AI: Is it one intellectual property too hot for the current provisions to handle?
The Centre for AI Safety released a statement last week that was signed by more than 350 of the world’s top experts in the field of Artificial Intelligence (AI). Surprisingly, the statement stressed that AI is a grave threat to human freedom and societies; in fact, the experts urged that the threat classification from AI be ratcheted up to that posed by pandemics and nuclear war.
The immediate cause for alarm, the recent release of the new model of GPT-4 chatbot, however, may seem rather innocuous. The AI community seems oddly exercised over their own best creations – ‘generative’ AI systems like GPT and Dall-E that are capable of generating text, audio, video and images that sound authentically ‘human’. The new model’s capacity to mimic human communication convincingly may be eerie but any threats posed by it can, at first glance, seem very far from existential.
Certainly, it may be difficult to understand the reasons behind the dire warning issued by Yoshua Bengio, one of the ‘godfathers’ of AI, who said that unregulated growth of certain types of AI is a ‘danger to political systems, to democracy, to the very nature of truth’ (Quoted in Henry Zeffman, Mark Sellman and Alex Faber, ‘We need “guardrails” to regulate AI, Rishi Sunak says at G7 summit’, The Times, 18 May 2023). Yet, the unease that has come to the surface now has been bubbling underneath for some time.
Human Peasantry and Digital Masters?
The speed of AI evolution has picked up enormously in the last decade. While AI, in general, can be an effective weapon for delivering improved healthcare solutions and ultimately solving problems like climate change, ‘generative’ or ‘general purpose’ AI systems are based in ‘machine-learning’ principles. They aim to ‘learn’ from the ‘big data’ that is fed to them, exactly like humans learn from experiencing data and communication in context.
These systems can be repurposed to perform a huge variety of different functions, limited only by the imagination and coding capacity of the humans in charge. Beyond a certain point, however, the systems’ algorithms mean that they work to a logic of their own, which may not always be clear to their creators either.
As a result, these systems can be used in widely varying ways, which is profitable for the companies that create them, however, with their logic often hidden, they may become agents for spreading disinformation even without a bad actor deliberately programming them to do so. The threat of bad actors capturing and misusing such systems is also real but that is beyond the purview of this post.
Bengio and others are worried that, backed by society’s respect for innovation and technology, the big tech companies are in an AI ‘race’ where the social and existential dangers to humanity are not accorded enough attention.
AI algorithms are complex and the general public currently is not informed enough to understand them even if they are shared. Instances have been reported where the human-like communication from AI has drawn trust from human actors even when the trust was not warranted. It is easy to see how a combination of an opaque systemic logic and lack of information amongst relevant human stakeholders can make an AI system an unwitting agent for spreading disinformation that can cause grave danger and harm to society.
Using data that is not always scientifically verifiable (see https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3896852, pp. 107-08), these systems may be put in positions where they control human thoughts and behaviour, effectively creating a new social division where human actors are manipulated by AI they trust but do not understand.
Regulation of AI Systems
The danger has not been lost on the AI industry. In the past decade concerted efforts have been made to stress the importance ‘ethical AI’. A basic and clear understanding exists that risks from artificial intelligence can come at one of the two stages: design and deployment.
A core list of ethical principles, similar to Australia’s Artificial Intelligence Ethics Framework, exists in most countries undertaking AI research in a considerable way to guide both stages. Like the Australian framework, these broadly include ‘generate net benefits, do no harm, regulatory and legal compliance, privacy protection, fairness, transparency and explainability, contestability and accountability’.
However, ‘generative’ AI systems such as the GPT-4 (it should be stressed that not all AI is ‘generative’ as most systems are programmed to function within limited logical and action parameters defined by their creators to perform only a fixed number of actions) in their current iteration present new challenges.
They are no longer merely their creator’s ‘intellectual property’ to be exploited for economic gains. They are entities capable of creating their own intellectual property – text, sounds, images, videos – which brings them nearer our own ranks as creators of ‘meaning’ in this world.
The European Union has been the first to recognise this special status of AI, separate from other digital products. The world’s first legal framework to regulate AI specifically, the AI Act, is currently under discussion and is likely soon to be passed into EU law.
Public consultation on the draft legislation has largely approved of the risk-based approach of the Draft AI Act that divides AI products into different risk categories and prescribes varying levels of regulatory regimes based on the risk posed to society by each one. The most dangerous applications, covering ‘biometric surveillance, emotion recognition, predictive policing AI systems‘ have been prohibited outright.
The U.K. and the United States have so far relied on broad principles and guidelines that set out ethical AI frameworks to be followed voluntarily. However, the U.K.’s Prime Minister Rishi Sunak recently spoke of the need for ‘guardrails’ to contain any threat to humanity from AI. In the United States the Supreme Court recently refused to award patents to AI who had generated inventions of their own. The court ruled that patents could only be granted to humans. The US Copyright Office has also refused to recognise art generated by AI.
Clearly, the authorities in these societies feel the need to assert the separation between human and artificial intelligence and to further impose the control of the human upon the artificial in line with Bengio’s warning about AI’s threat to truth, democracy and political systems.
Truth, Democracy and Moral Responsibility
Truth
Complex and highly-developed communication and creativity are, to the best of our knowledge, uniquely human qualities. In addition, human beings communicate not only ‘facts’ but also and, more importantly, ‘meanings’. For example, we recognise the leaves, roots, trunk and bark of a tree as ‘facts’, but also that together they make a ‘tree’, which may be ‘big’, ‘beautiful’, ‘inconvenient’, ‘a hurdle’, or many other things, depending on the speaker, the context, the speaker’s beliefs, wishes, experience and aims. To quote Luigi Romeo, ‘meanings change … depending on whether they are used by a philosopher, a literary critic, or an archaeologist‘.
Therefore, ‘meaning’ of facts in human communication is something more than its constitutive parts and can change, depending on context. Further, a fundamental, though implied, assumption in all communication we undertake is that ‘A person making an assertion … aims to say something true’.
Crucially, the truth of communication often relates not only to verifiable ‘facts’ but also to those underlying signals that ‘are the objects of mental states like belief, the bearers of truth and falsity as well as modal properties like necessity and possibility and epistemic properties like a prioricity and posterioricity‘, to use the language of theories of meaning. Similarity of beliefs and mental states hold groups and societies together and deep dissimilarities can tear them apart.
Generative AI, being used now in a variety of different contexts, may mimic this communication and sound ‘true’ because their communication follows a semantic logic that holds true by rules of logic we understand. Fed on ‘big data’ generated by humans this communication may use ‘correct’ facts and data but still convey an overall ‘meaning’ that may be dangerously untrue. For a simplistic example, one may think of an AI system used in making court decisions that accepts as valid evidence a picture of a ‘playful’ gesture as evidence of a genuinely harmful one and suggests a harsh sentence accordingly.
Further, there is the question of trust. Zhu et al. note that ‘humans and societies perceive trust in AI in intricate ways, which does not necessarily closely match the trustworthiness of a particular AI system’. Human actions are often enabled or constrained by things we believe to be true or false. For example, if people in a given area believed, based on information from a source they trusted, that the local polling station was closed on the day of the election, the majority may simply not go out to vote, thus affecting the election outcome. This is a simplistic scenario but one may build up to other, more complex, but equally realistic possibilities, involving genuine widespread social dysfunction, violence and harm.
Democracy
It follows from this that the greater is the human agency afforded to populations in a given system, the greater is the danger posed by the possible dissemination of untrue facts and beliefs. This is perhaps the reason why, despite some dissenting voices too, we are mostly hearing serious concern about generative AI being expressed by scientists and thinkers in democratic societies.
A democratic society affords its members considerable freedom to act, or affords them the exercise of Free Will. However, there is ample literature in philosophy and social science research asserting that our ‘Free Will’ itself is a product of ‘our upbringing and environment, our education and friendship circles, but also …[of] the social and political conditions we are born into‘.
Under the circumstances, actors that influence the socio-political environment are expected to act with a certain sense of moral responsibility. The greater the influence they exert upon the environment, the more stringently are the actors expected to abide by moral rules acceptable to the wider society; rules and codes that societies often enshrine in their laws and regulations.
Moral Responsibility
This issue, however, is complicated by the emergence of generative AI. When the logic followed by an AI system is not transparent, it is ‘autonomous’ of human intervention and acting of its own Free Will, but does this mean the AI system is morally responsible for the effects of its actions/products?
Philosophers have generally argued that an AI system cannot be a moral agent because ‘they cannot suffer and thus cannot be punished’. Who, then, is to be held responsible if an AI system disseminates disinformation? There are no clear answers at the moment.
A related and equally pertinent point, however, is the serious danger posed by even a human being following a cold rationality, as highlighted by Hannah Arendt about Adolf Eichmann. ‘Arendt believed that [Adolf] Eichmann’s being expressed a “banality” of evil, that is, an evil emanating not from a will (to do evil) but from a lack of thinking. … Eichmann followed the law but did so “blindly,” without ethical reflection [bold mine].’
Despite their fast speeds, and the capacity to learn, mimicking human reactions, AI systems do not have this extra dimension to human intelligence, which is the capacity to ‘reflect’ and which Luigi Romeo distinguished as ‘human intellect, not intelligence alone‘.
Further, entities that have, or may have, pure logic, embodied agency but a lack of ethics, are likely to be more akin to ‘nature’ that Nietzsche warned against: ‘… a being … boundlessly extravagant, boundlessly indifferent, without purpose or consideration, without pity or justice, at once fruitful and barren and uncertain … INDIFFERENCE as a power’. This nature was red in tooth and claw as all such entities are capable of becoming. To quote Nietzsche again, ‘Is not living valuing, preferring, being unjust, being limited, endeavouring to be different?’ It is a reasonable fear that our current AI race may be taking us on the path to populating our society with such beings.
Arguing for ‘Education-as-Defence’
Most AI experts are keen to stress that the ordinary members of society should find AI applications as accessible as possible. The U.K.’s guidance for developing responsible AI in the public sector exhorts experts to ‘Convey and communicate … algorithmically assisted decisions to the individuals affected by them in plain language’.
Scholars and organisations responding to the Draft AI Act of EU criticised the lack of popular or bottom-up control with many individuals and organisations echoing concerns expressed by Dr Michael Veale and Professor Fredrik Zuiderveen Borgesius at the ‘absence of affected individuals and communities‘ from the provisions of the Act. Moreover, even though the proposed Act makes guarding Fundamental Rights a central concern, there is a whole group of systems (not listed in Annex III) that is not covered by these provisions.
Even with these points addressed, due to the complex nature of AI use (for example, an AI system may provide the categorisation or background information on which a harmful decision is made downstream) and the ‘out-of-control’ race between different AI developers (Professor Stuart Russell quoted in Blakely, Rhys, ‘AI “Could be like an alien invasion” says British professor’, The Times, 13 May 2023), who are rushing to place competing products in the market as soon as they possibly can, there is need for careful and ongoing societal oversight of these powerful entities that are no longer mere ‘intellectual property’.
Ultimately, democracy’s best weapon is ‘education-as-defence’ and it must be an urgent task for all concerned to ensure the transparency of these systems and the ability amongst all groups of users to recognise, identify and report on possible banal issues and untruths, generated by the sterile logic of AI systems, that might end up costing us our very humanity.
Image courtesy Andy Kelly on Unsplash