Governments around the world are considering how they can – and should – regulate the development and deployment of increasingly powerful and disruptive artificial intelligence (AI) technologies. Australia is no exception. On 1 June 2023, the Australian government announced the release of two papers intended to help ‘ensure the growth of artificial intelligence technologies (AI) in Australia is safe and responsible’. The first of these is the Rapid Response Report: Generative AI, which was commissioned by Australia’s National Science and Technology Council at the request of the Minister for Industry and Science, Ed Husic, back in February. The Rapid Response Report assesses potential risks and opportunities in relation to AI, and is intended to provide a scientific basis for discussions about the way forward. The second paper is the Safe and Responsible AI in Australia Discussion Paper which, according to the Minister’s media release, ‘canvasses existing regulatory and governance responses in Australia and overseas, identifies potential gaps and proposes several options to strengthen the framework governing the safe and responsible use of AI.’
The discussion paper seeks feedback on how Australia can address the potential risks of AI. It provides an overview of existing domestic and international AI governance and regulation, and identifies potential gaps and additional mechanisms – including regulations, standards, tools, frameworks, principles and business practices – to support the development and adoption of AI. It focuses on ensuring AI is used safely and responsibly, but does not consider all issues related to AI, such as the implications of AI on the labour market and skills, national security, or military specific AI uses.
Another key area that is expressly excluded from this consultation is intellectual property. That is, in my view, a serious shortcoming. It appears to presume that IP is somehow separable from the other issues covered by the discussion paper. This is a flawed presumption, particularly in relation to business practices. In the contemporary world, IP is at the heart of many business practices, and the laws and regulations that we make around IP can be the difference between a business practice that is viable, and one that is untenable. And not every business practice that might be enabled by IP laws is necessarily desirable or of net benefit to society. If we fail to consider the interplay between IP laws, business practices, and other forms of regulation, then we risk making mistakes that might prove very difficult to undo in the future.
This article is prompted by, but is not primarily about, the Australian consultation process (although I will return to that at the end). It is about how IP rights, and other forms of regulation, could operate to concentrate increasing levels of power in the hands of the few big tech companies – such as Microsoft (through its partnership with OpenAI), Google and Amazon – that have risen in recent years as the dominant players in AI and its enabling technologies. Based on recent developments, I believe that the stage is already being set for implementation of exactly the kinds of laws and regulations that would most benefit these companies, under the guise of protecting innovators, content creators, and the general public against the various threats said to be presented by AI.
A perfect storm is brewing. Onerous regulation around the development, training and deployment of AI systems could combine with IP-based restraints on the use of training data, and on AI outputs, to bake-in an advantage for the world’s richest and best-resourced companies. The storm is being fuelled by hype and fearmongering which, even though much of it may be well-intentioned, plays to the interests of big tech.
IP Rights and the Art of the Steal
It seems that whenever IP laws are expanded or ‘strengthened’, this is acclaimed by the relevant legislators and authorities as being beneficial for creators and innovators. The reality, however, is quite different from this utopian ideal. In their recent book, Chokepoint Capitalism, Rebecca Giblin and Cory Doctorow dissect the myriad ways in which a miniscule number of mega-corporations are slurping up a lion’s share of profits across creative industries and the internet by creating ‘chokepoints’ between creators and consumers. Every single one of us is contributing our hard-earned currency to Google (including YouTube), Facebook/Meta, Apple, Amazon, Spotify, the three major labels that control music recording and publishing, the five companies that control trade publishing of books, the one company (iHeartMedia) that dominates US radio broadcasting and is increasingly spreading its tentacles around the world, and the one company (Live Nation Entertainment) that is busily mopping up the live music industry. And nowhere near enough of our money – and probably far less than you think – is ending up in the bank accounts of the creative artists we believe that we are supporting by our patronage.
(Full disclosure: Rebecca’s office is just around the corner from my desk at the Melbourne Law School. But this is not why I am telling you that Chokepoint Capitalism is an excellent and eye-opening book, and that you should buy it and read it. In case you are wondering, I paid for my copy myself – ironically, in the Kindle version!)
Extending the duration, scope or strength of copyright laws has never – at least in the modern era – been of any benefit to the overwhelming majority of creators of valuable copyright works. As Giblin and Doctorow pithily put it, ‘giving more copyright to creators who are struggling against powerful buyers is like giving more lunch money to your bullied kid.’ The bullies still steal your kid’s lunch money, but now it just makes those bullies richer! That is exactly what happened when it was established that sampling music was an infringement of copyright. You might think that this would have been great news for the artists whose work was being sampled when, in reality, it was mostly great news for the powerful labels which immediately claimed, or acquired, ownership of the relevant rights and started vacuuming up eye-watering licence fees, while using creative accountancy and exploitative contract terms to keep most of the proceeds for themselves.
In AI, Control of Content is Power
We are already seeing signs of how chokepoints may emerge in the field of AI technology. For example, in relation to the stock illustrations industry, there is some evidence that the business of licensing images is dominated by a relatively small number of providers: principally Shutterstock; Getty Images (including iStock); and Adobe Stock. Two of these companies are already engaging with generative AI systems, though in vastly different ways.
On the one hand, Shutterstock has ‘partnered’ with OpenAI in the training and use of OpenAI’s DALL-E text-to-image generative AI system. Under this arrangement, Shutterstock provides images (and associated descriptive text) from its database for use by OpenAI in training the DALL-E models. Meanwhile, Shutterstock users gain access to DALL-E to generate images based on text prompts. Although this may result in a loss of revenue to artists if users choose custom AI-generated images rather than the artists’ work, Shutterstock has promised to compensate artists for the use of their work in training the AI models (presumably from licence fees paid to Shutterstock by OpenAI).
On the other hand, Getty Images has sued Stability AI for (among other things) copyright infringement. Stability AI describes itself as ‘the world’s leading open source generative AI company’ with a goal to ‘maximize the accessibility of modern AI to inspire global creativity and innovation’. The first version of its text-to-image generative model, Stable Diffusion, was released in August 2022. Getty Images alleges that Stability AI has engaged in ‘brazen infringement of Getty Images’ intellectual property on a staggering scale’ by copying more than 12 million images from its database without permission or compensation.
The contrast between the business models of OpenAI and Stability AI could not be more stark. OpenAI is – despite its name – anything but open, and has not released DALL-E source code or trained models to the public. The contents of its training datasets are unknown, as are the details of its deal with Shutterstock. This has all the hallmarks of a chokepoint in the making. On one side we have artists seeking to make a living from their images, while on the other we have the prospective consumers of generative AI services. And, in the middle, OpenAI – the gatekeeper of its proprietary generative AI services – in partnership with Shutterstock – controller of one of the largest troves of stock imagery on the planet – with the potential opportunity to lock in both sides of the market, if only they can become sufficiently dominant.
Meanwhile, Stability AI (which also offers paid generative AI services) has opened up its source code and models, which have clearly been trained using text and images scraped from the internet (including, allegedly, Getty Images’ website) with apparent disregard for any copyright issues that this may raise. The release of trained models assists other companies to enter the market, even if they do not have the financial or computing resources to train models themselves (it has been reported that Stability AI’s operations and cloud expenditures exceeded US$50 million, prior to its raising US$101 million in venture funding in October 2022). But if Getty Images is successful in establishing that its copyrights were infringed in training the models, it is likely that they will have to be withdrawn, stranding anybody who had been trying to build a business based on Stability AI’s open models. This would hand the lion’s share of power back to the copyright aggregators, and the large, cashed-up, AI companies that are best–placed to offer them attractive deals.
I note, in passing, that class actions for copyright infringement have also been launched in the US by software developers against GitHub, Microsoft and OpenAI, and by artists against Stability AI, DeviantArt and Midjourney. The plaintiffs in these cases should perhaps be careful what they wish for – while they might be seeking control and compensation for creators, any success they may achieve is more likely to play into the hands of copyright aggregators.
Blessed are the Fearmongers, for They Shall Enkindle the FUD!
There is little that encourages hasty and ill-considered action much more effectively than fear, uncertainty and doubt (FUD). And there has been no shortage of FUD generated around recent developments in AI. Of course, much of this has been the result of an understandable level of general ignorance about these highly complex technologies, and how they work. But some of the fear is being sparked by people who, you might have thought, should know better, and whose credentials and reputations imbue their comments with enhanced authority and influence.
You may have heard a recent story about a speech made by Colonel Tucker ‘Cinco’ Hamilton, US Air Force Chief of AI Test and Operations, at the Future Combat Air & Space Capabilities Summit, in which he told an anecdote about a simulation in which an AI controlled drone ‘killed’ its (simulated) ‘operator’ when they interfered with its prime objective (‘kill the enemy’) by trying to order it not to kill the enemy. The only problem with this story is that it did not actually happen – the USAF has subsequently denied it, and Colonel Hamilton has since said that he ‘mis-spoke’, and that he was merely presenting a hypothetical ‘thought experiment’ that had originated outside the military. The problem with this problem with the story, however, is that some media outlets have chosen to play on public distrust to cast doubt on the denials – for example, The Register has gone with ‘who knows?’, while news.com.au, somewhat predictably, has kept its original alarmist headline (‘AI drone kills its operator in simulation’), and downplayed the denials as ‘damage control’, squarely placing the blame on the USAF as opposed to the journalism fail that is actually is.
Unfortunately, at this point, journalists and the public are already primed to believe these kinds of stories, rather than to treat them with the scepticism they deserve.
When we are told that over 1,000 people, including Elon Musk and many ‘AI experts’, have signed an open letter published by the Future of Life Institute including concerns that ‘nonhuman minds … might eventually outnumber, outsmart, obsolete and replace us’ and that ‘we risk loss of control of our civilization’, it is unsurprising that many of us might conclude that these are real and present threats.
When we are told about a Stanford University study finding that more than a third of researchers believe that AI could lead to a ‘nuclear-level catastrophe’, or we are told that one of the ‘godfathers of AI’, Geoffrey Hinton, has resigned from his role at Google in order to speak freely about the threats posed by AI, comparing them to ‘nuclear weapons’, it is not unreasonable to suppose that AI poses a threat that is comparable to nuclear weapons. After all, if Geoffrey Hinton doesn’t know his stuff on this topic, then who does?!
When we are told about Microsoft engineers – who worked with GPT-4 for months ahead of its public release while integrating it into the Bing search engine – concluding that GPT-4 exhibits early signs of ‘artificial general intelligence’ (AGI), then why would we not believe that we are genuinely now on the verge of developing AGI? Although, to be fair, there were a number of media outlets – including Vice, the New York Times, and Forbes, among others – that sought alternative views and treated these ‘eyebrow-raising’ claims with the scepticism they rightly deserve.
It is hard to fathom just why so many experienced and knowledgeable people have taken on the role of harbingers of our impending doom at the virtual hands of super-intelligent machines. They are certainly not all shills for the tech giants, and most of them appear to be entirely sincere. Perhaps there is something to the explanation offered by Cambrian AI analyst Alberto Romero, who posits compellingly that there is an element of quasi-religious group-think involved: ‘[t]hey're of an unusual kind with an uncommon faith … built on superficial but hard-to-debunk logic by which they've convinced themselves to be modern-day foretellers of a future that, at its best is singularly glorious and, at its worst, our closing chapter.’
Whatever the motivations, however, the fearmongering has been effective, and is playing nicely into the hands of the big tech companies that have, to date, invested the most – and naturally hope to gain the most – from AI technology.
Beware of Geeks Bearing Grifts
With all of the FUD now surrounding AI, governments are under increasing pressure to (be seen to) take action to protect concerned citizens against the potential harms that AI may cause. On 16 May 2023, OpenAI’s CEO Sam Altman, appearing before a US Senate Judiciary subcommittee, suggested that ‘the US government might consider a combination of licensing and testing requirements for development and release of AI models above a threshold of capabilities.’ That seems strange. Since when have US technology companies been openly in favour of being subjected to additional government regulation? Unless, of course, it benefits them. Indeed, whenever you see already powerful people or organisations lobbying for more regulation, it’s a generally a safe bet that they are rent-seeking. To see how that might apply here, we need to look a bit more closely at the state of the art in large language models (LLMs), like GPT-3 and GPT-4, and the direction in which the technology seems to be developing.
So far, the development of ever more impressive LLMs has mostly relied upon the principle that ‘bigger is better’. In their 2020 paper introducing GPT-3, Language Models are Few-Shot Learners, OpenAI researchers revealed that the model underpinning the initial free version of ChatGPT has 175 billion parameters, and was trained on a data set comprising 300 billion tokens (in this context, a ‘token’ is typically a word, part of a word, or a punctuation symbol, in the text used for training). Estimates of the actual training requirements vary, but it is certain that a hardware platform with thousands of high-end processing units (worth on the order of $10,000 apiece) is needed, with training times measured in weeks, and total compute costs measured in the millions of dollars – all, in OpenAI’s case, supplied and funded by Microsoft.
Very little is known for certain about GPT-4. As stated in OpenAI’s GPT-4 Technical Report, ‘[g]iven both the competitive landscape and the safety implications of large-scale models like GPT-4, this report contains no further details about the architecture (including model size), hardware, training compute, dataset construction, training method, or similar.’ But in a talk presented at MIT on 22 March 2023, Microsoft’s Sebastien Bubeck (who headed the team responsible for integrating GPT-4 into Bing) implied that the model size is on the order of a trillion parameters, i.e. around six times the size of GPT-3. With advances in hardware and training techniques, it is likely that the number of processing units and the training time would not have scaled up by this amount. Nonetheless, the compute costs for anybody else to train such a model would almost certainly be well in excess of ten million dollars. And that is not including prior development and testing – you do not successfully build and train one of these huge models at your first attempt!
While running a pre-trained model requires significantly fewer resources than training the model, the hardware requirements to run GPT-4 are also significant. A trillion parameter model (probably) requires two trillion bytes, i.e. two terabytes, of memory – not storage/drive space (though that also), but actual memory – when in use, plus additional memory to store inputs, outputs and other working data. If executed using the same high-end processing units employed in training (most likely NVIDIA A100 80GB GPUs), the total cost of components to build a suitable hardware platform to serve a single instance of the model would be on the order of a million dollars, and the resulting peak power consumption around 20kW (making decent air conditioning also an essential requirement). And to serve thousands of users simultaneously with fast response times, it is likely that this hardware setup would need to be replicated dozens, if not hundreds, of times. It is no wonder, then, that OpenAI is making users pay for premium access to ChatGPT, and for use of the API, and that non-paying users are made to endure downgraded performance!
The ultimate point, of course, is that very few end-users are going to be in a position to host their very own instance of a model such as GPT-4. Rather, they are going to be reliant on cloud providers to host the models, and to offer various services, at different price points, depending on the user’s requirements in terms of criteria such as performance, data privacy, and security. And the biggest cloud computing providers are, of course, Microsoft (via its Azure platform), Amazon (via Amazon Web Services, AWS), and Google (via Google Cloud Platform, GCP). That would give these companies a virtual stranglehold on the provision of these types of AI services.
But what if it were possible to build smaller, non-proprietary models capable of achieving comparable performance to monster models like GPT-4? As it happens, that is exactly the direction in which research on LLMs outside of OpenAI is going. In a paper published in February 2023, LLaMA: Open and Efficient Foundation Language Models, Meta (i.e. Facebook) researchers reported training models with as few as seven billion parameters to achieve performance comparable with the biggest LLMs. What is more, they did it using ‘publicly available datasets exclusively, without resorting to proprietary and inaccessible datasets’ (as OpenAI has done with its GPT models). And they claim that a version of the LLaMA model with 13 billion parameters outperforms the 175 billion parameter GPT-3 model on most benchmarks.
The training cost for these smaller models is still very high, at around US$5 million (assuming 21 days training on 2,048 A100 GPUs, as reported in the LLaMA paper). But in a massive ‘FU’ to OpenAI, Meta open sourced its training code, and released all of its trained models to the research community. While it intended to do this in a controlled manner, the trained model weights were quickly leaked on the internet, making them available to everybody.
The LLaMA models come with licensing conditions that restrict commercial use. However, there are already projects, such as RedPajama, underway to replicate Meta’s results and to create models that are unencumbered by these restrictions. At the same time, there are a number of efforts ongoing to enable these smaller models to be run on (relatively) cheap commodity and consumer-grade hardware. Right now you could, if you wished, run your very own LLM-based chatbot at home, on a PC worth under $3000. It would not be as good as ChatGPT (try out Intel’s Q8-Chat on the Hugging Face platform to get a feel for what is possible at the low end), but it would be all yours, completely private and secure, and much better than anything you had seen prior to November 2022!
I predict that these community-based research projects will win out, with the result that ‘Big AI’ (i.e. the likes of OpenAI/Microsoft and Google) will fail to establish the near-monopoly on AI-based services that they are clearly hoping to obtain. (In the interests of balance, however, I note that not everybody shares my optimism.) There is only one thing, in my view, that can stop this ‘democratisation’ of LLMs from happening: regulation. If the US government actually did introduce a licensing and testing regime for ‘AI models above a threshold of capabilities’ (whatever that means) it would pose no real obstacle to companies like OpenAI – they have already spent years building a reputation as ‘responsible custodians’ of their technology, and they clearly have the economic and political clout to cut through any regulatory red tape. Smaller companies, research groups, and community projects, on the other hand, would doubtless face significant hurdles in proving their credentials and bona fides in order to obtain regulatory approval to develop and deploy their own AI systems.
Worryingly, however, the FUD-mill seems to working for Big AI. On 6 June 2023, two members of the US Senate Subcommittee on Privacy, Technology, & the Law sent a letter [PDF, 187kB] to Meta CEO Mark Zuckerberg. The letter implies, unsubtly, that the ‘leak’ of the LLaMA model weights was a foreseeable and perhaps intended outcome, and requests ‘information on how [Meta] assessed the risk of releasing LLaMA, what steps were taken to prevent the abuse of the model, and how [Meta is] updating [its] policies and practices based on [the model’s] unrestrained availability.’ The implication of the the many questions put to Zuckerberg in the Senators’ letter are clear – if corporations cannot be trusted to act responsibly in relation to powerful AI models, then the government will need to step in to regulate their behaviour.
Who Will Benefit Most From AI-Related IP Rights?
Returning to the discussion around IP, much the same can be said of any proposals to ‘strengthen’ IP rights. This might be purported to protect the interests of creators and innovators, but in reality may just hand more power to the large, oligopolistic tech companies that will, once again, seek to acquire and accumulate all of the rights that they can, as cheaply as they can, and use them to extract rents from the rest of us every time we want to use the technology that they control.
If a licence is required to use copyright materials to train AI models, then the only ones who will be able to train AI models are those that have the money and clout to negotiate, litigate, and/or pay for licences to massive volumes of content. History – and Chokepoint Capitalism – teaches us that creators will not be the main beneficiaries of these licenses. The most valuable music rights are already controlled by a handful of recording labels; the most valuable book rights are already controlled by a handful of publishers; the most valuable news, television and movie content is already controlled by a handful of media companies; the most useful and well-structured image databases are controlled by a handful of stock image companies; and the bulk of the remaining content out on the internet is created by individuals and businesses that have, implicitly if not expressly, given it up to the world to do with as it pleases.
So, first and foremost, protecting the use of copyright materials to train AI models simply adds another source of revenue to companies already exploiting chokepoint business models, without significantly benefiting the creators of those materials. In other words, it just provides more lunch money to be stolen from the bullied kids!
Much the same goes for recognising AI as an author or inventor, and granting copyrights and patents for AI-generated creations. A naïve view – such as that promoted by Ryan Abbott and the Artificial Inventor Project – would have it that this would ‘incentivize the creation of intellectual property by encouraging the development of creative computers’, and recognise the potential for computers to ‘overtake human inventors as the primary source of new discoveries.’ But, again, experience tells us that the most valuable IP rights do not generally end up being held by their human creators, but by corporations that employ the human creators or that subsequently acquire the rights. I have already discussed examples of copyright aggregators, while in the patent sphere there are patent assertion entities (PAEs) – or sometimes, pejoratively, patent ‘trolls’ – that accumulate rights in order to extract licence fees from alleged infringers. So if AI can invent, and we decide to grant patent rights for AI-generated inventions, might we not just be creating a new business model for ‘patent generation entities’?
Abbott might be right. Granting copyrights and patents for machine-generated creations might incentivise the development of creative machines. But that is not the most important point. What matters is who owns and operates those creative computers, and who benefits from their creations.
Conclusion – the Need for an Integrated Approach
Just to be clear, in case it seems like I am advocating completely hands-off, regulation-free, AI anarchy – nothing could be further from the truth. We absolutely need to regulate AI. But what most needs to be regulated is how it is used, as opposed to how it is developed, trained and distributed. There is no question that AI poses real risks to human society and well-being, such as job losses, privacy violations, deepfakes, algorithmic bias, and weapons automation. These concerns need to be addressed by ethical and responsible use of AI, and we will undoubtedly require legislation and regulation to ensure that corporations, researchers, and individuals operate within agreed boundaries – and that there are consequences if they do not.
But it is futile to try to stuff the genie back in the bottle by attempting to control how, and by whom, AI is developed and trained. This simply cannot work. While most small companies, research institutions, and individuals lack the money and resources to build and train larger and more sophisticated AI models, there is no shortage of bad actors – including criminal organisations and rogue nation states – with the capacity to do so. If we criminalise training AI models, then only criminals will train AI models. Everybody else will be left out, and will be at the mercy of the relatively small number of dominant tech companies that are granted authorisation and/or are able to secure access to the large quantities of training data needed.
With the release of open source code and pretrained models – whether intentionally or otherwise – we are seeing the first signs of the democratisation of sophisticated AI. And yet two democratically-elected US Senators want Mark Zuckerberg to explain what steps Meta has taken ‘to track the distribution, repurposing, and end use of LLaMA … including any DMCA takedown notices, cease and desist letters, or similar efforts.’ And if existing laws are inadequate, then I am sure the Senators can come up with new ones. While I would not want to put ideas into their heads, what about making it compulsory to add tracking and kill code to all distributed models, and to apply encryption and digital rights management (DRM), so that even attempting to decrypt and reverse engineer the models would be a potentially criminal act? Would that stop criminals, or enemy agents? No. But it would stop just about everybody else.
So if this is the future we all want, then by all means we should regulate the development, training and distribution of AI models, and fine-tune our IP laws to lock down rights on training inputs, and on the outputs of deployed models. In the recent past, that kind of approach has worked out well for big tech, but not so well for the community more broadly.
But turning back to the Australian consultation, the government says that the exclusion of IP in the current discussion paper is OK, because it is addressing IP in other initiatives, such as the Ministerial Roundtable on Copyright, and the AI Working Group of IP Australia’s Intellectual Property Policy Group. While that is all very well, neither of these groups includes a broad representation of interested stakeholders in AI technology. Participants in the Ministerial Roundtable include most of the usual big players in the copyright industries, such as the public and private broadcasters and news organisations, representative bodies of the film, music, performing arts and publishing industries, and the Copyright Agency. Information on IP Australia’s Intellectual Property Policy Group is harder to come by. All I know about it, via professional colleagues, is that it includes representatives from the IP professional associations FICPI Australia and the Institute of Patent and Trade Mark Attorneys of Australia (IPTA).
You will therefore have to excuse my scepticism that these IP initiatives will give adequate consideration to the interplay between IP laws and other regulatory actions and business practices. Reviewing the role of IP laws in isolation is, in my view, a mistake. Intellectual property rights are implicated right across the AI pipeline, from the selection and aggregation of training data through to the outputs generated from the final trained model. They should therefore also be integrated into any approach to regulation of AI, to avoid unforeseen consequences and interactions between IP laws and other aspects of the regulatory framework.
By taking a piecemeal approach, we risk getting the regulation of AI wrong. For once, let’s try not to do that.
0 comments:
Post a Comment