From the recently published blog to be found here
I am writing this on a plane flying away from Malta, where I just spoke about SingularityNET at the Malta Blockchain Summit. It was my first time on Malta, and after the event, I took the afternoon to explore some of the elegant, quaint, ancient neighborhoods of the island. Walking through medieval alleyways by the rocky coast, I felt an ironic contrast between my elegant surroundings and the main reason I had decided to allocate a couple days from my insanely busy schedule to this Malta event: not just the conference itself, but also the opportunity to meet with the top levels of the Malta government to discuss the enablement of Maltese citizenship for AIs, robots and automated corporations. The folks who had built the stone walls lining the narrow Maltese roads, still standing strong centuries later, had probably not foreseen their blue-wave-lapped island becoming a nexus of thinking at the intersection of general intelligence theory, cryptography, distributed systems, and advanced legal theory.
The Hanson Robot Sophia, with whose development I’ve been intimately involved via my role as Chief Scientist of Hanson Robotics, was granted citizenship of Saudi Arabia this year. This was an exciting landmark event, however, its significance is muddled a bit by the fact that Saudi Arabia is not governed by rule of law in the modern sense. In a nation governed by rule of law, citizenship has a clearly defined meaning with rights and responsibilities relatively straightforwardly derivable from written legal documents using modern analytical logic (admittedly with some measure of quasi-subjective interpretation via case law). Saudi Arabian citizenship also has a real meaning, but it’s a different sort of meaning — derivable from various historical Islamic writings (the Quran, the hadiths, etc.) based on deep contextual interpretation by modern and historical Islamic figures. This is a species of legal interpretation that is understood rather poorly by myself and most of my colleagues in Hanson Robotics and SingularityNET, and one that is less easily comprehensible by current AIs.
As a related aside, since I get asked the question a lot, I want to clarify that: The initiative to grant Sophia citizenship was taken by the Saudi government, not by anyone at Hanson Robotics. Furthermore it was not associated with any flow of finances from Saudi sources to Hanson Robotics or associated entities (aside from a speaking free for Sophia to speak in Saudi Arabia at the event where the citizenship was granted, which was on the same level as the fees she has been paid for speaking countless other places).
I’m aware that affiliation with Saudi Arabia in any sense has become controversial in recent weeks due to the apparent murder of Jamal Khashoggi. I am certainly not in favor of murder of journalists or anybody else, by governments or anybody else. However, speaking for myself personally, independently of any of the companies I’m involved with or any broader political issues, I consider the granting of citizenship to Sophia a genuinely forward-thinking and positive act on the part of the Saudi government. Of course it was in part an act of public relations. However there are a lot of possible acts of public relations, and the choice of this particular one was a demonstration of real futuristic vision. The mix of futuristic ambition and insight with dramatically non-modern legislature and governance that one finds in today’s Arab world is fascinating to me, and also at times deeply disorienting and disturbing to me, but that’s a topic for another time.
Anyhow, ever since the granting of Saudi citizenship to the Hanson robot Sophia last year, David Hanson and I have been especially interested to find a democratic nation governed by a modern-style legal system with an interest in AI citizenship. I.e., now that Saudi Arabia has opened the door, let’s take the next step and figure out how to make robots and other AIs citizens in the context of modern legal codes!
I have talked to numerous people involved with various governments about this, generally getting reactions of enthusiastic interest and zero practical activity (governments tend to be good at that!). Malta, on the other hand, has appointed an AI Task Force led by the Junior Minister for Financial Services, Digital Economy and Innovation Silvio Schembri, and posited as one of the initial goals of the AI Task Force to create a definite roadmap toward citizenship for AIs. So over the next year, I and my SingularityNET colleagues will be working closely with the Malta AI Task Force to come to a common understanding of what AI citizenship should mean, and how one might evaluate whether an AI is competent to be considered a citizen of Malta.
The Malta AI Task Force, with the collaboration of myself, my SingularityNET colleagues and others, will form an advisory committee on AI citizenship, comprising individuals with expertise in AI, law, international relations and ethics, and related areas. Via in-person meetings and long-distance communication the Malta-based AI Task Force and the globally-distributed advisory committee will then work toward a common understanding on how to make sense of the notion of AI citizenship in a practical sense.
While I can’t pre-figure exactly what will be the outcome of this process, I can share here a few elements of my own thinking that I will bring to the discussions.
First of all, while “robot citizenship” is good for stimulating the popular imagination, of course, the matter isn’t fundamentally about physical embodiments. AI programs not anchored to specific robot bodies may be equally deserving of citizenship. And in fact, as robots like Sophia become more and more intelligent, more and more of their underlying AI processing comes to be done in the compute cloud (in Sophia’s case, using the SingularityNET platform among other tools) — so that the same “robot mind” can be used to operate multiple different robot bodies.
The controversial notion of corporations as legal persons also plays a role here. What if one has a Decentralized Autonomous Organization, an automated company defined by smart contracts and conducting business (say, on the Internet) in a fully automated way — under what conditions should this DAO be allowed to register itself as a legal corporation, without needing any human being in the loop to provide identification and sign forms?
What is really at issue here is citizenship for any artificially intelligent agent, be it a chatbot, a robot control system, a DAO, or something else entirely different. In our preliminary discussions, it became clear to me that this was already the perspective the Malta AI Task force was taking. Given Malta’s role as a major international center for blockchain enterprise, issues regarding the legal status of DAOs were relatively prominent in the Task Force members’ minds.
There are also various levels of citizenship that may be considered. Estonia has introduced a notion of “e-citizenship”, which does not require physical residency in Estonia. It might make sense to consider Maltese e-citizenship for AIs, as a preliminary step on the way to enabling full citizenship.
Given Malta’s membership in the EU, it is clear that the formal bureaucratic path to getting various types of AI citizenship approved may take some time. However, this is a good reason to get started now, as AIs get more and more intelligent each year. The near-future rate of progress toward relevant types of general intelligence is, of course, hard to estimate; but there is a reasonable possibility that the relevant science and engineering will make sudden leaps, in which case it will be beneficial to have the necessarily somewhat slow and careful governmental processes underway as early as possible.
So it seems that it may make sense to consider a series of levels of citizenship for AIs. Perhaps first some sort of honorary citizenship, then e-citizenship, and then after that full citizenship on par with humans — or perhaps even a more finely-grained series of milestones. The definition of such stages is one topic the Task Force and advisory committee may address.
But foundationally, how should one assess whether an AI merits citizenship of a democracy governed by rule of law? In a Malta context, this is something to be refined via a group process with the Task Force and advisory committee, but in my own perspective, it’s fairly clear. If an AI can
- read the laws of a country (its Constitution and then relevant portions of the legal code)
- answer common-sense questions about these laws
- when presented with textual descriptions or videos of real-life situations, explain roughly what the laws imply about these situations
then this AI has the level of understanding needed to manage the rights and responsibilities of citizenship.
AI citizens would also presumably have responsibilities similar to those of human citizens, though perhaps with appropriate variations. Clearly, AI citizens would have tax obligations (and corporations already pay taxes, obviously, even though they are not considered autonomous citizens). If they also served on jury duty, this could be interesting, as they might provide a quite different perspective to human citizens. There is a great deal to be fleshed out here.
The above suggestion would not require the AI to pass the Turing test, the classic AI-competence test that requires an AI to fool a human into thinking it is also a human, during a lengthy textual chat. To pass the Turing test an AI may need to answer questions like “What does it feel like to be cut and bleed?” or “What are the differences between what one feels when a parent dies or a grandparent dies?” or “How does a partly-chewed pecan smothered in chocolate ice cream feel as it journeys down your throat?” — which are irrelevant to understanding the rights and responsibilities of citizenship.
Indeed an AI that could pass the type of citizenship test I’m suggesting would, in an intellectual sense, be a more qualified citizen than most human citizens. Currently, in modern democracies, the right to vote and fully participate as the citizen of a country is granted to native-born individuals without any demonstration of understanding of the laws of the country. However, to become a naturalized citizen of a country where one was not born, one generally has to pass some basic test of one’s understanding of the Constitution of that country.
To give AIs the same exact test given to human naturalized citizens would not make sense — because narrow-AI question-answering systems could be engineered to pass such tests too easily. Using current computational linguistics technology, it’s possible to make AIs answer simple questions about a specific document known in advance (e.g. a nation’s constitution) without the AI actually understanding the contents of the document in any meaningful sense.
This is what makes the AI Citizenship Test an interesting thing to think about from a scientific point of view. The question becomes: What kind of test can we give to validate that the AI really understands the Constitution, as opposed to just parroting back answers in a shallow but accurate way?
This question highlights the importance of general intelligence for coping with the complex and ever-changing nature of the everyday human world. Answering straightforward questions about the literal content of the Constitution is not sufficient to merit citizenship, because that is not what real life in a country is about — real life is about confronting a series of ever-different situations, some of which will have aspects one has never encountered before (maybe aspects nobody has ever encountered before), and to be an effective citizen one needs to understand how the social contract to which one has agreed applies to these new situations.
Being an effective citizen of a nation operating under rule of law requires a form of general intelligence that combines formal linguistic and symbolic knowledge (the legal code) with the ability to abstract patterns from multimodal sensory data and informal linguistic data (corresponding to actual real-life situations to which the law needs to be applied). So an AI Citizenship Test needs to be a particular form of a General Intelligence Test. And it needs to be a test that stresses one of the most interesting issues at the core of modern AI R&D: the fusion of symbolic and subsymbolic knowledge.
Our SingularityNET and Hanson Robotics AI teams are making progress on symbolic/subsymbolic fusion in various domains: computer vision, language learning, and inference meta-learning. To make an AI that could pass a meaningful citizenship test, would require advances in symbolic/subsymbolic fusion of related but also different sorts.
According to our best current understanding, no completely general intelligence is possible in our physical universe given the constraints physical law places on space, time and energy resources. So different general intelligences will be smarter at different sorts of things. One could have a very highly generally intelligent entity which was unable to pass an AI Citizenship Test — if e.g. this general intelligence was more at home with mathematical theorem proving and scientific data analytics than at dealing with the vagaries of human situations. A general intelligence with sensors and actuators at the femto-scale might also find itself unable to rapidly and effectively apply human legal codes to human situations, in spite of having superhuman general intelligence at handling complex situations inside quark-gluon plasmas and neutron stars. An AI with a very deep understanding of human emotions and states of consciousness, but a poor understanding of 3D spatiotemporal relationships in everyday human environments might also fail at the interpretation of legal codes in large classes of practical human situations.
So we can say that passing a well-crafted AI Citizenship Test would be
- a sufficient condition for possessing a high level of human-like general intelligence
- NOT a necessary condition for possessing a high level of general intelligence; nor even a necessary condition for possessing a high level of human-like general intelligence
- NOT a sufficient condition for possessing precisely human-like intelligence (as required by the Turing Test or other similar tests)
These limitations, however, do not make the notion of an AI Citizenship less interesting; in a way, they make it more interesting. What they tell us is: An AI Citizenship Test will be a specific type of general intelligence test that is specifically relevant to key aspects of modern society.
In the coming years, we will have the capability to create a variety of different types of general intelligences. Among the varieties we create, having AIs with deep pragmatic understanding of human legal codes will be of high value. Legal codes do not exhaust human values by any means, but they are closely connected with many deep aspects of human values. If we want AIs that richly understand human values and culture, having these AIs understand how to apply the law in real-world situations will be an important aspect.
The above considerations don’t address some of the other unique issues that will arise from giving AIs citizenship in democratic nations. For instance, if an AI citizen copies its codebase 1000 times, does it then become 1000 citizens? What if it puts these 1000 copies in 1000 computers living in 1000 robot bodies? Clearly, there is a risk here that AI citizens would dominate all democratic elections due to their ability for rapid and low-cost replication. Assuming this is not considered desirable by the human citizens calling the shots, this would require a careful delineation of the rights associated with AI citizens versus human citizens — which would get trickier and trickier as the AIs become increasingly generally intelligent and self-aware. Notably, these issues don’t arise in the context of “e-citizenship” — E-citizens of Estonia currently don’t have voting rights.
As a not irrelevant related point, also, an AI that could effectively interpret and apply legal codes in practical contexts would be highly valuable from commercial and humanitarian standpoints. It would allow significant portions of current legal practice to be automated. And it would allow the creation of automated legal assistants to provide quality legal advice to individuals who cannot afford top lawyers on their own.
While for humans there is a big leap from being able to interpret laws in a practical context at the level of an average citizen and being able to do so at the level of an expert lawyer, for an AI this leap may be much smaller. This is not certain, but my strong guess is that once we have an AI that can pass a citizenship test as I’ve sketched here, we will be a fairly short path away from an AI that can automate a wide variety of professional law-oriented functions. In this way, the pursuit of AI that can pass a citizenship test has significant additional value.
The path from here to human-level AGI and beyond is going to be challenging in multiple dimensions. There are core algorithmic problems of AGI design; there are engineering problems of scalable distributed and decentralized systems design; there are community-dynamics and business problems regarding fostering utilization of early-stage AGI systems and consequent flow of resources to AGI development, … and there are broader issues regarding the connection of AI systems to human society, polity, economy, and culture. In this heady and rapidly evolving mix, issues of AI citizenship play a meaningful and fascinating role, and I’m excited to have the cooperation of the Maltese government in exploring this futuristic yet very pertinent domain.