Photocopier No More: The Reckoning with AI Creativity Has Arrived

Two esoteric programming events bubbled up this past week. If you’re not into computer science, they may appear irrelevant to you. They’re not. The arcane managed to bring to life our pressing questions about whether AI can create or is a regurgitation machine.

The first involves a face-off between the developers of “Chardet,” one of the building blocks many apps depend on to function. These underpinnings, known as libraries, are bits of code that do tasks common enough to justify packaging them in a reusable form.

Chardet’s job is to enable programs written in Python, the most popular present programming language, to look at a text file and figure out how it is encoded. Everything text-based we see or type has an encoding to it. Old style on PCs was known as “ASCII” plain text, but when you send a friend an emoji or use accents while communicating to an associate in Mexico, you’re using another one called Unicode. Various others also exist and if a program fails to check the encoding or gets it wrong, that’s when we see scrambled characters or blank boxes in lieu of the words and symbols we expected.

Mark Pilgrim wrote this particular library for Python several decades ago before deciding to retreat out of the public view. He disappeared, but the need for what his tool accomplished remained. Because he placed the library under a so-called copyleft open source license, others were able to keep developing it, under the proviso they similarly shared their changes. So it went for a decade and a half.

Last week, that changed. Those who had maintained his code unveiled a major new release of Pilgrim’s library they claimed was a clean room-style rewrite. Such an effort has been attempted with code — often by competitors — many times over the years, utilizing programmers who didn’t have access to the original code but were told what it needed to do. The idea is to replicate functionality without infringing on the original’s copyright.

The caretakers of Pilgrim’s Chardet certainly weren’t eligible to do clean room work themselves. Years of work in the code would undercut the credibility of any feigned ignorance. Their solution was to task an AI coding agent with the clean room work they couldn’t do. The agent, presumably primed with guidance from said maintainers, produced “new” code and they relicensed that code under a very different license.

Their maneuver in short order brought Pilgrim out of his self-imposed internet hermitage to assert his successors had breached his code’s share-alike requirement. His complaint raises a serious question: can someone circumvent a legal license by operating an AI agent as the “clean room programmer” or is such an effort inherently contaminated by the human puppeteer?

In other words, how independent is the AI “creator” from its human operator and other humans’ work it has been trained on? There are three possible answers and each has profound implications for the future of creativity and AI.

One could assert that a present day AI coding agent isn’t able to do a true clean room implementation of code that’s been publicly released because it has been trained on as much of that available code as possible. That’s how Large Language Models (LLMs) like Claude and GPT work: the companies behind them feed them vast quantities of known materials and they “learn” from that data. The AI may be more familiar with Pilgrim’s code — and thus less clean — than the human maintainers who invoked it to create the plausible distance needed to justify the claim the code was new and thus could be relicensed.

Would we solve anything if we could guarantee the AI itself somehow had been prevented from knowing about the particular code it was tasked with mimicking? Not necessarily.

The bigger question lies further downstream: who is the creator of new AI generated code when we accept it is, indeed, new code? The answer will dramatically shape the landscape of intellectual property — not just code, but writing, music and art — for our future.

If we say the maintainers of Chardet can’t legally do what they did because they are the actual authors of the AI output, that has huge implications for the future of copyrights. Following this logic, AI is a tool like a paintbrush or a keyboard. The person typing on the keyboard still brings the knowledge and holds the responsibility, even if they swap from one keyboard to another. Is trading keyboard for AI any different?

On the other hand, what if we say the AI was as independent as a series of programmers put in that proverbial isolated room to recreate something? This would put the maintainers of Chardet in the clear, even if we dislike what they did. We would also be saying a machine can now perform a genuinely creative act.

Is that true?

Critics of generative AI have argued against the notion for years now. Even setting aside whether Claude has studied Chardet itself as part of its training, is it merely imitating other code? The first appearances of significant AI artwork a few years back catalyzed the conversation. When one piece won an award over human created work, skeptics questioned whether it was actually a creative work at all or simply reenactment of others’ art.

The distinction isn’t as clear cut as it may first seem. I’ve long argued against the notion that we should prevent AI from “learning” via any and all human work. Why? Because that’s precisely how learning happens, whether human or machine. We know it would be absurd to insist kids never read or look at copyrighted works lest their first stories or pictures might infringe on copyrights.

The best creative minds of today probably wrote fan fiction or traced a beloved cartoon character as kids to get started, proudly holding it up as their own when they finished. Even once those gifted with creativity move well beyond being copycats, they don’t entirely forget their intellectual genealogy. We all absorb and build upon the work of those who influenced us.

No one writes in a void. No one paints in a void. No one sings in a void. To do so would be unintelligible. The fact that we don’t is why we can identify “eras” (not in the Taylor Swift sense): figures are breathing in the same cultural air and each other’s work and, in retrospect, we see the cross-pollination unmistakably.

Whether human or AI, a far better measure is whether the “creator” — and I’m not declaring AI creative here, just riding out a hypothetical — is imitating or inspired by the past. There are plenty of gray areas, but we all recognize the difference between mere parroting and creative extension.

Has AI managed to extend? With, say, art, it’s something we’re still resolving. Granted, early generative AI had an unfortunate penchant for slapping artists signatures on its work, encouraging the accusation it is glorified photocopier. But again, that’s like how we learn. The kid studying impressionism might also mistake a scribble in the lower right as an essential aspect before learning it was actually Monet’s claim as creator.

If only the creating happened in a less subjective realm so we could better measure if it had happened. That brings us to the second, more esoteric AI event of the last few days.

Don Knuth is a legendary computer scientist and author of the magisterial The Art of Computer Programming; he has stood large in his field since his first volume was published in 1968. A week or so ago, he opened an essay on AI with “Shock! Shock!” As he continues his work on the remaining planned volumes of The Art nearly sixty years after the first, he tucked an “open question” of a mathematical problem he mused on but had not resolved into his draft with the intention readers could consider it in the years to come. It would not remain “open” long enough to reach publication.

A colleague, Filip Stappers, felt inspired to plug Knuth’s open question into Anthropic’s Claude and, with Stappers’s gentle prodding, the AI solved the question the legendary scientist had not yet resolved himself. Just as a human programmer might do, it iterated through different plausible methods until it found a general formula that worked in all cases. The brilliant computer scientist had noodled on this for weeks without finding a satisfactory theorem; Claude reached one in an hour.

Knuth writes, “What a joy it is to learn not only that my conjecture has a nice solution but also to celebrate this dramatic advance in automatic deduction and creative problem solving.” Could a human of this man’s ability have solved the problem? I’m sure at some point. But had he — or anyone else — yet done so? It appears not.

Could this be one of the first provable examples of AI creating something new? Perhaps.

One might object that math is not an art, that it is defined and anything not known is simply undiscovered. That AI can be used to speed up discovery but not creation. Here again is a line blurrier than it may first appear. Intentionally or intuitively as it may be for the individual, formulae undergird so much of the structure of art.

What has this to do with Chardet? If a preeminent mathematician and programmer, someone who at 88 is still writing defining theory in his field, can see AI run with one of his questions and celebrate that AI did genuine discovery, is it so far fetched to think Claude or Codex could recreate some code legitimately?

We are going to have to dwell in gray areas more than we want to. A human being can create a new cartoon character and a human being can also create an illegal ripoff of Garfield. If AI can do the latter, how does that preclude it from also doing the former?

When courts have evaluated clean room work, how clean that clean room is weighs significantly. In the Chardet debacle, the prompts given to the AI should probably loom large.

Knuth’s essay seems to rule out the option of dismissing AI as a mere robo-parrot. The two remaining options both give AI a level of credit many have hesitated to grant. In this, well, era in which AI can be a collaborator to creative acts, the telling determination isn’t whether AI could have made something, but who the collaborators were on this particular act.

If Pilgrim were to sue over Chardet, the prompting humans gave the AI will necessarily be front and center at resolving the question. Philosophically, this would be helpful to explore, even if this particular incident never devolves into legal blows.

Did the maintainers feed Claude enough “prompt” information to taint its work? Or did they stay within the bounds those giving instructions to human clean room programmers — presumably under careful lawyerly supervision — would have kept?

The visceral reaction we have to this particular controversy suggests we recognize AI’s output cannot be neatly categorized as a mere remixing of the already known nor something necessarily independent of its human operator.

We can intuit the logical follow up: if AI is not merely spitting out a pattern matching a fancy search query, prompt engineering is a creative exercise, not just a pull of a slot machine lever. Combined with Knuth’s observations, we find ourselves faced with the idea that prompt-based generation is a collaboration of two creators, an augmentation of a human mind with a not-human something. That is what makes the attempted end run around Pilgrim’s copyright so disquieting.

Should we say prompting is not relevant to the output and it was indeed “clean room,” we will see a legion of copyrighted works upended by similar machinations of AI regeneration. But, if we say prompts are inextricably intertwined, we put to rest the idea that AI work is not copyrightable for lack of a creative human operator.

However we as a society land legally or culturally — AI as genuinely an independent creator or AI as a human augmentation or “it depends” — even the mildest AI as co-creator path will alter our world. These two examples over the last week are tremors; the seismic impact of the question is unavoidably ahead of us.