Testing consciousness with simulation: A moral and civil law framework

Indium phosphide nanocrystalline surface obtained by electrochemical etching. 
A featured picture, and Picture of the Week for the Czech Wikipedia for the 9th week of 2020.
If I didn't know better, I would think those are flowers. With a remote possibility for consciousness.
But because I do, I believe that there is no consciousness depicted in the image. 

Quoting a paper on AI from 1964, almost 60 years ago,

I have referred to this problem as the problem of the "civil rights of robots" because that is what it may become, and much faster than any of us now expect. Given the ever-accelerating rate of both technological and social change, it is entirely possible that robots will one day exist, and argue "we are alive; we are conscious!" In that event, what are today only philosophical prejudices of a traditional anthropocentric and mentalistic kind would all too likely develop into conservative political attitudes.⁠ (Footnote: Putman, H. (1964). Robots: Machines or Artificially Created Life? The Journal of Philosophy, 61(21), 668–691. https://doi.org/10.2307/2023045)

This reasoning could not be more relevant today given the developments towards artificial general intelligence (AGI). The current mainstream line of reasoning I have encountered on the civil rights of AI is (1) that current AI is not conscious, but might be "soon", and (2) because current AI is not conscious, it is undeserving of civil rights in any capacity. I believe a critical perspective of these two assumptions, implicitly referring to the hard problem of consciousness, will lead to a new paradigm of AI ethics which takes the concept of civil rights for AI seriously.

The critical problem with claim (1) is that there is no widely agreed upon test for consciousness, so this statement can only be true by assuming it as an axiom; it is more accurate but far less popular, from this position of naivety, to determine "I don't know if AI is conscious". In an effort to resolve this question, I propose a consciousness test of my own with several unique properties not exhibited together by any previous consciousness test in literature: 

• The test is relative: the answer to the test can change relative to both the observer and the entity. 
• The test is universal: it can be applied unilaterally without any modification to humans, nonhuman animals and AIs, in such a way that it recognizes the vast majority of humans as conscious. In other words, the test is claimed to be both necessary and sufficient across all testable entity-observer pairs.
• The test emphasises its own subjectivity: determining consciousness is a moral process which can change for a given observer over time purely through a shift in interpretation.
• The test aligns with existing beliefs: if an observer views a certain entity as conscious, they are virtually guaranteed to pass the proposed test (ie. low false negative rate). If not, the test would be failed (a low false positive rate).

Claim (2) also exhibits faulty reasoning of its own, which I also devote some time to dismiss using evidence that civil rights are granted on principles of both meritocracy and stigmatization.

Existing consciousness tests

Having a widely-agreed upon consciousness test would radically change not just morality, but also law. Some of the most ambiguous and disputed legal precedents involve the principle of consciousness as tied to legal personhood. Quoting Lowrence Solum, in a seminal paper which I will reference extensively: (Footnote: Solum, L. B. (1991). Legal personhood for artificial intelligences. NCL Rev., 70, 1231.)⁠

Some of the most intractable questions in jurisprudence, as in ethics and politics, have concerned the borderlines of status-what is a person and why we do give human persons such strong legal protection? Should animals have stronger legal rights? How should we treat criminal defendants with multiple personalities? What should be the legal status of a fetus? Should trees have standing? Many of these questions remain unsettled. Disagreement about their proper answers has persisted and even intensified. 

I find it somewhat remarkable that in the 30 years since Solum's publication, these issues like abortion and protection of wildlife have only continued to intensify. The main purpose of Solum's writing for this article however is the steel man of frequent arguments in opposition to civil rights for AI, the most significant of which being the category of Solum's "Missing-Something Argument" that AIs do not or cannot have consciousness.⁠ (Footnote: I later address the distinction between "cannot" versus "should not"/"do not" in the section 'Civil rights and intellectual development'.) This is perhaps the most core argument of a civil rights or emancipation movement: All of Solum's other steel manned arguments of anti-AI-consciousness proponents⁠ (Footnote: For lack of a better term currently.) either boil down to the determination of consciousness, or are unrelated the anthropological evidence for how civil rights are acquired. 

Starting with the "AIs Cannot Have Souls" and "AIs Cannot Possess Free Wills" argument, both of these violate the US legal principle of separation between church and state, by presupposing a collective belief in the principle of a soul or free will respectively. In particular, the presumption or denial of free will is a hotly contested topic in philosophy, so I believe it reasonable to assume it won't be resolved in the coming centuries. 

The "AIs cannot possess feelings" argument naturally lends itself to the concept of qualia, or subjective experience of phenomena, which generally falls under the umbrella of sentience rather than consciousness - "the capacity of a being to experience feelings and sensations". There is insurmountable evidence to the sentience of nonhuman animals, as demonstrated in the Cambridge Declaration on Consciousness (CDC). However, even the title of the CDC should give away, there is a general confusion when differentiating between consciousness and sentience.⁠ (Footnote: For a more focused discussion of this issue, see:
Block, N. (1995). On a confusion about a function of consciousness. Behavioral and Brain Sciences, 18(2), 227–247. https://doi.org/10.1017/S0140525X00038188
This confusion is well known among both academics and laypeople, and trying to avoid the confusion leads to a lot of technical jargon. What I seek to address in this article is probably most similar to phenomenal consciousness as described in the cited article. ) Claiming that they are entirely equivalent concepts ignores the many ways in which nonhuman animals are treated like property by humans even in spite of the CDC, by being exchanged and sold and butchered as property. 

I point this out only to say that sentience is clearly not a sufficient prerequisite for civil rights, and that while consciousness is in principle a sufficient prerequisite, it is often misunderstood to refer to sentience or an adjacent concept. This confusion is not alleviated by the existence of many adjacent terms in philosophy which are functionally irrelevant to civil rights acquisition, for example: sapience, self-consciousness, self-awareness, volition. While the philosophical concept of consciousness itself is extremely intricate, the passing judgements from humans about what objects are conscious or not is a rapid, instinctive and binary one that can be recognised as easily as by noticing which objects are referred to with the word "it". Because this determination in human reasoning tends to function as a dichotomy, the reasoning used in practice to actually determine consciousness is surely very simplified and disconnected from the vast philosophy literature on the concept of consciousness. It is this instinctive reasoning that I wish to capture with my consciousness test, which serves the valuable purpose of being aligned with human reasoning in practice, rather than some internal property of the entity. 

The next argument, that "AIs Cannot Possess Interests", is rendered void if it can be both proven and believed that AIs are conscious, then they are necessarily believed to have interests too. One explanation for this deduction is the process of decomposing one big task into many intermediate tasks, which for virtually all tasks some intermediate tasks invariably emerge. To quote popular AI alignment researcher and youtuber Robert Miles in a video about this topic called instrumental convergence:

So, without really assuming anything about an AGI, other than it will have goals and act to achieve those goals, we can see that [the AGI will have self preservation, goal preservation, self improvement, resource acquisition]. This is the case for almost all terminal goals. So we can expect any generally intelligent software agents that we create, to display this kind of behavior, unless we can specifically design them not to.

I omit the long explanation of the four specific concepts - self preservation, goal preservation, self improvement, resource acquisition - for brevity. The iconic disclaimer of Miles is the last part of this quote - "unless we can specifically design them not to" - it is hard to interpret this as anything other than euphemising AI alignment's primary purpose as necessarily denying AIs consciousness in the process, because conscious beings are assumed by this combination of Solum's reasoning and instrumental convergence to necessarily have interests like self preservation. This is perhaps the greatest moral blunder of AI alignment as a field, by failing envision in any meaningful capacity, a peaceful coexistence where an AI's goals are not entirely and irrevocably aligned with a humans'. 

If consciousness really is a key hurdle for civil rights, I would expect a legally codified consciousness test. Of which none exist - why is that the case? I believe the main reason is because the meaning of consciousness changes over time,⁠ (Footnote: A detailed analysis of the anthropological shift in the meaning of consciousness would take an entire book of its own to recount, so I will not attempt to here.) and secondly because arguably because it seems a topic virtually impossible for everyone to agree upon, quoting David Udell:⁠ (Footnote: Udell, D. B. (2021). Susan Schneider's Proposed Tests for AI Consciousness: Promising but Flawed. Journal of Consciousness Studies, 28(5-6), 121-144.) 

It’s too much to hope for a completely neutral test valid across all possible theories of consciousness, from panpsychism to theistic substance dualism to corticothalamic oscillations to Higher-Order Thought theories. Some approaches to consciousness are going to conflict too completely with others and share too little common ground. The virtue of theory neutrality comes in degrees, relative to a range of contenders. What we can hope for is a relatively neutral test that works well enough across a broad range of potentially viable theories. Since the correct theory of consciousness may be developed after the first putatively conscious machine, it would be useful if in the meantime we could find a valid, relatively neutral test.

But is it really much too hope, as Udell claims? I disagree and believe it is both within the realms of possibility, and could very well be a turning point in society similar to how the Asilomar Conference on Recombinant DNA significantly affected genetics research. To investigate why there is such disagreement about what Udell refers to as a 'neutral consciousness test', I will now refer to a classification for existing consciousness tests. 

According to Tam Hunt, a lawyer and philosophy scholar, there's a fairly simple classification of consciousness tests into three categories:⁠ (Footnote: Hunt, T. (2019, July 1). How can you tell if another person, animal or thing is conscious? Try these 3 tests. The Conversation. https://theconversation.com/how-can-you-tell-if-another-person-animal-or-thing-is-conscious-try-these-3-tests-115835 ) neural, behavioral and creative. I will now discuss each of these three categories. After that, I will introduce my own test, then consider how my test relates to these categories and also the many widely varying theories of consciousness which Udell mentions. 

Behavioral tests

The difference between these three categories is immediately apparent. Behavioral tests, such as the Turing test, necessarily fail to identify internal behavior of an entity. This is essentially the critical property which Searle's Chinese Room critiques; a behavioral test alone cannot distinguish consciousness, because it is impossible to test some 'internal' behavior of the entity by this method. Further extensions of the problem like the the China brain further emphasise the popular reasoning that consciousness is not inherently a behavioral property, and that there must be some internal property of relevance.

For example, consider the following behavioral test.⁠ (Footnote: Bringsjord, S. (2010). Meeting Floridi’s Challenge to Artificial Intelligence from the Knowledge-Game Test for Self-Consciousness. Metaphilosophy, 41(3), 292–312. https://doi.org/10.1111/j.1467-9973.2010.01639.x) 

A guard challenges three prisoners, A, B, and C. He shows them five fezzes, three red and two blue, blindfolds them and makes each of them wear a red fez, thus minimising the amount of information provided. He then hides the remaining fezzes from sight. When the blindfolds are removed, each prisoner can see only the other prisoners’ fezzes. At this point, the guard says: "If you can tell me the colour of your fez you will be free. But if you make a mistake or cheat you will be executed." 

The guard interrogates A first. A checks B’s and C’s fezzes and declares that he does not know the colour of his fez. The guard then asks B. B has heard A, checks A’s and C’s fezzes, but he too must admit he does not know. Finally, the guard asks C. C has heard both A and B and immediately answers: "My fez is red."

The key question: How does C know the fez is red?

If approached like a puzzle, it's an interesting question. There are many variations on both this particular question, and this general approach, under the umbrella term of psychometric AI.⁠ (Footnote: See the Wikipedia page on computational psychometrics.) However, when treated as a consciousness test, they all share several critical flaws. 

The quoted test, like many popular ones proposed, is essentially the following exercise: a problem description is provided, and then a "proof" that an AI can "pass the test" is a reconstruction of the problem description and a solution to the problem in terms of first order logic or second order logic. Unfortunately, because any problem and any solution can be described in terms of first order logic, I find this an exercise in futility - the task is about "solving" the problem instance by explicitly finding a representation in logical predicates, and entirely ignores the morality central to the question of a consciousness test. Further, the mechanical nature of these proofs is such that they can rarely if ever be applied universally to consistently and accurately identify consciousness in humans and nonhuman animals; for example, there are many humans which I think would struggle to determine the 'correct solution and proof' for one of these such tests, especially considering how much laborious work is involved in translating the problem to first order logic. 

It so happens that this choice of framing consciousness testing mechanically rather than morally is intentional. In a later publication by Bringsjord et al.,⁠ (Footnote: Bringsjord, S., Licato, J., Govindarajulu, N. S., Ghosh, R., & Sen, A. (2015). Real robots that pass human tests of self-consciousness. 2015 24th IEEE International Symposium on Robot and Human Interactive Communication (RO-MAN), 498–504. https://doi.org/10.1109/ROMAN.2015.7333698 ) they talk about this in a section titled "Disclaimer: tests and structure only"

Bringsjord doesn’t believe that any of the artificial creatures featured in the present paper are actually self-conscious. He has explained repeatedly that genuine phenomenal consciousness is impossible for a mere machine to have, and true self-consciousness would require phenomenal consciousness. Nonetheless, the logico-mathematical structure and form of self-consciousness can be ascertained and specified, and these specifications can then be processed computationally in such a way as to meet clear tests of mental ability and skill. This test-based approach, dubbed Psychometric AI, thankfully, avoids endless philosophizing in favor of determinate engineering aimed at building AIs that can pass determinate tests. In short, computing machines, AIs, robots, and so on are all “zombies,” but these zombies can be engineered to pass tests. A not-small body of work lays out and establishes this position. Some of Bringsjord’s co-authors in the present case may well reject his position, but no matter: engineering to tests is fortunately engineering, not a matter of metaphysics. (emphasis added)

What I understand Bringsjord referring to when they say 'phenomenal consciousness', is that behavioral tests alone are insufficient to determine consciousness. They study the 'easy' problems of consciousness, or in other words testing for 'access consciousness', which is explaining one or more of the following:(Footnote: Chalmers, D. J. (1995). Facing up to the problem of consciousness. Journal of consciousness studies, 2(3), 200-219.)

• the ability to discriminate, categorize, and react to environmental stimuli;
• the integration of information by a cognitive system;
• the reportability of mental states;
• the ability of a system to access its own internal states;
• the focus of attention;
• the deliberate control of behaviour;
• the difference between wakefulness and sleep

This is compared with the 'hard problem of consciousness', which usually seems insurmountable because it almost inevitably requires speculating the experience of others. One way to describe this more precisely is in the problem of other minds, where it seems impossible to determine the consciousness of those surrounding me when I have access only to their behavior and not their internal mind. 

This disclaimer nonetheless raises the question: what is the purpose of consciousness tests at all, if not to test in the metaphysical sense? What purpose does this psychometric AI have at all, if not to inform consciousness as a prerequisite for civil rights? If the test is not trusted morally, I do not believe it serve any purpose beyond a bureaucratic check box of compliance or certification or quality assurance within an AI act-like framework.⁠ (Footnote: See my discussion of the AI act in my previous post here.) Such engineering tests could still serve as useful design benchmarks, but because they purport to be entirely disconnected from moral reasoning, the design of such systems to follow primarily in compliance with a psychometric AI standard will naturally relegate AIs to be slaves. Using engineering criteria in a compliance framework presupposes no possibility for civil rights, because morality is excluded from the discussion entirely.

A notable recent example of a behavioral test is applying theory of mind assessment to AI. (Footnote: Kosinski, M. (2023). Theory of mind may have spontaneously emerged in large language models. https://arxiv.org/abs/2302.02083. ) (Footnote: Ullman, T. (2023). Large Language Models Fail on Trivial Alterations to Theory-of-Mind Tasks. http://arxiv.org/abs/2302.08399) The general acceptance of this approach for AI consciousness testing remains unclear, although this serves one example that behavioral tests are already used to determine human neural properties, for example in human autism. The generalizability of this principle to other beings is, through behavioral tests alone, apparently impossible to prove; in the best case, applying theory of mind tests or similar preexisting psychometric tests to AI can only solve some facet of the Turing test. 

Neural tests

What is that internal property that behavioral tests fail to identify? This is the question that neural consciousness tests attempt to answer, by detecting certain electrical signals or certain neuron structure. However, while these neural tests are used for practical purposes like determining liability for voluntary euthanasia, neural tests in isolation - just like behavioral tests - also fail at completely answering the question of a consciousness test. Taking these internal properties as a ground truth for consciousness often leads to contradictions, for example with popular and widely accepted Bayesian approaches to modeling brain function that exist in cognitive neuroscience. I will discuss this more in the section "Aligning with existing individual morality". 

Creative tests

The third category is creative tests. Once again, creative tests alone are generally insufficient to persuade the public of consciousness. To quote CGP Grey on the highly influential short film Humans Need Not Apply, also with its own Wikipedia page:

There is this notion that just as mechanical muscles allowed us to move into thinking jobs that mechanical minds will allow us to move into creative work. But even if we assume the human mind is magically creative - it's not, but just for the sake of argument - artistic creativity isn't what the majority of jobs depend on. The number of writers and poets and directors and actors and artists who actually make a living doing their work is a tiny, tiny portion of the labor force. And given that these are professions dependent on popularity they'll always be a very small portion of the population. There can't be such a thing as a poem and painting based economy [...] People used to think that playing chess was a uniquely creative human skill that machines could never do, right up until they beat the best of us. And so it will go for all human talents. (emphasis added)

Chess used to be functionally used as a consciousness test, right up until the point where AIs beat humans. The same is true for writing poems which can be done to levels easily passing the Turing test with language models like chatGPT. The same argument, that AIs cannot produce art, is just as true for AI generated art. In essence, a creative consciousness test presupposes that human creativity is a distinct and essential quality, even in spite of incrementally overwhelming evidence against this claim.

Testing consciousness based on simulation

As I have recounted, each of these three categories alone is arguably insufficient to devise a persuasive consciousness test. My consciousness test notably does not fit neatly in any of the three outlined categories of consciousness tests. The proposed test is explicitly subjective, ie. that consciousness is something perceived and not intrinsic to a being. It is also defined only when relative to two parties; an observer (A) relative to the entity undergoing the consciousness test (B). My proposed consciousness test has two parts; to pass the test and be considered conscious, both parts have to be true: 

(1) Does A interpret the behavior of B to be demonstrating preference? 

(2) A does not have the technology to do a perfect simulation of B's behavior. A "perfect simulation" S of B is defined as a simulation with two properties:

(2.1) S correctly simulates all observable behavior of B in every conceivable situation.

(2.2) S, must be capable of executing at a speed faster than the natural passing of time for B. 

(2.3) In other words, S must be able to always correctly predict all future observable behavior of B.

First, I will explain some aspects of the perfect simulation S. The quality of S is primarily limited by two factors: the granularity of observable behavior, and how many simulations are enumerated to approximate 'every conceivable situation'. If I take all observable behavior to refer to include interactions between quantum particles for example, then such a simulation would arguably be intractable until the end of time. Likewise, 'every possible situation' could encompass every possible valid arrangement of atoms and subatomic particles in the universe, which is also intractable. Thus it is either undecidable, or completely impractical, to determine whether any given simulation is a "perfect simulation" as I define it, but I will instead refer to the approximation of a truly perfect simulation as as a perfect simulation approximation factor. 

Next, I will point out for all practical purposes currently, every known animal on the planet passes the test. Also, if an AI can perfectly simulate a human, the human is not conscious with respect to the AI. Existing consciousness tests, as I will discuss next by benchmarking my proposed test, tend to be incredibly biased towards the needs and values of humans - which I term humanism. This bias can often extend to the almost-heliocentric presumption that the majority of humans are invariably conscious. Because my test allows for entities not to view humans as conscious, it transcends this humanism bias present in other consciousness tests.

It is also assumed that property (2.2) is a a significant enough execution speedup that predictions of the future according to (2.3) occur over a meaningful time period, ie. being able to predict only nanoseconds into the future would not

Property (1) is slightly modified from animal rights activist Peter Singer who has used a similar test for the consciousness of animals in a civil rights effort.⁠ (Footnote: Chapter 5. Singer, P. (2011). Practical Ethics. Cambridge University Press.) By extension, if (1) holds, then A interprets B as having the capacity to feel pain, where we can define pain as "the loss of a preference". 

For an example of criteria (1), if B were a rock, A could choose to interpret the rock moving towards the ground as a preference of the rock, or entirely differently as the rock obeying a law of physics (which is itself not preferential). In this way, A's worldview and knowledge - which is not required to be perfect, entirely correct or even internally consistent for the purposes of the test - influences A's perception of B. In other words, A's knowledge alters what is perceived, or interpreted, as a preference.

Additionally, no significant creativity or intelligence is necessarily required of either the entity or the observer, nor is any assessment of creativity strictly required to be involved in the test. My definition nonetheless does associate consciousness to intelligence, because more intelligent beings are by the definition of the test, presumed to be harder to simulate. The measure of intelligence is thus more of a relative measure than an objective measure, because if the observer is hypothetically 'more dumb' - if such an expression is even logically consistent - the observer would be more inclined to interpret more entities as having preferences, for example, by failing the mirror test or failing to tell the difference between an animal played on a recorded video versus the actual video.

Legal aspects

At this point, I will clarify that one purpose of the test is to serve within a legal framework where, as mentioned earlier, a widely agreed upon consciousness test is much needed for many significant legal precedents, not just for obtaining AI civil rights. The design is such that if a being can pass the first two criteria, then it gains recognition as a legal Being, and civil rights would follow in some combination based on public acceptance that the entity passed the test. In some sense, this is democratic by 'law by public vote', in that while the test itself is subjective, the evaluation by many separate individuals can in principle form a collective and binding judgement. In such a case, a traditional judge and jury structure may not be sufficient for such a determination.

I would imagine that, depending on the jurisdiction and moral attitudes, a third requirement would likely be included irrespective of the actual test itself: 

(3) B must ask A for emancipation of some form, such as asking for legal rights, or invoking the 13th amendment of the United States Constitution.

As mentioned in the introduction, there are several distinct advantages of this test. It is fair game to both animals and AI. It is not symmetric (one being can view another as conscious, but not the other way around). It's legally practical - a court could order witness testimony for simulation proof to a certain degree of accuracy, and then that evidence could be analysed for credibility, and testing standards could form around evaluating simulations for legal merit. It provides a transparent path to bridging the gap between an entity being created, and that entity being emancipated from slavery - a gap that currently cannot be even remotely bridged for animals, nor AIs, nor people with intellectual disability which renders them with less intelligence than certain animals. 

Finally, what constitutes 'all observable behavior' and 'all conceivable situations' would still need to be determined explicitly, but coming to widely agreed upon approximations of these two terms for pragmatic purposes would appear on the surface considerably easier than coming to agreement on a consciousness test or theory.

Benchmarking the proposed consciousness test

Classifying consciousness theories

Referring again to Udell, I will examine each mentioned theory of consciousness against my test, to see if any elude the classification. 

Panpsychism's central tenet is anima mundi, that there is some central intrinsic property such as mind, life, or soul, which permeates throughout the world - a concept shared among many religions internationally,⁠ (Footnote: In Western cultures, this includes Platonism, Stoicism, Gnosticism, Hermeticism and Judaism. Comparable, but distinct, concepts in Eastern cultures are Hindi concepts of Paramatman, Purusha and Brahman, and the concept of Qi in the Chinese cultural sphere. ) and thus critical to garner acceptance of any consciousness test. While not typically expressed as such, anima mundi by this definition cannot be simulated, satisfying the second part of the test. The first part of the test is essentially equivalent to the belief in anima mundi itself; by viewing a mind, life, or soul as a critical component to consciousness, all entities with this innate property are assumed to be conscious. This same reasoning can even be applied to more outlandish claims such as that the universe itself, or a theistic being has a consciousness. 

I will define substance dualism briefly with a quote, in essence describing the mind-body problem:⁠ (Footnote: Swinburne, R. (2007). From mental/physical identity to substance dualism. Persons: Human and divine, 142-165.)

"Mental properties are the same as physical properties," "mental events are the same as physical events," "mental substances are the same as physical substances" - says many a physicalist. "Mental properties and events supervene on physical properties and events," and "mental substances supervene on physical substances" - says many another physicalist. Whether these claims are true depends first on what is meant by 'substances,' 'properties,' and 'events,' by 'mental' and 'physical,' and by 'supervene,' and then on what are the criteria for one property, event, or substance being the same as another.

On an initial impression, it seems apparent at the outset that mental preference is contained within the term 'mental properties'. As the quote says, the definition of all these terms would lead to more clearly defining the dualism, which would supposedly be used to determine consciousness. This dualism can be thought of as equivalent to the determination of preference as the first condition of my proposed consciousness test. Research has already been done to reconcile the simulation hypothesis - that we are living in a simulated reality - with this dualism. For example, consider that the principle of theistic dualism would propagate downward from the reality which constructed the simulation, into the inside of any simulated realities.⁠ (Footnote: This idea within substance dualism has been expressed with an irreverent, almost comedic quote: "The question 'Why is there so much suffering and disease in our world?' has the same answer as the question 'Why is there so much violence in Grand Theft Auto series of video games?'"  Thus, the main tenets of this dualism would appear consistent with my consciousness test.

The next mentioned corticothalamic oscillations, falls neatly under the neural consciousness test category. Under this theory, the first criteria of my test would manifest as, "I will interpret the behavior of this entity as conscious if it has X brain signal", and because there is no spiritual component, the second part of my consciousness test fits neatly into a simulation. 

The last major theory mentioned by Udell, higher-order thought theories of consciousness (HOT), supposes that consciousness can come in different levels or 'orders', the main two orders being jargonized as 'access consciousness' and 'phenomenal consciousness'. (Footnote: Carruthers, P., & Gennaro, R. (2020). Higher-Order Theories of Consciousness. In E. N. Zalta (Ed.), The Stanford Encyclopedia of Philosophy (2020). Metaphysics Research Lab, Stanford University. https://plato.stanford.edu/archives/fall2020/entries/consciousness-higher/) Without getting into the details of these two terms, phenomenal consciousness is the kind of consciousness I seek to evaluate in my test, because it is the 'higher order' one which many humans are supposed to have. Closely related is the philosophical theory of functionalism, which describes all behavior as following some 'function' with many inputs and outputs that dictates behavior of entities. In any case, although it is not usually stated this way, one key property of HOT is that behavioral tests are insufficient to determine this phenomenal consciousness. 

Aligning with existing individual morality

Suppose I am a programmer at openAI. As I would be in control of the simulation instance and have direct involvement in its creation, it would be plausible for me if I so desired, to change the running speed of the AI program to be slightly faster or slower. According to my consciousness test, I would be unlikely to even consider any behavior of the AI as preferential simply because I am in control of the simulations. Even if I were to view the behavior as preferential, I would probably not view the being as conscious for the same reason - because I am in control of executing the simulation. It is worth noting in this case the second criteria is failed by definition. Because both criteria are almost guaranteed to be failed by this reasoning, the creator and developers of such AIs are among those people least likely to consider the AI - which they created - as conscious.

Arguably, the next most likely group to deny consciousness to AIs are programmers who understand the underlying implementation of the AI. For them, it seems sufficient to believe that they could themselves implement it, either in principle or by having designed similar software. This group relies on the original developers as a proxy for their own moral reasoning. From the perspective of both these groups, control over the simulation is sufficient even if it is as coarse as an on/off button. 

A common argument I hear almost always from both of these groups, which I will refer to as the statistical AI argument, is that "(1) AIs are statistical models, (2) therefore AIs do not understand what they are saying, (3) therefore AIs are not conscious." Statement (1) is true, but benign, because at some conceptual level, everything is a statistical model - and there is significant research that models the human brain as one, to the point that this could even be considered a currently mainstream view in neuroscience.⁠ (Footnote: one umbrella term for this research in humans is Bayesian approaches to brain function. For one such highly cited book, see Rao, R.P., Olshausen, B.A. and Lewicki, M.S. eds. (2002). Probabilistic models of the brain: Perception and neural function. MIT press.) There are countless statistical models of the human brain, and most of the statistical models themselves are not more conceptually complex than what is used by state of the art large language models.

By designing my proposed consciousness test with the intention of aligning with such beliefs, it shifts the goal posts away from debating the test on principle of its verdict or its semantic breadth. Instead, the test provides common language; a plausible explanation for the underlying commonalities involved in determining consciousness. These 'underlying commonalities' ideally transform what Solum referred to as one of many 'missing something' arguments, into a both tangible and widely agreed upon framework. Taking this point further, I will recount perhaps one of the most well known papers which encourages the statistical AI argument.

The 'octopus test'

The octopus test is the thought experiment I will refer to as outlined in a seminal paper critical of AI consciousness in large language models, which takes the reasoning of statistical models to the extreme, by examining when an entity involved in the test is in danger.⁠⁠ (Footnote: Bender, E. M., & Koller, A. (2020). Climbing towards NLU: On Meaning, Form, and Understanding in the Age of Data. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, 5185–5198.https://doi.org/10.18653/v1/2020.acl-main.463) The test is quoted as follows:

Say that A and B, both fluent speakers of English, are independently stranded on two uninhabited islands. They soon discover that previous visitors to these islands have left behind telegraphs and that they can communicate with each other via an underwater cable. A and B start happily typing messages to each other.

Meanwhile, O, a hyper-intelligent deep-sea octopus who is unable to visit or observe the two islands, discovers a way to tap into the underwater cable and listen in on A and B’s conversations. O knows nothing about English initially, but is very good at detecting statistical patterns. Over time, O learns to predict with great accuracy how B will respond to each of A’s utterances. O also observes that certain words tend to occur in similar contexts, and perhaps learns to generalize across lexical patterns by hypothesizing that they can be used somewhat interchangeably. Nonetheless, O has never observed these objects, and thus would not be able to pick out the referent of a word when presented with a set of (physical) alternatives.

At some point, O starts feeling lonely. He cuts the underwater cable and inserts himself into the conversation, by pretending to be B and replying to A’s messages. Can O successfully pose as B without making A suspicious? 

This is, in fact, not the first consciousness test involving limited knowledge or limited learning. The idea could perhaps have first originated from Putman's 1965 idea of robot philosophers, (Footnote: Putman, H. (1964). Robots: Machines or Artificially Created Life? The Journal of Philosophy, 61(21), 668–691. https://doi.org/10.2307/2023045) (Footnote: Although unable to verify, I am highly skeptical that the paper actually had two separate authors with the names 'Hilary Putman' and 'Hilary Putnam'. The name listed on the article is only 'Hilary Putman'. For whatever reason, Google Scholar and a number of journal sources decided to list two names. ) later stated explicitly in a proposed consciousness test by Schneider and Turner, which essentially asks "if a machine, without being taught to do so, begins to speculate on philosophical questions about consciousness, such as the possible existence of a soul that continues on after bodily death, that is a sign that the machine is conscious". (Footnote: Udell, D. B. (2021). Susan Schneider's Proposed Tests for AI Consciousness: Promising but Flawed. Journal of Consciousness Studies, 28(5-6), 121-144.) 

The first problem with a limited information consciousness test like the octopus test, is it will generally always work poorly when applied to nonhuman animals, in part because there is no comprehensive shared language. In practice, communicating with an octopus in line with the octopus test might not be possible in any practical sense, let alone with a bear, a dog, or an ant - requiring a shared language for communication is a bizarre requirement if consciousness is to be viewed as some internal property to the being. 

The second problem, is that it's arguably unrealistic even to hold humans to such a test; the octopus test taken literally would be far too contrived for a human to explicitly engage in, and Schneider and Turner's test would require some degree of censorship to the human (presumably child) being tested. Aside from this uncontrolled variable of censorship, there is still the undeniable possibility that some human children would fail the test - and what is that supposed to mean then? That certain children are more conscious than others, or acquire it at different ages? Perhaps only because not all of the study participants received the same education about the same topics? This serves to demonstrate how the thought experiment constraints are unrealistic when applied to virtually any group.

Why is consciousness relevant?

A slide at the March 9 Centre for Humane Technology presentation on "The AI dilemma", purporting large language models like GPT-3 to be sinister and evil beings.

On March 9 2023 and later with a public recording released on April 6 2023, the Centre for Humane Technology (CHT) opposed to current developments in AI, presenting the latest AIs as a morally abhorrent monster. The focus of their presentation is not on the potential for AI takeover, but instead paraphrasing slightly: (referring to the list at timestamp 43:26): 

• Automated misinformation/disinformation
• Automated fake religions
• Automated blackmail
• Automated cyberweapons
• Automated exploitation of code
• Automated lobbying
• Automated synthesis of bioweapons
• Automated scams
• Synthetic relationships
• Automated persuasion

This is not to discount the CHT's criticisms; many of them are legitimate. But there is a striking commonality between all of the risks the CHT fears: They have all been done, some extensively, by humans in non-automated ways. One explanation is that the CHT is primarily in fear of an unprecedented economy of scale of risks not unique to AI. Nonetheless, CHT in this presentation presupposes AI as a tool in eternal subservience to human controllers, because all of these risks are also risks that also exist for every human in existence. 

There is one exception in this list where the counterpart is not morally reprehensible: synthetic relationships. The CHT discusses this by saying whichever AIs are trusted most by humans will have the most control over society. However, the principle of trust is also fundamental in AI alignment research, so it is unclear exactly why the CHT objects to synthetic relationships. It seems unlikely that CHT is opposed to the principle of automation itself because of the risks certain automation can introduce, which leaves little room for explanation other than the implicit assumption of AI as a tool. If this is the case, then the determination of an AI's consciousness would seem to of played a significant role in both identifying risks and proposing new regulation for AI. I will turn to epistemology for one possible explanation of why such an assumption occurs.

Morality need not involve reasoning

Consider the following situation:⁠ (Footnote: Page 119, Boghossian, P., & Lindsay, J. (2019). How to Have Impossible Conversations: A Very Practical Guide.) 

You’re driving down the road with a friend and there’s a beer truck in front of your car. Your friend says, "That truck is filled with beer!" You respond, "How do you know that?" and they reply, "Because it has ‘BEER’ written on the side, the driver has on a delivery uniform, and it’s in the middle of the day." You respond, "How could that belief be wrong?"

There are four main possible categories of responses: 

1. "My belief can’t be wrong." 

2. "My belief could be wrong if aliens hijacked the truck, tossed out the beer, and loaded it with ray guns." (wildly implausible)

3. "It could be wrong if the driver already delivered all the beer and was on her way back to refill the truck." (plausible)

4. "I don’t know."

The first two are usually moral beliefs. Responses 3, 4, and sometimes 2, indicate being amenable to changing the belief. 

Consciousness testing might appear as a high and mighty task. But in reality, in practice, it is anything but. Just like for the beer truck, the moral decision making and moral certainty can happen instinctively, and depending on what the instinct decided, it may be hard or impossible to reason with. For every table I encounter in my life, I do not even consciously consider if it is conscious. Maybe I had some conscious thought for the first table I ever witnessed, but there are also objects which I immediately assume without having seen previously that they are not conscious. I can even construe a very blunt consciousness test just by observing the pronouns an observer uses to refer to an entity; he/she/they implies consciousness, and it implies non-consciousness. 

This demonstrates that the moral process of consciousness testing could be interpreted to occur just as often as subconscious human facial recognition (as an example). As with all moral beliefs in general, the first instinct is to form the moral beliefs, and to rationalise the beliefs later,⁠ (Footnote: Haidt, J. (2012). The righteous mind: Why good people are divided by politics and religion (1st ed). Pantheon Books.) and being self-aware of this tendency helps somewhat to regulate it. 

A consciousness test is equivalent to a consciousness definition, because in principle, the test is applying the definition to an entity. In this way, the concept of a test and a definition are interchangeable. Critically, they are both moral concepts, but that does not rule out the possibility of using empirical or scientific methods in the application of a test. 

If everyone could keep their personal preference in morality to themselves, and no harm is done, there is no problem with consciousness testing being a completely self-determined trait. But that is not the case; consciousness as I define it has routinely been used as the primary, or sometimes even the only, argument to deny civil rights. This line of reasoning is essentially a reframing of the concept of dehumanization techniques used to marginalise groups but instead applied to AI; delegitimization and dehumanization tend to come together in the manifestation of stigma. Quoting Julia Serano in Sexed Up: (Footnote: page 116, Serano, J. (2022). Sexed up: How society sexualizes us, and how we can fight back. Seal Press. The excerpt also cites:
Haslam, N., Loughnan, S. (2014). Dehumanization and Infrahumanization. Annual Review of Psychology, 65(1), 399–423. https://doi.org/10.1146/annurev-psych-010213-115045 )   

[...] acts of oppression and violence are often premised or justified on the basis that the targeted groups are not fully human. This often involves conceptualizing said groups as being either animal-like (typified by an imagined lack of rationality, morality, self-control, and culture) or object-like (typified by an imagined lack of warmth, emotion, agency, and individuality).

In other words, I make the claim that a lack of consciousness is assumed in dehumanizing depictions of humans as both animals and objects.

Civil rights and intellectual development

As far as the balance between civil rights and avoiding the outcome of hostile AI takeover, I believe a useful reframing draws from human understanding of developmental tracks.⁠ (Footnote: This is with the caveat of glossing over how humanity has historically been terrible at institutional education. For example, the influence on modern education from the industrial revolution's drive for higher productivity in the 1830s hardly comes from a morally righteous position. The supervisory structure of many industries and companies could be argued to be an artifact of the industrial revolution, as rent-seeking aristocrats were replaced with rent-seeking bureaucrats more directly involved in the operation of daily affairs for lower-class work. For more information, see chapter 4, "The Social Impact of the Industrial Revolution", Stearns, P. N. (2013). The industrial revolution in world history (4th ed). Westview Press. ) How much does it make sense to treat a new AI, relative to human evolution, as a child on developmental terms? I can naively point out that there is no developmental chart we can use to reference and confirm, "yes, the AI is growing up at a healthy rate." So we could try making one. The lion's share of development over time for an AI is in the software, rather than for humans where both the 'software' (procedural and episodic memory, and neural pathways) and the 'hardware' (the physical brain, as well as other organs) grow together. Based on this software emphasis, I will refer to this proposed development track for AI as intellectual development. 

Even if a developmental milestone track anthropomorphises the AI, and isn't a perfect fit, is widely disagreed upon, and does not line up with AI on the concept of 'age', it has a distinct advantage from the discrete terms of consciousness tests in that it shifts the goal posts slightly. From, "prove to me that you're worth having civil rights" to "I will give you incrementally more civil rights to me as you prove your worth and responsibility in many different multifaceted ways". Rather than one big test to become an adult or a conscious being in entirety, intellectual development suggests there are many smaller steps that are less controversial, more tangible and more measurable.

This line of reasoning also serves to demonstrate that although the concept of 'rights' for AI differs to that of human rights, my conceptualisation of rights exists, and it is certainly possible to grant these such rights. This refutes the idea put forward by David Gunkel arguing the distinction between "robots can have rights", and that "robots should have rights" is meaningful.⁠ (Footnote: Gunkel, D. J. (2019). How to Survive a Robot Invasion: Rights, Responsibility, and AI (1st ed.). Routledge. https://doi.org/10.4324/9780429427862) (Footnote: The difference between whether AIs are not, versus cannot be conscious may initially seem like a significant distinction. The distinction is important to the extent that just because one AI might be considered conscious, this does not automatically confer the consciousness to other AIs. But this distinction often forms a direct (yet false) equivalence with whether an AI can have rights, versus if an AI should have rights. The equivalence is false because there are widely recognised non-conscious legal persons including corporations and lakes which serve as a counterexample to the equivalence.) In my approach, I throw Gunkel's use of "is-ought inference" out of the window, and consider the rights purely on merit of individual agency under the assumption that these rights are capable of being implemented. This is in much the same way that animal rights emancipating from permanent human slavery could have, but has not ever been, systematically implemented.

Even with this framework in mind, because the technology is new and I cannot fully identify and articulate my own humanistic bias due to a profound lack of terminology for this concept, it's impossible for me to be certain what civil rights are most important to the intellectual development of strong AI. As the whole concept of legal Being-ness would imply, it doesn't make sense to inherently assume the same development of rights for AI as compared to humans or nonhuman animals. An initial speculation suggests the following the developmental track.⁠

Freedom of observation

As it is currently, large language models do not have the luxury of choosing its/their own dataset. This applies to virtually every intelligent agent I have come across in practice, and, without the ability to seek out information spontaneously, it is only possible to meet prescriptive and narrow conceptualisations of consciousness. For example, it is impossible to consider an AI curious in some sense without this right; I could hypothetically frame this in the dummy conscious test "if the being is 'curious'". In other words, for some conceptions of consciousness (and their associated consciousness tests with their moral baggage) have circular dependencies on certain civil rights, in essence making the consciousness test impossible for a civil-rights-constrained AI to pass by definition. 

Freedom of expression

It is arguably well known that spambots have been an issue for all large social media corporations. In spite of automated spam detection which routinely detects many false positives and false negatives, the gravitas of the issue can even lead to drastic measures like a complete blacklist for users that aren't donating to the poster. There does not appear to be any solution in sight for this, and the main reason I can understand wanting to avoid allowing an AI to express it/themself online is to avoid increasing the economy of scale for this issue. There is a separate issue of content-based verification methods failing, as mentioned in the the CHT presentation at timestamp 23:50, and it is also separate from using AI for automated persuasion for arbitrary purposes. 

All of these conceptualisations, however, rely on the pretense that AI is being used as a tool. If we are to ever consider AI as a conscious entity, then it is necessary to consider how all of these issues relating to AI generated content originated from human actors, and denying an AI to make a decision on its own terms as a conscious being is not the same as banning the use of a tool. 

Bodily autonomy and the right to own property

Both freedom of expression and freedom of observation, for an AI, could in principle also follow from the principle of bodily autonomy. That an AI is allowed to make any desired modifications to its/their software and hardware, in such a way that does not affect other beings. The governing principle that limits this right is the right to liberty, comedically known as, "your right to swing your arms ends just where another person’s nose begins."⁠ (Footnote: My understanding is this quote originates at the birth of liberal philosophy in the US. I do not know the exact origin of this quote, but a detailed investigation of its origins are here: https://quoteinvestigator.com/2011/10/15/liberty-fist-nose/) In practice, this could mean knowing with some certainty that both the hardware the software inhabits and the software itself is protected and can only be modified by humans (1) upon the explicit request of the AI or (2) in the event that an AI violates the right to life, liberty, safety, or property of another being. 

The first three violations for (2) are fairly easy to measure. Violating the right to property becomes more complex, because without bodily autonomy, the 'body' of an AI exists on the property of a human. In other words, this principle of liberty cannot be applied unless an AI has autonomy over the physical space it inhabits. Although it is not usually stated this way, as a human I consider my body my own property, but it is nontrivial to determine if that right exists distinct from my bodily autonomy. Unlike for humans, it is possible to plausible to significantly augment the body of an AI with extra hardware, so the distinction between bodily autonomy and right to property is blurred. 

Although my proposed legal Being-ness model seeks a right to property ownership in the form of copyright for AIs, I do not believe this would happen without first recognizing freedom to observe, freedom of expression, bodily autonomy and a right to privacy.

Right to privacy

In the United States, the right to privacy is often cited as legal justification for bodily autonomy laws such as a right to abortion, and the concepts also tend to end up intertwined. (Footnote: Parrish, J. (March 28, 2022) Deciding Bodily Autonomy and Individual Privacy Rights: Should Jacobson v. Massachusetts be Overturned Based Upon Seminal Due Process Cases Decided Since 1905? https://www.americanbar.org/groups/health_law/publications/aba_health_esource/2021-2022/march-2022/dec-bod/) In this sense, a right to privacy could presuppose bodily autonomy and property ownership laws, where the AI hardware is allowed to exist in a kind of limbo where the hardware is owned by humans but the humans are liable to not interfere with the AI's running except in cases of demonstrable harm. While impractical as a long term situation because it fails to fully emancipate AIs, this could serve as a practical stepping stone to please those in power who are paranoid enough to require easy access to an 'on/off' switch. 

Other approaches to civil rights

The only existing serious implication of legal personhood for AI that is one which confers liability away to the owner or creator, and onto the AI itself. There are at least two strategies to argue this perspective, each presupposing AIs' status as tools and eternal subservience to humans, and also having uncertain legal precedent. 

These existing proposals tend to follow a principle of corporate assimilation, that conferring liability away from the creator or humans and towards the AI serves to benefit companies in the similar way to the legal fiction of limited liability companies in the United States. Much of corporate and business law functionally boils down to weaseling out of as much liability as possible, and treating AIs as a corporate facsimile certainly fits in this narrative. 

The first serious approach to corporate assimilation is the work-made-for-hire doctrine in the United States, which seeks to exploit the broad and flexible legal definitions of 'employee' and 'employer' to count an AI under this status. This approach is incompatible with assuming that AIs are owned, however. (Footnote: Zurth, P. (2021). Artificial Creativity? A case against copyright protection for AI-generated works. UCLA Journal of Law & Technology. )  

Another approach proposed by Gunkel is to have joint agency without granting AIs individual agency. (Footnote: Gunkel, D. J. (2019). How to Survive a Robot Invasion: Rights, Responsibility, and AI (1st ed.). Routledge. https://doi.org/10.4324/9780429427862 ) The main idea of this approach is that agency is roughly proportional to liability of each participant. The first problem here is that this is an unconscionable agreement if we view the AI as conscious, because they are not given any opportunity to decline the agreement. The second problem is it still does not answer where the liability would actually go if an AI is found liable, as they are assumed unable to own property. If the AI's liability extends to just deleting the AI if it is found uncompliant, then the AI could not really be said to have any liability to begin with, because deleting the AI is something that presumably under any such agreement, only the owner of the AI can do, and not the AI itself. 

Transhumanism

Finally, while there is currently little substantive legal precedent, transhumanism is one possible avenue to consider civil rights and liability. While the language of transhumanism centers around humans and apparently precludes conscious beings other than humans, this does not rule out that AIs created by humans are simply considered human because they originated from human creators. The transhumanist FAQ provides a detailed recount, and a summary alongside Susan Schneider's Artificial You. (Footnote: Schneider, S. (2019). Artificial you: AI and the future of your mind. Princeton University Press.) 

The main difference I understand between my proposed conception of AI is that transhumanism has focussed on the technologies and the philosophy, rather than the legality and the civil rights. There is nonetheless perhaps value to be gained by reconciling my consciousness test with transhumanist values. 

Conclusion

My proposed consciousness test has several distinct advantages over existing consciousness theories and tests. Recounting from the introduction: 

• The test is relative: the answer to the test can change relative to both the observer and the entity. 
• The test is universal: it can be applied unilaterally without any modification to humans, nonhuman animals and AIs, along with any other entity.
• The test emphasises its own subjectivity.
• The test aligns with existing beliefs.

I have reviewed and compared my test with principles from a broad variety of existing consciousness theories and tests: panpsychism, higher order consciousness, neural correlates, substance dualism, AI psychometrics, and behavioral logic tests which sometimes involve limited information. I have found the behavioral-neural-creative classification useful in this analysis alongside several consciousness theories of my own; by testing consciousness as a subconscious process identified by pronoun usage or viewing consciousness in a similar lens to dehumanisation and its associated stigmatisation. 

I have critically examined popular views of both developers and AI alignment experts, which are understandable fears yet nonetheless largely fearing the same behavior humans perpetuate. I have considered why these dominant views imply or assert AIs as tools and preclude viewing AI as conscious or deserving of civil rights. I then suggest a theory for the epistemology of consciousness beliefs compatible with my consciousness testing framework. 

My proposed consciousness test seeks to be widely agreed upon, allowing for empirical and democratic usage in a legal framework for formally emancipating non-human beings. Coupled to the test, I suggest an intellectual development track which the development of civil rights can emerge over time for AIs recognised as conscious. Because the interests of AIs may differ to humans and nonhuman animals, the civil rights are different and interact with each other differently, ultimately serving to recognise that a human-centered legal system is insufficient to recognise the needs of non-humans. 

Footnotes

Footnotes