Philosophical Topics

4.1 SCALE: MECHANICALLY INTERLOCKED MOLECULAR ARCHITECTURES

Mechanically Interlocked Molecular Architectures (MIMAs) are a class of large molecular complexes wherein molecules that are not necessarily (covalently) bonded to each other are nonetheless linked due to their topology. Examples include catenanes and rotaxanes, as well as molecular analogues of knots and rings. A catenane is any molecule wherein cyclic chains are linked three dimensionally such as a pair of linked rings. A rotaxane is any molecule with a “ring” and “axel”, i.e. where a closed circular molecule (the ring) wraps a chain of another molecule (the axel) such that it can rotate but will not “fall” off the chain.

Click for larger view
View full resolution

Figure 1.

Rotaxane. The long molecule spanning horizontally is the “axel”; the molecule slightly offset to the right is the “ring”. “Caps” or “dumbbells” on each end prevent the ring from falling off the axel.

MIMAs have applications ranging from fluorescent tagging, contracting molecular “muscles”, and nanoscale “cars” with directional control to logic-gates and molecular data storage devices (Chen et al., 2003; Green et al., 2007; Ma and Tian, 2010; Bruns and Stoddart, 2014). In the latter case, chemical, electromagnetic or optical energy can be used to change the configuration of interlocked molecules, storing information. Indeed, MIMAs like rotaxanes have been heralded as potentially the most dense data-storage medium possible, outstripping conventional memory circuitry by orders of magnitude (Cox 2001).

These components of molecular machines show that function at the nanoscale is not specific to the parts of organisms. With more complex arrangements of MIMAs, the circuitry and circuit-like systems used in AI research can be constructed at the molecular scale. Insofar as sentience requires this scale of living activity, this places molecular machines in a category potentially ripe for proto-cognition and subjective experience, provided they can also operate in a similar context and exhibit a similar level of stochasticity to living systems. Examples of such latter systems come from the field of aqueous and DNA computing, discussed next. [End Page 229]

4.2 SCALE AND CONTEXT: AQUEOUS COMPUTING

Aqueous computing is computation that occurs in water (Head et al., 2002). Two major advantages of this form of computing are its information density and economy of energy (Gibbons et al., 1997). One of the most successful forms of aqueous computing used thus far is DNA computing. DNA computing was conceived by the physicist Mikhail Samoilovich Neiman in his 1960s discussion of polymeric “molecular memory” systems (Brunet, 2016). Thirty years later, the computer scientist Leonard Adleman pioneered experimental DNA computing by showing that he could use it to solve a version of the “traveling salesman problem” (Adleman, 1994). Adleman’s experiments used DNA assembly in vitro to compute the Hamiltonian path of a directed graph—i.e., the path in which each node of the graph is visited exactly once. The power of DNA computing to engage in massively parallel searches with much greater energy efficiency than a supercomputer became apparent thereafter (Adleman, 1994; Gibbons et al., 1997).

As a polymer, DNA has a number of properties that make it a good candidate for molecular computing experiments. Its manipulation, synthesis and sequencing are well understood from past molecular biological research; it can store arbitrary information appropriately encoded in its sequence (Church et al., 2012; Goldman et al., 2013); it can be organised to create fast logic gates and signal transmission lines (Chatterjee et al., 2017); and it can fold into an immense variety of shapes which can themselves bind to neighbouring folded molecules to give rise to even more complex gates (Sha et al., 2005; Seeman, 2007).

Folded DNA structures, sometimes called ‘DNA origami’ (Castro et al., 2011), can take on a rich variety of forms, from images of smiling faces or dolphins to flexible machine-like components and knots. In these systems, folding and computation are intimately related. One way to perform computation with DNA is to synthesize a sequence that folds into the three dimensional structure of Wang tiles (Wang, 1961; Woods et al., 2019). These are a formal system for implementing algorithms where computational processes are converted into problems of matching adjacent faces of a pre-defined and pre-synthesized set of “tiles”, converting computation into molecular origami. Moreover, at sufficient levels of complexity, some such DNA systems are provably Turing complete (Varghese et al., 2015).

These sorts of chemical computational processes do not need to be based in DNA. What is essential to this sort of computation is that one uses a polymer that can engage in sequence-dependent binding to neighbours, which has been achieved with nucleic acid analogues and other more exotic chemicals. For instance, Yamamura et al. (2001) report a method of aqueous computation using peptidyl nucleic acid (PNA)—an analogue of DNA where the usual backbone of DNA is replaced by linked peptides. Regardless of the particular molecular constitution, whether derived from biomolecules or synthesized ab initio, these alternative varieties of computation are embodied at the nanoscale and in [End Page 230] contexts where components are variously dissolved. As Feynman noted in his lecture, matter behaves differently at this scale and context; engineers have had remarkable success harnessing these behaviours to solve problems in new and efficient ways.

4.3 SCALE, CONTEXT AND STOCHASTICITY: MOLECULAR RATCHETS

For any system that must do work under the influence of chaos there are two options: find a way to mediate the effects of this influence, or find a way to turn chaos to your service. Evolution is remarkably good at producing molecular structures deploying the latter strategy, that function by “biasing tendencies” of the molecular storm inside cells at physiological conditions. Molecular ratchets are one important way this is achieved (Oster, 2002; Ait-Haddou and Herzog, 2003). Like the common tool of the same name, ‘ratcheting’ is a general process where the effects or outputs of a system are directional despite the fact that the cause or input is not. Importantly, ratchets do not function despite non-directional inputs, by mediating or reducing them, but by having particular causal processes that take advantage of them.

A prime example of a molecular ratchet is ATP-synthase, a protein complex at the heart of metabolism. In this system a channel between membrane compartments allows the diffusion of hydrogen ions, which is a stochastic or Brownian process. This passage of ions tends to result in a circular rotation of the F0 component of ATP-synthase and drives downstream production of ATP from ADP (Nakamoto et al., 2008; see Wideman et al., 2019). Ratcheting is a way of generating biased tendencies from an unbiased or chaotic external influence, but it is not specific to biomolecules. At the macro-level we can see similar energy generating ratchet processes in self-winding watches, where the ambient movement of the wearer’s body causes an oscillating weight or rotor to swing. These chaotic movements wind a coiled strip of metal (the mainspring), which then has directional (clockwise) downstream effects, providing the watch with a constant supply of mechanical energy. Ratcheting is an essential process guiding the fine-grained stochastic effects of biomolecules, and on Godfrey-Smith’s (2016a) view is important to subjective experience of the sort enjoyed by organisms (see section 3), but it is also a feature of the functional profile of many machines.

At the macro-level again, artificial neural networks (ANNs) are digital ratcheting devices in many ways analogous to biological molecular ratchets. Godfrey-Smith notes that biological causal processes “are based in networks with many redundancies and small effects. These have consequences for robustness and adaptability” (Godfrey-Smith, 2016a, 503). This is also true for ANNs. They can be fed large quantities of fairly disorganized information, which is passed along the network while being “nudged” by weights in directions deemed useful during the largely statistical processes of training. This creates a biased tendency to produce a certain kind of output, despite consisting of a large number of meaningless or dispensable processes. [End Page 231]

It is somewhat unsurprising that ANNs work this way. Both learning and evolutionary processes tend to result in satisfying or “good enough” solutions to complex problems and, in complex systems especially, biasing a tendency is often an easier solution than overcoming or eliminating the chaos involved in existing processes. In the words of the biologist François Jacob (1977), evolution is a tinkerer; it produces novelty by constrained alteration of existing structures. The statistical processes involved in training ANNs are sophisticated tinkering, methodically adjusting parameters until the computational process outputs a desired result to a satisfying degree of accuracy.

ANNs are particularly good examples of ratcheting processes that enable computation despite stochastic inputs, since they are also prima facie candidates for artificially intelligent and sentient systems, provided Godfrey-Smith is correct about the importance of scale, context and stochasticity. Nonetheless, the need to manage stochastic processes is not specific to ANNs and occurs whenever computation takes place. Godfrey-Smith acknowledges that low-level changes in computers are inevitable, but these are “engineered to be as small as possible” and this makes computers “different from living systems in ways that make engineering sense” (Godfrey-Smith, 2016a, p. 503). However, this is not always the route taken in computer engineering, nor always the best method for achieving effective computation, as the examples in this section show.

Instead of being limited or engineered away, chaotic effects are regularly exploited in the field of stochastic computing. For von Neuman (1956), the important point of stochastic computation is the same as for probabilistic logic: that unavoidable error in parts must somehow be accounted for in the production of useful wholes. Indeed, von Neumann is clear that this stochasticity is also a non-accidental part of advanced computer automata, which he and many in his cohort fondly referred to as “organisms”. He writes that error in computer automata is seen, “not as an extraneous and misdirecting accident, but as an essential part of the process under consideration” (von Neumann, 1956, p.43). No computational process—digital, DNA or quantum—produces zero entropy. All entail the production of at least some heat or delay, and the consequent potential for “error” in computation. A large part of computer science deals with how to cope with computational errors due to lower-level stochastic processes. Though the aim— identified by von Neuman—is not simply to limit low-level changes but to recognize and exploit their fundamentality to the system.

As a final example, Brownian computation methods in theoretical computer science have taken the route of turning chaos into an advantage. As the physicist and information theorist Charles H. Bennett writes:

In these [Brownian] computers, a simple assemblage of simple parts determines a low-energy labyrinth isomorphic to the desired computation, through which the system executes its random walk, with a slight drift velocity due to a weak driving force in the direction of forward computation [p. 905] . . . the Brownian computer makes logical state [End Page 232] transitions only as the accidental result of the random thermal jiggling of its information-bearing parts.

(Bennett, 1982, p. 912)

This description of Brownian computation sounds remarkably like Godfrey-Smith’s characterisation of the ratcheting, biased tendencies of metabolism. In order to distinguish the operations of fundamental metabolic components from familiar machines, Godfrey-Smith says that the processes typical of such biomolecules, a “storm-like collection of random walks influenced by friction, charge, and thermal effects’’ are “non-mechanistic” (Godfrey-Smith, 2016a, p.486). However, it is precisely those features of biomolecules that Godfrey-Smith identifies as “nonmechanistic” and essential to the operation of metabolic processes, that are essential to the operation of machines like the Brownian computer. Like Godfrey-Smith, Bennett also notes that, however unlikely such processes are in the macroscopic world, this way of getting things done is common at the level of molecular reactions (Bennett, 1982, p. 912).

More recent work on Brownian computation has continued to draw inspiration and theoretical principles from Bennett and ratcheting in biomolecular motors. Peper et al. (2013) note that “organisms have mechanisms in place, not only to cope with fluctuations, but also to exploit them as a driving force in biological processes” (p. 2), which is the aim of Brownian computation. Peper et al. (2013) and Lee et al. (2016) focus on designing universal Brownian circuitry capable of extracting useful computation from nano-scale fluctuations. These fluctuations are “not just a nuisance” (Lee et al. 2016 p. 342) but “actively exploited” [ibid]. Moreover, to make sure that computation proceeds at an appreciable rate despite inevitable backward processes, speed is increased by “ratcheting devices” that are “well-known in nature” (Peper et al. 2013, p. 19) and work by biasing probabilistic circuit transitions. Importantly, like many biomolecular ratchets, Brownian circuitry is unable to achieve certain effects unless there is sufficient stochasticity in the system.

Complex bio-derivative forms of Brownian computation are also exemplified in DNA computing. What remains to be seen is the achievement of full-scale AI in these alternative mediums. Qian et al. (2011), and later Cherry and Qian (2018), have taken some of the first proof of concept steps: they fabricated neural networks using DNA-based computational techniques. As Qian et al. (2011) point out, “Stochastic simulations suggest that the four-neuron DNA associative memory would function reliably [in] a volume of roughly 1μm³, that is, small enough to fit inside a bacterium” (p. 372). Here is a system that functions stochastically, dissolved in water, at the scale of nanometers, and at the same time has the potential to realize the coarse-grained computational “information processing” operations characteristic of traditional AI projects.

We have reason to believe this sort of AI research will continue; the practical benefits of such technology have been steadily accruing. Chen et al. (2015) review a number of biomedical applications of “dynamic” DNA nanotechnologies, i.e. those, like circuits, with time-dependent and analytic functionality. Moreover, the [End Page 233] prospects raised by some of these technologies—that they might overcome limitations on size and energy-efficiency that currently limit familiar circuitry (Livstone et al. 2003)—suggest that their realization may be driven by demands within AI and robotics. This makes it more likely that these forms of AI will be achieved on a not-too-distant path from current AI projects.

We can look to these unconventional but well-theorized cases of machines for precisely the sorts of fine-grained functions thought relevant for the achievement of sentience. Perhaps our 18th or 19th century complement of machines was impoverished in functionality, but modern examples surpass these in both degree and kind. A DNA-scale computer stochastically checking multiple solutions to a graph-theoretic problem with dissolved components in parallel is about as far from a microprocessor as the latter is from the macroscopic ‘difference engines’ of Babbage or Lovelace. What fine-grained functional or material bases might be lacking in clockwork or contemporary metal-and-silicon computers can be found in these alternative embodiments of artificial intelligence.

If Godfrey-Smith’s view that the metabolic side of life is relevant for sentience is correct, then molecular machines represent promising candidates for artificial sentience. In some cases, these machines exhibit all of the fine-grained functional properties—scale, context, and stochasticity—that Godfrey-Smith identifies as relevant for the emergence of sentience in living systems. We hope the examples discussed here help demonstrate that the answer to the empirical question, “are current AI and robotic systems likely candidates for sentience?” might be “yes.”

5.1 OBJECTION FROM EVOLUTION

One major objection to the view that molecular machines are compelling candidates for sentience is that they lack a critical feature characteristic of living systems: biological organisms depend not only on metabolisms for their continued existence, but also on evolution for their origin. As Godfrey-Smith notes: “Life has a metabolic side and a side that has to do with reproduction and evolution” (Godfrey-Smith, 2016a 484). The molecular machines described above share features with metabolic systems, but are largely products of human engineering, not biological evolution. One might thus argue that, without this evolutionary component, artificial systems are not serious candidates for sentience.

We see three ways of proposing this objection. First, perhaps evolution is necessary for sentience. Second, perhaps evolution is required for sentience in some sense that is strong but nonetheless weaker than nomic or logical necessity; perhaps it is extremely unlikely for humans to engineer systems with functions analogous to those that, in organisms, required billions of years of evolution to arise. Third, one might argue that molecular machines lack the foundational materials [End Page 234] for evolving sentience, that even if they were subjected to evolutionary forces, they would lack what is needed to evolve proto-cognition or minimal subjectivity.

In response to the first two arguments, some molecular machines indeed were and are subjected to evolutionary processes. The molecular machines designed by synthetic biologists have a mixed history, like dairy cows, owing their existence in part to artificial and in part to natural forces (Lewens 2013). If merely having some evolutionary component is required then DNA computers are safe from objections on the basis of past evolution. DNA itself and the enzymes used in computational protocols are, in part, outcomes of evolution.

If it is rather present evolvability that is required, we can look to the algorithms and experimental processes analogous to evolution that engineers use, for example, for solving optimization problems and discovering new molecular structures (Hu & Banzhaf, 2010). In a recent study, researchers used a genetic algorithm, machine learning, and in vitro analysis to explore the space of possible antimicrobial peptides or AMPs (Yoshida et al., 2018). AMPs are short (10–50 amino acids) peptides found in almost all living organisms, which kill microbes by breaking bacterial membranes, inhibiting DNA, RNA, and protein synthesis, and other methods. Researchers have long been exploring ways of designing new peptides for clinical use (Loose et al., 2006). The technique developed by Yoshida and colleagues shows how one can artificially evolve functional biomolecules by randomly generating a population of peptides and evaluating the “fitness” of individuals in silico and in vitro. A subset of this population is then selected, forming the basis of a new population in which mutations and crossovers are introduced (Yoshida et al., 2018).

Although the rates of biological and artificial evolution are difficult to compare, much work is dedicated to increasing the speed of artificial evolution for the benefit of practical applications (Hu & Banzhaf, 2010). Moreover, artificial evolution can be combined with methods like deep learning and reinforcement learning to efficiently develop new systems with surprising capacities. The program AlphaStar—the first system to defeat top human professionals at the real-time strategy game StarCraft—was developed in this way (Arulkumaran et al. 2019). If sentience depends on cumulative evolution (in either the stronger or weaker senses above), this is unlikely to place it beyond the reach of our current engineering efforts.

Regarding the third argument, we think it is important to be clear about the difference between gradualist and saltationist accounts of the emergence of fully-fledged consciousness. Our response is inspired by Samuel Butler, a contemporary of Charles Darwin. Butler proposes the descent of conscious machines as a disjunction elimination from the assumption that consciousness either emerges gradually or saltationally.

If consciousness is seen as saltational—occurring in a leap or “definite step”— then we are left with the mystery of explaining that leap. However, we know that such a leap is possible. As Butler put it, if consciousness emerges by a leap, then [End Page 235] “the race of man has descended from things which had no consciousness at all” (Butler, 1872, p. 197). That is, if consciousness evolved saltationally, then something need not have any consciousness at all for it to evolve consciousness. In this case, current molecular machines might be strong candidates for consciousness without themselves having any consciousness at all. At least, so far as we know, given that we have no satisfying explanation for that leap in either machines or organisms.

If instead consciousness emerges gradually, Butler proposes, then much of the “action that has been called purely mechanical and unconscious”, such as that of our distant and not obviously conscious ancestors, must have contained “elements of consciousness” (Butler, 1872, p. 197). If so, we are left without a reason that other “purely mechanical” actions might not have elements of consciousness. Butler writes: “There is no security . . . against the ultimate development of mechanical consciousness, in the fact of machines possessing little consciousness now. A mollusk has not much consciousness” (Butler, 1872, p. 194). Godfrey-Smith and Butler share the position that consciousness can be found in diverse entities and that it either evolves gradually or exists widely in embryo or proto-cognitively. We imagine Godfrey-Smith largely agreeing with Butler about molluscs and other metabolizing cases, while denying that contemporary machines possess even a “little consciousness” now.

This claim about machines arises not because Godfrey-Smith denies the possibility of a system having a little consciousness. Godfrey-Smith believes one can have more or less consciousness and does not see the origin of consciousness as a “definite step” (Godfrey-Smith, 2016b, p. 68); he supports the view that subjective experience can come in “minimal but non-zero” proportions (Godfrey-Smith, 2016a, p. 495). Of course the notion of increments of consciousness or minimal degrees of subjectivity is not uncontentious (Shevlin 2021). However, gradualist evolutionary proposals of the sort endorsed by Godfrey-Smith and Butler require some such notion to get off the ground. On the assumption of gradualism, whether or not machines can evolve consciousness turns on whether they now possess the right foundations or building blocks for consciousness.

Butler’s analysis of consciousness in machines is motivated by epistemic considerations and concerns of risk. At a sufficiently proto-level, consciousness might be too subtle to notice, and thus to avoid or provide “security against” (Butler, 1872, p. 194). This leads him to attribute “germs of consciousness” to “many actions of the higher machines” (ibid., p. 197), as he would to molluscs. However, without some principled way of assessing “germs of consciousness” in both higher machines and molluscs, Butler provides little guidance for future analysis or standards for evaluating his own account.

Happily, Godfrey-Smith provides an account of what is required for minimal forms of subjectivity, and thus permits a deeper assessment of putatively similar candidates for the gradual evolution of consciousness. Godfrey-Smith is partial to biopsychism: the view that low-level conscious capacities are exhibited not by ordinary [End Page 236] matter (classical panpsychism) or mechanical activities, but by “the simplest forms of cellular life” (Godfrey-Smith, 2016a, p. 495). Presumably, what is relevant about simple and evolutionarily early forms of life are just those features of “living activity” thought relevant for sentience in organisms today. If we accept this framing, then the question of low-level conscious capacities in current machines becomes whether we can find “living activities” responsible for the “actions of higher machines”, as we would in ancient life, and our arguments in section 4 show that we can. If indeed these are the right fine-grained features of living activity required as a foundation or for the building blocks of minimal consciousness, then gradualist narratives for the origin of machine consciousness can get off the ground.

Butler, Godfrey-Smith, Canguilhem and Hacking would all agree that any inference to the descent of conscious machines from the present stock should be an a posteriori affair. Indeed, to Canguilhem’s or Hacking’s bare bones wait-andsee approach mentioned in section 3, Godfrey-Smith adds flesh in the form of the metabolic criteria required for living activity in simple cellular life.

5.2 OBJECTION FROM LIFE

Finally, one might object to molecular machines as candidates for sentience on the grounds that they are not alive.⁵ Perhaps something other than the metabolic and evolutionary side of life is required for sentience—some property of living systems not highlighted in Godfrey-Smith’s account. This is a reasonable objection, given the nascent state of research on subjectivity. Godfrey-Smith weaves a rich tapestry of dependencies, linking the scale, context and stochasticity of molecular metabolisms to simple and complex forms of cognition. However, he is also quick to point out that the story provided is a preliminary sketch. Such a sketch requires revising and filling in, guided by our growing theoretical and empirical understanding of the relevant systems and processes. Given this, it is reasonable to expect new features relevant to sentience to emerge, and to reassess the question of artificial sentience in light of this research. We agree with this approach and offer our analysis of artificial sentience as one based on the current state of the art of research on the biological and evolutionary conditions for sentience.

We would like to emphasise, however, that if one objects to our claim that AI is a candidate for sentience on the basis that it lacks some other property of life, then one should also specify what this property is and why it is relevant. It will not suffice to say that these machines are not alive. The concept of life is ambiguous and can be filled out in different ways. One could thus appeal to this concept to get a desired outcome in an ad hoc manner. If one wanted to exclude AI from the class [End Page 237] of sentient beings, for example, one could select precisely those living properties— particular aspects of growth or reproduction, for instance—that current AI lacks. Such an account would resemble what Deborah Mayo calls a “use-constructed” or “rigged” hypothesis or one constructed on the basis of known data, such as a marksman hypothesizing that she is a good shot by shooting holes in a board before proceeding to draw a target around those holes so that a bullseye is scored (Mayo 1996, p. 201). This method of theory construction should be avoided, as it cannot fail to fit the data, even if it were false.

Instead we think one should follow Godfrey-Smith’s lead and specify those fine-grained functions and broader evolutionary processes that we have reason to believe resulted in systems’ sentience. Once these are specified, they can be applied to other systems, as we have done here. We should however also keep in mind what Shevlin (2021) refers to as the “specificity problem” or “the problem of determining the appropriate level of detail that should be adopted by theories of consciousness in applying them to non-humans” (p. 298; see also Birch 2020). The more finely we spell out the criteria for sentience, the more committed we are to excluding systems that differ from those used to build our account (Sprevak, 2009). Explaining why a particular feature or function is relevant for sentience helps avoid this problem to some degree. In providing such explanations, one justifies the inclusion of a feature or process (at some given level of detail) in one’s account. Godfrey-Smith provides a compelling case that scale, context and stochasticity are relevant to the evolution of sentience. Any additional properties—whether features of life or not—should be similarly justified.