates are in canon.md; the coin and retire record for the impact verb is decisions D030 and D033 in project/development_log.md, with the full audit in documents/evaluation/peace_linguistics_vocabulary_audit.md. The wia regloss and the kewai to whekai recoining are settled decisions in canon.md. The validator is scripts/validate_examples.py, run by continuous integration on every change (.github/workflows/validate.yml). The Morris source witness, all 80,908 words of it, is texts/news_from_nowhere/source.txt.