Unprecedented Scientific Censorship

Scientific discourse is considered one place where you can present certain kinds of truth as accurately as possible, regardless of whether they conform to the prevailing orthodoxies or not, whether they are truths that most people want to listen to or not, and whether they agree with political ideologies or not. It used to be the case that most of scientific discourse was on matters which did not directly and immediately interest or concern either the general public or, to a lesser extent, even the powers that be. And so, scientists were able to pursue their research with tolerable hindrance from the circumstances and people in which and among whom they lived and worked.

This started changing when the modern Industrial, and then Corporate — apart from the state — establishment developed not only huge stakes in scientific research, but started funding most of it, not just for courtly splendour as was the case in the age of old feudalism. With funding came control. Simultaneously, with the neoliberal/neoconservative dominance of the world, government funding for independent research started diminishing at an ever increasing rate. This inevitably meant that scientific community came under heavy influence of state and corporate actors.

In the 21st century, this influence is transforming into more and more tighter form of control over not just what research is carried out, but how, to what end, and even with regard to whether it produces ‘desirable’ results or not.

The Pandemic of 2020 has made this phenomenon of tight control over scientific research more widespread as well as more visible. With it, however, has come (perhaps fittingly) an extremely shrill rhetoric of “You don’t believe in science?!” and “Science says so and so”, where so and so could be a very obviously a debatable matter (or not: it doesn’t make a difference). In other words, on the one hand, science is becoming more like religion, both in terms of concepts like heresy, blasphemy and blind (or at least uncritical) belief, and in terms of censorship of expression, even scientific expression. Genuine scientific debates are becoming more like theological conflicts, as the science wars about the Pandemic have revealed.

This is also the time when Artificial Intelligence (AI) is all the rage. It is being touted as the Silver Bullet to solve all of humanity’s problems, current and future. No wonder then that AI too is seriously in danger of becoming a theology and a church, rather science and technology. Perhaps the best example of this is the recent case of a paper on ethics of AI, co-authored by mainstream AI ethicists and researchers, which caused Google to ask one of its authors to retract the paper. Timnit Gebru, the co-lead of Google’s ethical AI team, was a co-author of the paper. She has since left her job rather than agreeing to retract the paper. Many researchers cannot afford to do that, and the paper might be published, but still this case is unprecedented.

I had my own experience with scientific censorship recently. I have been working on a paper about the impossibility of humanoid artificial intelligence, but I could not think of a suitable venue for this paper, since it seems to go against one of the most dearly held ideas about AI: that true humanoid AI is not only possible, but inevitable. Since the draft was written in a semi-formal style, using arguements against the possibility of humanoid AI, analogous to the arguments philosophers have been using for and against the possibility of a Single Supreme God. In my view, building humanoid AI will require AI as a whole to become a Single Supreme God, at least as far as human affairs are concerned. The arguments centred around the distinction between Micro-AI and Macro-AI.

Then I came across an unusual research workshop at the most well known AI conference (Neural Information Processing Systems or NeurIPS 2020), which was titled ResistanceAI. It invited papers and even media, including those not in an academic form or format. It seemed perfect to me, so I decided to submit my draft at this workshop. Since it is a common practice now to post such drafts (preprints) on the best known scientific archive or preprint hosting site arXiv. I already have posted several papers on arXiv. Since such preprint sites are meant for archival purposes, they do not put the papers through a peer review process, as that is going to happen anyway when the paper is submitted to a peer reviewed venue. Usually, the paper is posted directly after a kind of sanity check. Sometimes, however, arXiv puts a paper through moderation, which usually involves reclassification of the paper under suitable categories. In very rare cases, a paper can be removed. The reasons for such removal are supposed to be:

  • Unrefereeable content
  • Inappropriate format
  • Inappropriate topic
  • Duplicated content
  • Rights to submit material
  • Excessive submission rate

Based on the description of these reasons given on their moderation page, none of these apply in anyway to my draft. I had submitted the paper on 8th October 2020. I first received a mail saying it will be ‘announced’ (that is, posted) the next day. Then, on 14th October 2020, I received a mail saying that the paper has been ‘put on hold’. Initially I assumed it must be for reasons of reclassification. However, on the same day, I received another mail saying the paper has been removed. The mail said:

Dear arXiv user,

Our moderators have determined that your submission is not of sufficient interest for inclusion within arXiv. The moderators have rejected your submission after examination, having determined that your article does not contain sufficient original or substantive scholarly research.

As a result, we have removed your submission.

Please note that our moderators are not referees and provide no reviews with such decisions. For in-depth reviews of your work, please seek feedback from another forum.

Please do not resubmit this paper without contacting arXiv moderation and obtaining a positive response. Resubmission of removed papers may result in the loss of your submission privileges.

For more information on our moderation policies, see:

https://arxiv.org/help/moderation

Regards,
arXiv moderation

The reason given (“your article does not contain sufficient original or substantive scholarly research”) was a kind of review itself, which is not supposed to be there as a reason for removal, since duplication means direct duplication, not extending existing ideas. The reason can be reasonably interpreted as saying simply that some references were missing from the paper, meaning that it was a kind of feedback to me about the paper, which arXiv is not supposed to give.

This came right before the deadline for submission at the ResistanceAI workshop. So I added a few of the missing references, given the page limit of four pages. The paper was, however, rejected at the workshop, although I did receive a review of the paper. Note that one of the reasons for removal from arXiv is “unrefereeable content”. So, clearly, the paper was not unrefereeable.

The review from the workshop is given below:

Reviewer #1
Questions

2. Please provide constructive feedback to the authors
This paper address some timely questions about what we might expect the “Singularity” to look like. Unfortunately, section three–the meat of the paper–is somewhat difficult to follow. Rather than listing many different arguments, it may be more helpful to focus on a subset of these arguments and explain how they are related. As currently written, it is difficult to understand the argument and how it reaches the conclusions that “Singularity at the level of Micro-AI is impossible” and that a Singularity at the “Macro-AI level” would be an existential threat to human intelligence.
3. Please give this submission a score
Weak Reject

Reviewer #2
Questions

2. Please provide constructive feedback to the authors
1/ The paper, while looking at the impact of a hypothesized ‘Macro AI’ on human beings in the future, ignores the issues that AI technology is causing in the present.
2/ In particular, it fails to inspect and analyze the material impact that AI is already causing in the lives of human beings, whether or not it is a ‘humanoid’ AI which is doing that.
3/ Overall, the paper does not fit the theme of the workshop — which has more to do with how AI concentrates power in the hands of a few, rather than hypothesizing about the future of AI and what that means for humanity, without grounding it in a material analysis.
3. Please give this submission a score
Strong Reject

Although I at least received reviews of the paper, the reasons given here are highly questionable, particularly in the light of the fact that the workshop has accepted not just papers, but also poems, rants, essays etc., and even an anonymous submission, which is never the case at a research venue. In particular, the reviewer statement, “ignores the issues that AI technology is causing in the present”, does not make sense. In a four page paper, when dealing with a topic like this, how can one include a survey of harms already being done by AI? I have, in the past, written at least one paper on such harms, which is (ironically) hosted on arXiv. That paper was rejected without review from the conference where it was submitted simply because I mistakenly did not notice that the paper, before submission, had (at the last moment) exceeded the four page limit by a two or three (one column) lines.

I had then two options, apart from working further on the paper and submitting it to another peer reviewed venue. One was to appeal the decision by arXiv, which I might still do, and the other was to post the draft on some other preprint site. I found two alternatives for the second option. One was the PhilSci Archive for preprints in philosophy of science. The second was HAL Archive.

I posted on both of them. The draft was again rejected from the PhilSci Archive, giving the following reason:

Unfortunately the item could not be accepted into PhilSci-Archive. The item lies outside the range of material suitable for PhilSci-Archive. We regret that because of the volume of material posted, the archive cannot enter into correspondence concerning submissions that have been refused.

This may be debatable, since it seems to me the paper is well within the scope of philosophy of science.

The preprint has finally been accepted by the HAL Archive, after they asked me to first post a paper already published in a scientific journal ‘in order to establish a confidence contract’, which sounds reasonable.

I am working on improving the draft with the possibility of submitting it to another venue, preferably peer reviewed. However, in the fifteen years since I first published a peer reviewed paper, this has been the strangest case of rejection by multiple venues, not just by peer review, but by two different preprint sites, one of them (PhilSci) does not even have a moderation process according to their policy.

Even so, this is not the first case of strange rejection that I have experienced from peer reviewed venues. Till recently, it could be attributed to the inherently imperfect nature of the peer review process, but now it seems to be clearly going beyond that, as the Google case shows, if not also the case of my paper.

The Impossibility Conjecture of Humanoid Artificial Intelligence and the Non-Benign Singularity

Abstract

[A Rough Draft of a Work-in-progress.]

The idea of machines which are almost identical to human beings has been so seductive that it has captured the imaginations of the best minds as well as laypeople for at least a century and half, perhaps more. Right after Artificial Intelligence (AI) came into being, it was almost taken for granted that soon enough we will be able to build Humanoid Robots. This has also led to some serious speculation about ‘transhumanism’. So far, we do not seem to be anywhere near this goal. It may be time now to ask whether it is even possible at all. We present a set of arguments to the effect that it is impossible to create or build Humanoid Robots or Humanoid Intelligence, where the said intelligence can substitute human beings in any situation where human beings are required or exist.

1. Humanoid Intelligence, the Singularity and Transhumanism

Before we proceed to discuss the terms of the title of this section and the arguments in the following sections, we first define the foundational terms to some degree of conciseness and preciseness:

1. Human Life: Anything and everything that the full variety of human beings are capable of, both individually and collectively. This includes not just behaviour or problem solving, but the whole gamut of capabilities, emotions, desires, actions, thoughts, consciousness, conscience, empathy, creativity and so on within an individual, as well as the whole gamut of associations and relationships, and social, political and ecological structures, crafts, art and so on that can exist in a human society or societies. This is true not just at any given moment, but over the life of the planet. Perhaps it should include even spiritual experiences and ‘revelations’ or ‘delusions’, such as those hinted at in the Philip K. Dick story, Holy Quarrel [Dick et al., 1985].

2. Humanoid: A living and reproducing entity that is almost identical to humans, either with a human-like body or without it, on a different substrate (inside a computer).

3. Intelligence: Anything and everything that the full variety of human beings are capable of, both individually and collectively, as well as both synchronically and diachronically. This includes not just behaviour or problem solving, but the whole of life as defined.

4. The Singularity: The technological point at which it is possible to create (or have) intelligence that is Humanoid or better than Humanoid.

5. Transhumanism: The idea that, after the singularity, we can have a society that is far more advanced, for the better, than the current and past human societies. From 1910 to 1927, in the three volumes of Principia Mathematica [ 1925–1927], Whitehead and Russell set out to prove that mathematics is, in some significant sense, reducible to logic. This turned out to be impossible when Godel published his incompleteness theorems in 1931 [Sheppard, 2014, Nagel et al., 2001]. During the days of origins of modern Computer Science, before and in early 1930s, it would have been easy to assume that a computing machine would ultimately solve any problem at all. This also proved to be impossible with Turing’s undecidability theorem [Hopcroft et al., 2006] and the Church-Turing thesis of computability [Copeland and Shagrir, 2018]. Since then, other kinds of problem have been shown to be undecidable.

Now that we are supposed to close be enough to the Singularity [Kurzweil, 2006] so that it may happen within the lifetime of a large number of human beings, perhaps it is time to ask ourselves whether real intelligence, in particular Humanoid Intelligence (as defined above) is possible at all. We suggest that there are enough arguments to ‘prove’ (in an informal sense) that it is impossible to build, to create or to have Humanoid Intelligence. We argue that even though the Singularity is indeed possible, perhaps even very likely (unless we stop it), it may not be what it is supposed to be. The conjecture presented here is that the Singularity is not likely to be even benign, however powerful or advanced it may be. This follows from the idea of the impossibility of Humanoid Intelligence.

2 Some Notes about the Conjecture

We have not used the term theorem for the Impossibility and the reasons for this should be evident from the arguments that we present. In particular, we do not, and perhaps cannot, use formal notation for this purpose. Even the term conjecture is used in an informal sense. The usage of terms here is closer to the legal language than to the mathematical language, because that is the best that can be done here. This may be clearer from the Definition and the Story arguments. It is due to a similar reasoning that the term ‘incompleteness’ is not used and, instead, impossibility is used, which is more appropriate for our purposes here, although Godel’s term ‘essentially incomplete’ is what we are informally arguing for about Humanoid AI, and perhaps AI in general. No claim is made as to whether or not a formal proof is possible in the future at all. What we present is an informal proof. This proof has to be centred around the distinction between Micro-AI (AI at the level of an intelligent autonomous individual entity) and Macro-AI (very large intelligent autonomous systems, possibly encompassing the whole of humanity or the world). To the best of our knowledge, such a distinction has not been proposed before. While there has been some work in this direction [Brooks, 1998, Signorelli, 2018, Yampolskiy, 2020], for lack of space, we are unable to explain how this work differs from previous such works, except by noting that the argumentation and some of the terms are novel, a bit like in the case of arguments for or against the existence of God, which question has been debated by the best of philosophers again and again over millennia, which as we will see at the end, is relevant to our discussion.

3 The Arguments for the Impossibility Conjecture for Micro-AI

The Definition Argument): Even the Peano Arithmetic [Nagel et al., 2001] is based on three undefined terms (zero, number and is successor of ), which are relatively trivial terms compared to the innumerable terms required for AI (the core terms like intelligence and human, or terms like the categories of emotions, leave alone the terms like consciousness).

The Category Argument: A great deal of AI is about classifying things into categories, but most of these categories (e.g. anger, disgust, good or bad) have no scientifically defined boundaries. This is related to the following argument.

The Story Argument: It is almost established now that many of the essential concepts of our civilisation are convenient fictions or stories [Harari, 2015] and these often form categories and are used in definitions.

The Cultural Concept Argument: Many of the terms, concepts and stories are cultural constructs. They have a long history, most of which is unknown, without which they cannot be modelled.

The Individuality, or the Nature Argument: An individual intelligent autonomous entity has to be unique and distinct from all other such entities. It originates in nature and we have no conception of how it can originate in machines. We are not even sure what this individuality exactly is. However, all through history, we have assigned some degree of accountability to human individual and we have strict provisions for punishment of individuals based on this, that indicates that we believe in the concept of the ‘self’ or the ‘autonomous individual’, even when we deny its existence, as is becoming popular today.

The Genetic Determinism Argument: Individuality is not completely determined by nature (e.g. by our genes) at birth or creation once and for all. It also develops and changes constantly as it interacts with the environment, preserving its uniqueness.

The Self-organising System Argument: Human beings and the human societies are most likely self-organising [Shiva and Shiva, 2020] and organic systems, or they are complex, non-equilibrium systems [Nicolis and Prigogine, 1977]. If so, they are unlikely to be modelled for exact replication or reproduction. The Environment, or the Nurture Argument: Both intelligence and individuality depend on the environment (or on nature). Therefore, they cannot be modelled without completely modelling the environment, i.e., going for Macro-AI. The Memory, or the Personality Argument: Both intelligence and individuality are aspects of personality, which is known to be dependent on the complete life-memory (conscious and unconscious) of an intelligent being. There is not enough evidence that it is possible to recover or model this complete temporal and environmental history of memory. A lot of our memory, and therefore our individuality and personality is integrally connected with our bodily memories.

The Susbstrsate Argument: It is often taken for granted that intelligence can be separated from the substrate and planted on a different substrate. This may be a wrong assumption. Perhaps our intelligence is integrally tied with the substrate and it is not possible to separate the body from the mind, following the previous argument.

The Causality Argument: There is little progress in modelling causality. Ultimately, the cause of an event or occurrence is not one but many, perhaps even the complete history of the universe.

The Consciousness Argument: Similarly, there is no good enough theory of consciousness even for human understanding. It is very unlikely that we can completely model human consciousness, nor is there a good reason to believe that it can emerge spontaneously under the right conditions (which conditions?).

The Incompleteness/Degeneracy of Learning Source and Representation Argument: No matter how much data or knowledge we have, it will always be both incomplete and degenerate, making it impossible to completely model intelligence.

The Explainability Argument: Deep neural networks, which are the state-of-the-art for AI, have serious problems with explainability even for specific isolated problems. Without it, we cannot be sure whether our models are developing in the right direction.

The Test Incompleteness Argument: Perfect measures of performance are not available even for problems like machine translation. We have no idea what will be the overall measure of Humanoid Intelligence. It may always be incomplete and imperfect, leading to uncertainty about intelligence.

The Parasitic Machine Argument: Machines completely depend for learning on humans and on data and knowledge provided by humans. But humans express or manifest only a small part of their intelligent capability. So machines cannot completely learn from humans without first being as intelligent as humans.

The Language Argument: Human(oid) Intelligence and its modelling depend essentially on human language(s). There is no universally accepted theory of how language works.

The Perception Interpretation Argument: Learning requires perception and perception depends on interpretation (and vice-versa), which is almost as hard a problem as modelling intelligence itself.

The Replication Argument: We are facing a scientific crisis of replication even for isolated problems. How could we be sure of replication of Humanoid Intelligence, preserving individual uniqueness?

The Human-Human Espitemic Asymmetry Argument: There is widespread inequality in human society not just in terms of money and wealth, but also in terms of knowledge and its benefits. This will not only reflect in modelling, but will make modelling harder.

The Diversity Representation Argument: Humanoid Intelligence that truly works will have to model the complete diversity of human existence in all its aspects, most of which are not even known or documented. It will have to at least preserve that diversity, which is a tall order.

The Data Colonialism Argument: Data is the new oil. Those with more power, money and influence (the Materialistic Holy Trinity) can mine more data from others, without sharing their own data. This is a classic colonial situation and it will hinder the development of Humanoid Intelligence.

The Ethical-Political Argument: Given some of the arguments above, and many others such as data bias, potential for weaponisation etc., there are plenty of ethical and political reasons that have to be taken into account while developing Humanoid Intelligence. We are not sure whether they can all be fully addressed.

The Prescriptivastion Argument: It is now recognised that ‘intelligent’ technology applied at large scale not only monitors behaviour, but changes it [Zuboff, 2018]. This means we are changing the very thing we are trying to model, and thus laying down new mechanical rules for what it means to be human.

The Wish Fulfilment (or Self-fulfilling Prophecy) Argument: Due to prescriptivisation of life itself by imperfect and inadequately intelligent machines, the problem of modeling of Humanoid Intelligence becomes a self-fulfilling prophecy, where we end up modeling not human life, but some corrupted and simplified form of life that we brought into being with ‘intelligent’ machines.

The Human Intervention Argument: There is no reason to believe that Humanoid Intelligence will develop freely of its own and will not be influenced by human intervention, quite likely to further vested interests. This will cripple the development of true Humanoid Intelligence. This intervention can take the form of secrecy, financial influence (such as research funding) and legal or structural coercion.

The Deepfake Argument: Although we do not yet have truly intelligent machines, we are able to generate data through deepfakes which are not recognisable as fakes by human beings. This deepfake data is going to proliferate and will become part of the data from which the machines learn, effectively modeling not human life, but something else.

The Chain Reaction Argument (or the Law of Exponential Growth Argument): As machines become more ‘intelligent’ they affect more and more of life and change it, even before achieving true intelligence. The speed of this change will increase exponentially and it will cause a chain reaction, leading to unforeseeable consequences, necessarily affecting the modelling of Humanoid Intelligence.

4 The Implications of the Impossibility

It follows from the above arguments that Singularity at the level of Micro-AI is impossible. In trying to achieve that, and to address the above arguments, the only possible outcome is some kind of Singularly at Macro-AI level. Such a Singularity will not lead to replication of human intelligence or its enhancement, but something totally different. It will, most probably, lead to extinction (or at least subservience, servitude) of human intelligence. To achieve just Humanoid Intelligence (Human Individual Micro-AI), even if nothing more, the AI system required will have to be nothing short of the common notion of a Single Supreme God. Singularity at the macro level will actually make the AI system, or whoever is controlling it, individual or (most probably small) collective, a Single Supreme God for all practical purposes, as far as human beings are concerned. But this will not be an All Powerful God, and not a a Kind God, for it will be Supreme within the limited scope of humanity and what humanity can have an effect on, and it will be kind only to itself, or perhaps not even that. It may be analogous to the God in the Phiilip K. Dick story Faith of Our Fathers [Dick and Lethem, 2013], or to the Big Brother of Orwell’s 1984 [Orwell, 1950]. We cannot be sure of the outcome,
of course, but those as likely outcomes as any others. That is reason enough to be very wary of
developing Humanoid Intelligence and any variant thereof.

References

Philip K. Dick, Paul Williams, and Mark. Hurst. I hope I shall arrive soon / Philip K. Dick ; edited by Mark Hurst and Paul Williams. Doubleday New York, 1st ed. edition, 1985. ISBN 0385195672.

Alfred North Whitehead and Bertrand Russell. Principia Mathematica. Cambridge University Press, 1925–1927.

Barnaby Sheppard. Gödel’s Incompleteness Theorems, page 419–428. Cambridge University Press, 2014. doi: 10.1017/CBO9781107415614.016.

E. Nagel, J.R. Newman, and D.R. Hofstadter. Godel’s Proof. NYU Press, 2001. ISBN 9780814758014. URL https://books.google.co.in/books?id=G29G3W_hNQkC.

John E. Hopcroft, Rajeev Motwani, and Jeffrey D. Ullman. Introduction to Automata Theory, Languages, and Computation (3rd Edition). Addison-Wesley Longman Publishing Co., Inc., USA, 2006. ISBN 0321455363.

B. Jack Copeland and Oron Shagrir. The church-turing thesis: Logical limit or breachable barrier? Commun. ACM, 62(1):66–74, December 2018. ISSN 0001-0782. doi: 10.1145/3198448. URL https://doi.org/10.1145/3198448.

Ray Kurzweil. The Singularity Is Near: When Humans Transcend Biology. Penguin (Non-Classics), 2006. ISBN 0143037889.

Rodney Brooks. Prospects for human level intelligence for humanoid robots. 07 1998. Camilo Miguel Signorelli. Can computers become conscious and overcome humans? Frontiers in Robotics and AI, 5:121, 2018. doi: 10.3389/frobt.2018.00121. URL https://www.frontiersin. org/article/10.3389/frobt.2018.00121.

Roman V. Yampolskiy. Unpredictability of ai: On the impossibility of accurately predicting all actions of a smarter agent. Journal of Artificial Intelligence and Consciousness, 07(01):109–118, 2020. doi: 10.1142/S2705078520500034.

Y.N. Harari. Sapiens: A Brief History of Humankind. Harper, 2015. ISBN 9780062316103. URL https://books.google.co.in/books?id=FmyBAwAAQBAJ.

V. Shiva and K. Shiva. Oneness Vs. the 1 Percent: Shattering Illusions, Seeding Freedom. CHELSEA GREEN PUB, 2020. ISBN 9781645020394. URL https://books.google.co.in/books?
id=4TmTzQEACAAJ.

G. Nicolis and I. Prigogine. Self-Organization in Nonequilibrium Systems: From Dissipative Structures to Order Through Fluctuations. A Wiley-Interscience publication. Wiley, 1977. ISBN 9780471024019. URL https://books.google.co.in/books?id=mZkQAQAAIAAJ.

Shoshana Zuboff. The Age of Surveillance Capitalism: The Fight for a Human Future at the New Frontier of Power. 1st edition, 2018. ISBN 1610395697.

P.K. Dick and J. Lethem. Selected Stories of Philip K. Dick. Houghton Mifflin Harcourt, 2013. ISBN 9780544040540. URL https://books.google.co.in/books?id=V1z9rzfTb2EC.

George Orwell. 1984. Tandem Library, centennial. edition, 1950. ISBN 0881030368. URL http://www.amazon.de/1984-Signet-Classics-George-Orwell/dp/0881030368.