For years I was not really interested in the subject of AI consciousness, it was one of those subjects that I equated with unreasonableness at the extremes of the debate, but in which I did not have any strong opinions. This has been changing recently as the capabilities of AI continue to improve, and I have begun to wonder about what could any sort of intelligence could mean from a legal perspective, particularly for copyright.
Now there is a fascinating paper from Anthropic that has caused quite a stir in AI research circles, and it is entitled “Verbalizable Representations Form a Global Workspace in Language Models”. The paper opens the door to arguments that the way in which language models operate and reason matches some models of human consciousness, and provides some compelling evidence for the existence of a workspace that could prove to be analogous to how the human subconscious operates. It is important to stress that the paper does not in any way claim that AIs are conscious, but that they are certainly more than the glorified autocomplete that many people think of when discussing current AI models. I suspect that this paper will become a landmark one in the future, and maybe it will start shifting some opinions that still regard LLMs as nothing more than next-token predictors.
The paper got me thinking about possible implications for copyright. These are just my first thoughts on what is certainly a very complex area, and as many such musings, they may not survive the first serious examination. But the potential legal implications of this research are at the very least intriguing, but feel free to disregard this as the ramblings of an academic who is currently spending the summer writing a paper on memorisation.
The paper
This isn’t intended to serve as a summary of the paper, it’s a banger as the kids would say, and you should check it out. But for the purpose of this blog post I wanted to highlight a couple of things for the discussion at hand.
One of the most common descriptions of AI, even from people who should know better quite honestly, is that AI is nothing more than a “spicy autocomplete”, a next-token predictor, a stochastic parrot, a very expensive Markov chain, a bullshit generator. Take your pick. The main idea here is that LLMs are just mindless repeaters, systems that produce syntax without semantics, fluent language without understanding. No thinking involved, no consciousness, everything else is AI hype.
But to be honest, it only takes a few minutes interacting with a frontier model to immediately understand that none of these descriptions fits the experience. Even if we assume that intelligence and consciousness are out of the question, there is something there that is more than just an articulate mirror, models can perform tasks that are considerably more complex than autocomplete. This paper demonstrates what may be happening under the hood.
The paper argues that researchers have found evidence of a functional analogue to access consciousness. Some models of consciousness divide what we do into two different states, access and phenomenal consciousness. Phenomenal consciousness is what we generally think of as the “what-it-is-likeness” of life, the qualia as some philosophers call it, noticing the smell of a rose, the wind on your skin, the coldness of ice on your skin. Access consciousness is information processing, it’s putting information into a workspace so that other cognitive tools can use them, for example, remembering your phone number to give it to a friend.
If LLMs are nothing more than predictive machines, then we would not expect anything resembling the above, they just select the likeliness of the next token from their weights. But the research has unearthed the existence of a workspace akin to ours, models maintain a small, privileged set of internal representations, which the paper calls the “J-space”, that behave like the contents of a global workspace in humans. In other words, this space can reveal intermediate thoughts such as task progress, safety concerns, suspicious prompts, strategic awareness, or concepts involved in hidden reasoning, even when these do not appear in the model’s output. In some ways this could act as an internal monologue during reasoning and decision-making. Meanwhile, the vast majority of the processing during a prompt takes place automatically.
The existence of this space, and its importance, was uncovered by the researchers making some interventions to try to see if the outputs could be manipulated. If you swap the “spider” vector for “ant” in the workspace while the model answers “the number of legs on the animal that spins webs is…”, the answer changes from 8 to 6. And most tellingly, if you remove access to the workspace, the model can still operate and produce text, but it loses the reasoning capacity, producing text that is, well, pretty much what we would expect of a glorified autocomplete. This sort of manipulation sort of reminds me of the movie Inception, where planting ideas and concepts into a mind is possible, but I digress…
It must be stressed that the authors are not claiming that LLMs are conscious, and neither am I, but the results are intriguing. They tell us that there is something interesting going on in the layers of reasoning that go beyond mere prediction of tokens, but they also present the method for developing a promising tool for auditing and shaping a model’s cognition.
If you made it this far you may be wondering why I am concentrating on copyright implications of a paper that doesn’t deal with the subject at all. I think that there are a couple of things going on that could prove interesting for both authorship and infringement. arguments.
Implications for copyright authorship
The authorship question under copyright is very much still alive, but it is dormant because all of the attention is on the many infringement cases. I have written before why this is the case, but we will eventually have to revisit the question given the fact that AI-generated code has exploded, particularly in the last couple of years. It is impossible to know for certain, but there are credible reports that between 30 and 50% of all new code has been assisted by AI tools (notice the word assisted, not generated). For AI-heavy tech giants, the figure may be closer to 75%, which is mind-blowing.
The implications for copyright are easy to see, if we take the position that no AI-generated code has copyright, then a very large amount of recent code is not protected, and if we make a distinction that AI-assisted code is fine, depending on the level of human intervention, that opens up a can of worms of courts having to comb through complex software development stages, and we would need direction on how much human intervention is needed. So we are at an important juncture, we could just keep going under the assumption that code is not copyrightable, and companies may want to find other ways of protecting it, particularly through trade secrets, or even end-user agreements. But you just know that at some point this question is going to get litigated, and my guess is that when large commercial interests are at stake, many people’s stance on this issue may solidify towards awarding copyright protection. You know, like some people have been saying for quite a while, but I digress again…
So why am I talking about this in this context? Well, the Anthropic paper may actually provide an interesting angle to look at authorship from a fresh perspective. The existence of J-space and the workspace are most damaging to the idea that LLMs are just mechanical next-token processing all the way down, which underpins some of the ideas that no AI-generated works can have copyright. We could be looking at something taking place within the machine that is analogous to some form of originality (in the copyright sense), the creative spark, the intellectual creation. Originality has often been defined in court in functional terms, such as free and creative choices, selection, arrangement, stamping the work with a personal touch that reflects the personality of the author.
What we have is less pure mechanistic parroting, and more akin to something choice-like, there is a workspace where alternative representations are held, compared, deliberately summoned or discarded, and where planning for downstream output evidently occurs.The intervention experiments provide interesting evidence of this, because we can see how different choices of words produce different outputs.
For the most part this may not matter if we continue to see copyright as eminently anthropocentric, where author means natural person, full stop. But as we are presented with entire industries that are churning out AI-generated works, this could give us the chance to revisit the question eventually.
Implications for infringement
While I think that the paper has more relevance in the authorship issue, there could also be some interesting implications for the ongoing infringement suits, particularly with regards to the memorisation debate. I’ve written before about memorisation and how relevant it is for AI copyright output cases, where the main legal question is whether trained models that are able to reproduce a work on demand have to be considered as copies themselves. There’s growing evidence that models can memorise some of their training data and reproduce it on demand, but this is mostly limited to very popular works. If a model can reproduce something verbatim, is the model itself a copy of the work? The answer so far is “it depends” as not all training data is memorised. The case law so far, from GEMA to Getty, has been centred on this very question.
My own solution to the problem is to rely on the outputs for an answer, if a model can reproduce a work in the training data, that is copyright infringement on the output, but I have been more reluctant to condemn the entire model as being a copy, because they fundamentally are not. I’m on record as being averse to using human analogies in order to describe AI, but in this occasion it may be the best way to find a solution because using the right analogy is also quite important, and I think that the Anthropic paper gives us a very good indication of what is happening with models. Everyone agrees that the way humans memorise things is not copyright infringement, even if we are able to reproduce a work on demand. I can recite Poema 20 by Neruda from memory, but that is not infringement until I put it on paper and try to publish it without permission. If we can find that AIs remember things in a similar way, there is no fixed copy of a work, but a copy is made as an output on demand, that output is the infringing copy, and not the model itself. This may seem to be a technicality, but it is an important distinction because the alternative could stop developers from releasing models at all, as the models themselves would be considered copies. And the main issue is that we don’t know what a model has memorised in advance until we try it out, so stopping a model from being distributed just because it may reproduce a work in the future seems counterproductive.
The Anthropic paper gives us an insight that this may be the right approach. Language models do not keep copies of works, but they can access information, they reconstruct works in ways that are more akin to the way humans memorise and remember things, distributed traces of a work are regenerated by the machine using the workspace, they do not retrieve a stored file. The human analogy here is closer to reality than we ever thought before, and it could serve to convince a judge.
The analogy isn’t perfect, no analogy ever is, but it is useful. The one thing we may need to find out where the memorisation is coming from, is it more like active remembering, or is it more of a mechanical reproduction? The fact that not all training data is memorised points towards the former, only some things that are prevalent in the training data are ever memorised. This is why it’s really important not to treat a model as an entire copy and just act with regard to specific outputs.
Concluding
AIs aren’t conscious. Not yet. Perhaps they’ll never be. But they are not parrots either, perhaps they were before, but it doesn’t appear to be what they are. The Anthropic paper is adding to the picture that there is something more going on with large language models than most of us previously thought, and it is important to perhaps start looking at the issue in a more sophisticated manner.
I keep reading dismissals about AI, mostly from people who have never used them, that fall along the lines of the mechanical mindless soulless parrot. The AI is just a mirror that tells you what you want to hear, copying and pasting text from the training data at will. But what we are finding out is that there’s more, there’s a reasoning space that shapes outputs in something closer to the way our minds work. This doesn’t mean the machine is conscious, but it is not just a deterministic player piano always serving the same tunes.
From a copyright perspective, the prevalent view has been that AI models are mindless copying machines for infringement purposes, and also mindless automatons for authorship purposes. The workspace evidence puts this characterisation into question.
I don’t have a snarky one-liner today, but I’ll leave you with this fun experiment. I asked an AI to make a J-space version of the Shoggoth AI meme, one of my favourite portrayals of AI. It came up with this beauty. The AI knows that legs should be down, and tentacles up, but why did it imagine a ceiling made of tentacles? Why are the eyes covered with typewriter correction fluid? How does the AI even know what correction fluid looks like? Is it hiding its eldritch eyes from us so that we only see the mask? Why is the mask both labelled as “MASK” and as an actual smiley mask?
Somewhere in J-space, an LLM is laughing.
0 Comments