Invertebrate Sentience: A Useful Empirical Resource

Executive Summary

Rethink Priorities reviewed the scientific literature relevant to invertebrate sentience. We selected 53 features potentially indicative of the capacity for valenced experience and examined the degree to which these features are found throughout 18 representative biological taxa. These data have been compiled into an easily sortable database that will enable animal welfare organizations to better gauge the probability that (various species of) invertebrates have the capacity for valenced experience. This essay details what we’ve done, why, and the strengths and weaknesses of our approach.

Project Overview

This post is the second in our series on invertebrate^[1] welfare. In the first post we examine some philosophical difficulties inherent in the detection of morally significant pain and pleasure in nonhumans. In the third, fourth, and fifth posts we explain in detail the features we believe to be most relevant for assessing invertebrate sentience. In the sixth, seventh, and eighth posts, we present our summary of findings, both in narrative form and as an interactive database. In forthcoming work (to be published late July), we analyze the extent to which invertebrate welfare is a promising cause area.

We focus on invertebrates for two reasons: (1) We are already reasonably confident that mammals, birds, reptiles, amphibians, and most fish^[2] feel morally significant pain and pleasure, and hence must be included in our moral calculations, but we are unsure if more distantly related animals warrant similar concern, and (2) The subject of invertebrate welfare, though recently gaining traction both in the scientific literature and the effective altruism community, appears neglected relative to the sheer number^[3] of potentially suffering invertebrates.

To develop accurate cost-benefit models that can be used to allocate resources across the animal welfare movement, we need to take the possibility of invertebrate pain and pleasure seriously. But determining whether invertebrates have the capacity to experience pain and pleasure in a morally significant way is an extraordinarily complex and difficult undertaking. There is tremendous uncertainty at virtually every level at which one might investigate the matter. We don’t expect to conclude with high confidence that invertebrates do or do not experience morally significant pain and pleasure. Such an outcome is too ambitious. Rather, our goal is to clearly map out the problem so that we can begin to systematically reduce key uncertainties in a cost-effective manner.

One tractable way to improve our credences with respect to invertebrate welfare is to create a comprehensive collection and analysis of the extant scientific studies relevant to the subject. While some scientific studies directly address the issue of animal pain, the vast majority of studies that are relevant at all to invertebrate welfare are relevant only tangentially. For example, studies on cockroach navigation and place memory don’t directly address the issue of invertebrate welfare. Nonetheless, the ability to accomplish certain navigational feats might be decent evidence that creatures with that ability are conscious, which is itself a necessary condition on experiencing pain and pleasure. However, a search for “invertebrate welfare” on Google Scholar doesn’t deliver any studies on cockroach navigation. Gathering all the relevant scientific literature in one place is thus a nontrivial and, to-date, unaccomplished task.

Rethink Priorities has spent the last ten months completing just such a task. We have analyzed the degree to which more than 50 features potentially indicative of phenomenal consciousness^[4] are found throughout 18 representative biological taxa. For each pair of taxa (ants, say) and features (protective behavior, say), we first determined whether there was sufficient scientific data to make a call as to whether the taxon in question possesses the feature. If so, we evaluated the likelihood that the taxon possesses the feature. The studies we consulted are linked to each cell, with accompanying clarificatory quotes and/or commentary. These data have been compiled into a sortable, user-friendly database.

Project Rationale

In the remainder of this post, we aim to explain four things: (a) why we built the database, (b) how we chose the features that we chose, (c) how we chose the taxa that we chose, and (d) what the limitations of the database are.

The project was inspired in part by the work of Luke Muehlhauser at the Open Philanthropy Project. In his comprehensive 2017 Report on Consciousness and Moral Patienthood, Muehlhauser briefly describes the potential usefulness of a large reference work that collects scientific data pertaining to consciousness from disparate fields into one place. This suggestion was the causally proximate impetus for the current project.

The project is important because it will allow grantmakers, charity entrepreneurs, animal activists, and legislators to make better judgments about which animals have the capacity for valenced experience.^[5] Although other researchers have occasionally tried to gauge the distribution of consciousness throughout the animal kingdom, the project is unique in both breadth of taxa examined and the number of features investigated. Additionally, the project is open to continuous improvement. It is designed to allow both for the incorporation of new scientific evidence as that evidence becomes available and for the inclusion of additional biological taxa (or even non-biological entities) should such additions be deemed useful.

Argument by Analogy and Inference to the Best Explanation

The project can be used to help facilitate two important arguments often deployed to investigate consciousness in nonhuman animals. The first is argument by analogy. The basic structure of an analogical argument is as follows (where E1 is the source domain and E2 is the target domain):

(1) Entity E1 has some properties P1 … Pn
(2) Entity E2 has the same properties P1 … Pn
(3) Entity E1 has some further property Pn+1
∴
(4) Therefore, entity E2 likely has the same property Pn+1

Our database can aid in the construction of analogical arguments. First choose some taxon from the database to serve as your source domain. This should be a creature you are reasonably confident is conscious (cows, say). Then select some features that the source domain taxon possesses that you think are relevant to consciousness. Next select some taxon to serve as your target domain. This should be a creature for which you are reasonably uncertain whether it is conscious (octopuses, say). Finally, check to see how many of the source domain taxon features are shared by the target domain taxon. This sort of analogical reasoning should also be run in reverse. Start with some taxon you are reasonably certain is not conscious (plants, say), then compare the feature profile of this taxon with your original target domain taxon (in this example, octopuses).

The other main type of argument for which our database could be useful is called inference to the best explanation. Inference to the best explanation is related to argument by analogy, but for complicated theoretical reasons, it is often more powerful.^[6] To see how inference to the best explanation works, consider how we know our fellow humans are conscious. When I cut my hand, I cry out, I move my hand away from the sharp object, and I later treat the wound with a clean bandage. When other humans cut their hands, they also cry out and attend to the subsequent wounds in similar ways. There are a variety of hypotheses which, if true, could explain this behavior. Perhaps they are sophisticated robots programmed to behave as I do. But the simplest and best explanation of the behavior of other humans is that they feel pain like I do.^[7]

Of course, this explanation might be mistaken, and we might come to know it is mistaken. If I examined the heads of many fellow humans and in each case found not a brain but a crude artificial device receiving signals from a robotics factory, that would constitute a defeater for my prior explanation. I would then no longer be able to rationally endorse the view that other humans have mental states like I do. Inference to the best explanation tells us that, in the absence of defeaters, we are licensed to prefer the simplest explanation of a phenomenon.^[8]

Inference to the best explanation can also be applied to nonhuman animals. First one ought to determine which combination of features correlates to the behavioral patterns that, in conscious creatures like humans, are typically caused by conscious mental states like pain and pleasure. Then see which taxa possess this combination of features (taking care, of course, to allow for the behavioral patterns to be expressed in different ways by different taxa). Next, check for defeaters. Defeaters might come in any number of guises. For example, one might think that organisms below a certain neurological complexity (measured by brain size or neuron count) are incapable of conscious states. One might think that failure to alter behavior in order to avoid injurious stimuli is a telltale sign that the organism in question doesn’t feel pain. One might think that nociceptors (specialized cells used to detect potentially harmful events) are necessary for pain sensation, in which case organisms that lack nociceptors are unable to experience pain. No matter one’s views, having all the relevant data easily manipulable and accessible in one place will make adjusting one’s credences about the distribution of consciousness easier and better informed.

Which Features?

Our goal at the outset was to create a database with the most comprehensive set of features potentially relevant to valenced experience yet published. Even so, we had to make some hard choices to keep the list in the manageable range of approximately 50-60 features. In this section we attempt to justify the composition of the list. To simplify the matter, we discuss the list in terms of eight broad categories: anatomical features, noxious stimuli reactions, learning indicators, cognitive sophistication, motivational tradeoffs, mood state behaviors, drug responses, and navigational skills. Before considering these categories in detail, it’s helpful to first look at the current state of the literature to see which features other similar projects have investigated.

The State of the Literature

The most influential recent discussion of nonhuman pain in the literature is Sneddon et al.’s 2014 “Defining and Assessing Animal Pain.”^[9] The authors present 17 features potentially relevant to pain perception, then discuss the extent to which these features are found across 7 very broad taxa (see Table 1 below.)

Table 1, recreated from Sneddon et al. 2014

To our knowledge, this is the most comprehensive, authoritative table in the literature. Nonetheless, the Sneddon table is limited in several ways. One way issue is the breadth of its seven taxa: mammalia, aves, amphibia/reptilia, agnatha/osteichthyes, cephalopoda, decapods, and insecta. According to the paper’s methodology, so long as at least one species in the taxon exhibits the feature in question, the relevant cell receives a checkmark. This methodology is problematic for two reasons. First, the taxa considered are enormous. The last taxon, insecta, contains more than six million species. A single species isn’t possibly representative of the whole taxon. More importantly, this methodology can produce a misleading table. By Sneddon et al.’s own lights, we should only think a species experiences pain if they exhibit all or most of the features. But if 17 different species within a given taxon each exhibit a single different feature, the column would be full of checkmarks, giving the misleading impression that the members of the taxon generally satisfy the presented definition of animal pain.

More broadly, we find that 17 features is still not enough data points to give a relatively theory-neutral picture of animal consciousness. Every investigation, including our own, has to make difficult decisions about which features to examine, and these decisions will always be informed, to a certain extent, by theoretical considerations about the nature of consciousness. (Nobody is going to include number of appendages as a feature to be examined because no plausible theory of consciousness holds that appendage number is relevant to consciousness.) In general, though, the fewer features included in an investigation, the more theory that has to go into deciding which features are important. The more features an investigation includes, the fewer theoretical decisions have to be made in advance. We aimed to outdo the Sneddon table both in specificity of taxa and number of features to be investigated.

We turn, then, to the features we included in our own investigation and their (broad) justification.

Anatomical Features

There are several anatomical features which may be relevant to the capacity for valenced experience. Nociceptors are specialized peripheral sensory cells used by the body to detect potentially harmful stimuli. We investigated which of our targeted taxa possess nociceptors, including both a strict definition of nociception and a loose definition of nociception. Other anatomical features are neurological. We know that, in humans, pain and pleasure are processed in the brain, so we included brain size and brain neurons as features to be investigated. The presence of opioid-like receptors is potentially important because opioids are analgesics (painkillers). Opioids are especially significant because, unlike other analgesics, opioids can reduce the affective (valenced) component of pain without a corresponding reduction in the sensory aspect (e.g., the burning/throbbing/cutting/stinging/aching quality) of pain.^[10] Although opioid receptors play many roles, their presence is at least mild evidence that the creatures which possess them have the capacity for valenced experience. We included average lifespan for each taxon because shorter-lived creatures in general encounter fewer novel situations and thus may be able to survive and reproduce utilizing only innate, pre-programmed behaviors.

Noxious Stimuli Reactions

One obvious strategy for detecting valenced experience in nonhuman animals is to see how they respond to various potentially positive and negative stimuli.^[11] There are far more scientific data relevant to potential animal pain than to pleasure, so for this broad category of features we focused on reactions to negative (hereafter “noxious,” in keeping with the literature) stimuli. If a creature doesn’t seem to react at all to a noxious event (such as getting eaten by another creature), that is evidence that the creature in question doesn’t experience the event as painful. However, as we’ll see, just because a creature does react (by squirming away, say) doesn’t mean that the event is necessarily experienced as painful.

One evolutionary function of conscious pain experiences—perhaps the evolutionary function of conscious pain experiences—is to promote long-term bodily integrity. Although not all damaging injuries are painful and not all pain experiences result from damaging injuries, there is nonetheless a fairly tight correlation between tissue damage and pain sensation. When I prick my hand on a rose bush, it hurts. I don’t like the feeling, so I learn to be more careful around rose bushes in the future. In this way I protect the integrity of my hand.

However, there is a different system the body also utilizes to avoid tissue damage, a system that operates below conscious awareness: nociception. Nociception is widely conserved throughout the animal kingdom. Even relatively simple creatures, like the roundworm C. elegans, possess nociceptors. Nociception enables the detection and avoidance of potentially noxious stimuli, and it does not have an attendant phenomenology with whatever evolutionary cost such a phenomenology incurs. So why would a creature opt for both a nociceptive system and conscious pain experiences?

The answer may be long-term memory. Nociception is fast and reflexive, but it is not normally associated with a long-lasting memory. On the other hand, pain experiences, with their attendant felt badness, tend to leave a memorial imprint. Because pain experiences are often stored in long-term memory, pain often induces long-term behavioral and motivational changes. For long-lived animals in complex environments, pain is thus potentially more effective at protecting the animal from damage than mere nociception.

What pain and nociception have in common is that they are triggered, in the normal case, by noxious stimuli. We included several features in our database which give evidence that the creatures who possess them are aware of noxious events. Movement away from noxious stimuli is a very basic feature, found, we discovered, even among some plants. Protective behavior (e.g., limping, wound guarding, wound rubbing, wound licking) is evidence that an animal knows it has been injured. Noxious stimuli related vocalization (e.g., screaming, crying, groaning) is sometimes taken to be a good metric of farmed animal welfare, and so we investigated the extent to which it is present elsewhere in the biological world. But the best evidence that nociception leads to conscious pain experience is not that an animal reacts to noxious stimuli but rather that these reactions lead to behavior alteration—that is, the creature learns from these encounters.

Learning Indicators

Some short-lived animals are so well-adapted to their evolutionary and ecological niche that nociceptive responses are sufficient to keep the animal safe from harm. Animals in more complex situations (either ecologically or socially) need more plastic responses, which requires learning from the environment and/or other animals. According to some researchers, behavioral plasticity is a hallmark of consciousness. For example, Paula Droege, a philosopher at Penn State, argues that “animals capable of flexible responses are conscious animals, animals that feel conscious pain.”^[12] For this reason we looked at a number of learning indicators, including classical conditioning, operant conditioning, and contextual association. Most importantly, we looked to see if exposure to potentially painful experiences led to long-term behavior alteration.

Learning indicators are potentially relevant in another way. On one common understanding of consciousness, a conscious mental state is just a mental state one is aware of being in. Being aware of one’s own mental states would plausibly help a creature understand the mental states of others. Understanding the mental states of others plausibly aids in the project of social coordination. Thus, the evolutionary role of consciousness might be to help creatures interact, cooperate, and communicate in more advanced and adaptive ways. For that reason we also investigated the extent to which different animals engaged in observational or social learning. Such behavior is also evidence of cognitive sophistication.

Cognitive Sophistication

There are a number of ways in which evidence of cognitive sophistication might also be evidence for the capacity for valenced experience. At the most general level, there appears to be a correlation between the capacity for valenced experience and cognitive sophistication. Humans have the capacity for valenced experience. Humans are the most cognitively sophisticated animals on the planet. Other cognitively sophisticated animals, like apes and cetaceans, are also often thought to possess the capacity for valenced experience.

Additionally, some theories of consciousness put prerequisites on conscious experience which seem able to be satisfied only by relatively sophisticated creatures. According to higher-order theories of consciousness, conscious experience requires meta-mental self-awareness. To have a conscious desire for pineapple, say, it is not enough that one desires pineapple. (Some desires, as Freud taught us, are unconscious.) One must also possess a higher-order mental state that represents, in some form, one’s desire for pineapple. If consciousness indeed functions in this way, then cognitive sophistication of a certain degree may be a necessary condition on conscious experience.^[13]

Elements of cognitive sophistication can also be more direct evidence for conscious experience. Genuine deception, for instance, is a complex cognitive and social skill.^[14] To deceive, one must be capable of appreciating the distinction between pretense and reality. One must also be capable of representing, to some extent, the mental states of the deception target. To deceive one must have a cognitive grasp on how things appear, both to oneself and to one’s target. To be aware of appearances, one must experience them. A complex form of behavior in which an animal uses appearances to deceive other animals is thus potentially fairly direct evidence that the animal is conscious.^[15]

Another example: tool use is often considered a benchmark of cognitive sophistication. Genuine tool use requires the acquisition of a foreign object to be utilized at some later time.^[16] In most cases of tool use, the acquisition or transportation of the tool inflicts some cost on the animal, for which the animal is compensated by some benefit when the tool is later deployed. Tool use therefore requires a certain degree of foresight and planning, two elements of cognitive sophistication.^[17] On many understandings of consciousness, such abilities are difficult to imagine without a central, unified mental space from which to evaluate competing interests and demands. Such considerations bring us to motivational tradeoffs.

Motivational Tradeoffs

This category of features is characterized by behavior that indicates the creature in question is capable of weighing competing interests and demands. For instance, when deciding whether to enter a particular foraging site, an animal might weigh the threat of predation against the need to find food to survive. Similarly, when confronted with a competitor, an animal might weigh the benefits of maintaining possession over a particular territory against the danger of suffering physical harm in a fight. These types of tradeoffs indicate the creature has some sort of underlying unified utility function capable of incorporating different streams of information. Such integration is often considered a distinctive attribute of consciousness.^[18]

Such tradeoffs can be replicated, controlled, and monitored in the lab. Many animals, including at least one species of invertebrate, are willing to pay a cost to receive a reward. For example, fruit flies will endure electric shock in order to attain the cue associated with ethanol, indicating that they are prepared to tolerate punishment to obtain the drug. Conversely, many animals, even some invertebrates, are willing to pay a cost to avoid a noxious stimulus. Shore crabs generally avoid well-lit areas, preferring to hide in dark environments such as those found under rocks. Given the choice between two chambers in a laboratory setting, one brightly lit and the other dark, the crabs will universally choose the dark chamber. However, if the dark chamber is rigged to deliver a mild shock, the crabs will begin to opt for the normally avoided well-lit chamber. The crabs do so in increasing numbers (and increasingly quickly) as the number of trials increases. A compelling explanation of this behavior is that the crabs feel pain, then learn to avoid the pain by choosing the opposite, otherwise undesirable chamber.

Some motivational tradeoffs take the form of **predator avoidance **strategies. For example, in a laboratory setting, hermit crabs will abandon their shells if they are subjected to a mild shock. Initially, such behavior was thought to be purely reflexive. However, recent experiments show the crabs are significantly less likely to abandon their shells after shock if the odor of a predator is present. The fact that the crabs remain in their shells when the odor of a predator is present suggests that the behavior is not reflexive. A natural explanation is that the crabs weigh the pain of the shock against the fear of a predator. But to truly understand whether invertebrates can experience fear (and other negatively valenced emotional states), we investigated a variety of mood state behaviors.

Mood State Behaviors

Just as conscious pain is plausibly correlated with a certain behavioral profile, so, too, should we expect various emotional states to generate behavioral indicators. We examined the extent to which stereotypic behavior and displacement behavior, two common indicators of welfare problems among captive vertebrates, can be found in invertebrates. Stereotypic behavior is a type of repetitive activity that is unvarying and apparently purposeless. Examples include excessive grooming leading to self-mutilation, swimming in circles, pacing, mouthing cage bars, and chewing without anything in the mouth (so-called “sham-chewing”). Stereotypic behavior is probably caused by impoverished states, especially situations where the normal functional response of an animal is blocked. Displacement behavior is another response to stress. Common examples of displacement behavior include pacing, fidgeting, grooming, scratching, and object manipulation. Displacement activities are thought to occur when an organism is caught between conflicting motivations. (For instance, a bird unable to decide whether it ought to flee or attack an opponent might instead quickly preen or peck the ground.) Notably, displacement behavior is increased in a dose-dependent way when an organism is given anxiety-provoking drugs and decreased in a dose-dependent way when an organism is given anxiety-relieving drugs.^[19]

We also looked at fear-like and anxiety-like behavior in invertebrates. One example: stress-induced avoidance behavior in crayfish bears a striking resemblance to mammalian anxiety. A 2014 study demonstrated that shocked crayfish develop an extended, context-independent aversion to light.^[20] (The shocks were not associated with levels of illumination.) In contrast, unshocked crayfish, though preferring the dark, were happy to explore both illuminated and unilluminated areas of their environment. In humans, anxiety is often associated with danger that is perceived to be unavoidable or situations in which the threat is ambiguous or unknown. The electric shocks applied to the crayfish fit this description. In humans, anxiety is associated with generalized fear, that is, increased fear of unrelated stimuli. The shocked crayfish appeared to exhibit increased fear of light that is unrelated to the source of stress. Most importantly, injecting the crayfish with the anxiolytic drug chlordiazepoxide (used to treat anxiety in humans) eliminated the aversion to light.^[21] The fact that anxiety-like behavior in crayfish is reduced by a drug used to treat anxiety in humans significantly bolsters the case that crayfish experience negatively valenced emotional states. Thus, looking at drug responses may help us determine which animals are capable of valenced experience.

Drug Responses

Some drugs alter the valence of experience in humans. Analgesics and anxiolytics, for example, are used to treat negatively valenced experiences, like pain and anxiety. Many recreational drugs, including alcohol, normally induce positively valenced experiences. If when exposed to these drugs nonhuman animals exhibit behavior relevantly similar to the behavior of humans exposed to the same drugs, then, in the absence of defeaters, we have some evidence that those nonhuman animals undergo similarly valenced experiences. Thus, we investigated which organisms are affected by analgesics in a manner similar to humans. We were particularly interested in studies that tested whether an animal will self-administer analgesics becauses these studies can help us distinguish cases in which an analgesic genuinely reduces pain sensation from cases in which an analgesic merely diminishes response to stimuli in general.

We looked at studies which tested various recreational drugs and antidepressants on nonhuman animals and tried to discern which animals exhibited relevantly similar behavior (including self-administration). A wide variety of recreational drugs have been tested on invertebrates, including amphetamine in crayfish, marijuana and caffeine in spiders, and MDMA in octopuses. One study on the effects of cocaine in honey bees found that bees quickly become addicted to the drug despite the fact that it severely damages their coordination, locomotion, and motor systems. This is a significant finding, given the importance of navigational abilities to foraging honey bees. Indeed, the impressive navigational skills of honey bees is one of the main reasons to think they might be conscious.

Navigational Skills

To survive, animals need to avoid environmental hazards and predators. They need to locate food, water, and mates. Many animals need to return to fixed sites (such as dens, hives, breeding grounds, food caches, watering holes, foraging sites, or nesting beaches) multiple times over the course of their lives. To do so requires navigational skills.

Navigational skills are relevant for the detection of valenced experience in virtue of their possible connection to conscious experience more generally. According to some researchers, the evolutionary function of consciousness is to produce an integrated and egocentric spatial model to guide an animal as it navigates a complex environment. If consciousness evolved specifically to aid motile animals in complex environments, then various navigational feats could plausibly be interpreted as evidence of consciousness in evolved creatures.

Peter Carruthers, a philosopher and cognitive scientist at the University of Maryland, writes, “at least some invertebrates (specifically honey bees and jumping spiders) possess a belief-desire-planning cognitive architecture much like our own, as revealed by their sophisticated navigation abilities” (282). Moreover, sophisticated navigational abilities, especially those requiring detour behavior, are evidence of cognitive complexity, which, as was discussed above, is often taken to be relevant to consciousness.^[22] Spatial memory is potentially relevant to consciousness because of its relationship to what Michael Trestman, a philosopher of biology at Indiana University, calls a “complex active body” (CAB). A CAB is capable of independent, perceptually-guided, powered-motion (e.g., swimming, flying, crawling). According to Trestman, the evolution of CABs requires the capacity for integrated, embodied spatial cognition, including spatial memory. If one adopts a global workspace theory of consciousness, a sophisticated, integrated form of cognition (of which spatial memory forms a part) of this sort is evidence for consciousness.

Which Taxa?

To determine which taxa to include in our database, we faced two pertinent questions: (1) Which parts of the biological world would help us understand invertebrate sentience the best? and (2) At which taxonomic rank ought we to investigate those parts of the biological world? I’ll consider these questions in turn.

Which Lifeforms?

Ultimately, we want to know whether invertebrate welfare is a cause area worth prioritizing within the animal welfare movement. In order to make that determination, we need to be able to appropriately estimate the likelihood that invertebrates are capable of valenced experience. To get a handle on that likelihood, it’s not enough to merely investigate invertebrates.

Invertebrates represent (at least as far as we are concerned) an edge case of consciousness. There are plausible reasons to think that conscious experience is confined to a relatively small corner of evolutionary space (which excludes invertebrates), but there are also plausible reasons to think that conscious experience evolved very early and is conserved even among most invertebrate phyla.^[23] When investigating an edge case, it’s a good idea to compare the edge case to clearer cases on either side. We wanted a database with the breadth of taxa required to do that for consciousness. To that end we included some familiar vertebrate species: cows (Bos taurus), chickens (Gallus gallus domesticus), and of course humans. These species give a sense for how common the features we investigated are in creatures widely considered to be conscious. We also included plants (kingdom Plantae), protists, and prokaryotes to illustrate how frequently these features are found in entities widely believed to lack conscious awareness.

Including both positive and negative examples will help identify which features are relevant for detecting invertebrate consciousness. For instance, there is some evidence that plants have the capacity for associative learning. If that’s true, one might discount the relevance of associative learning in invertebrates, which is widespread.^[24] On the other hand, even cows fail the classic mirror test for self-recognition. With that knowledge in mind, one might think the fact that no invertebrate has ever passed the test isn’t especially important for assessing whether invertebrates are conscious.

When it came time to decide which invertebrates to investigate, there were several factors that went into our decision: (a) how much scientific data is available for a given type of animal?, (b) how numerous is a given type of animal?, (c) how likely is it that we could find actionable interventions to prevent suffering in a given type of animal?, (d) how prima facie likely do we think it is that a given type of animal is conscious?

We chose to investigate several taxa that include model organisms from the invertebrate world, including the common fruit fly, Drosophila melanogaster; the California sea hare, Aplysia californica; and the nematode Caenorhabditis elegans. These species afforded us a wealth of scientific data to analyze. We investigated ants (family Formicidae) in part because they are so numerous, constituting 15-20% of terrestrial animal biomass. We investigated honey bees (genus Apis) in part because they are extensively exploited by humans.^[25] We investigated two decapod crustaceans, crabs (infraorder Brachyura) and crayfish (family Cambaridae), for similar reasons. We investigated octopuses (family Octopodidae) because octopuses are widely considered the most cognitively sophisticated invertebrate. We wanted representatives from other common invertebrate phyla, so we investigated earthworms (Lumbricus terrestris) from the phylum Annelida and moon jellyfish (genus Aurelia) from the phylum Cnidaria. To round out the invertebrate section of the database, we also investigated cockroaches (genus Periplaneta) and spiders (order Araneae).

One final ‘taxon’ merits special discussion. It is an empirical question which behaviors are caused or accompanied by conscious experience and which are not. Even for sentient creatures such as humans, a great many behaviors, such as the patellar reflex, will be mediated solely by unconscious processes. If a behavior can be performed unconsciously by humans, that’s some evidence that the same behavior does not operate consciously in nonhuman animals. Distinguishing which features can be performed unconsciously (and to what extent) from those behaviors which are always performed consciously is not easy. Thus, we decided to include a column in our database dedicated to answering the question, for every feature, ‘can this behavior be performed unconsciously in humans?’ For each cell in this column we summarize the best available scientific evidence for the feature in question, often with mixed results. This column may make determining the relative importance of the features easier.

What Taxonomic Rank?

After deciding which types of animals we wanted to investigate, we next needed to decide the taxonomic level at which we would investigate those animals. There are competing considerations at play in this decision. On the one hand, there is pressure to drill down to a fairly narrow taxon (genus or species, say). The higher up the taxonomic hierarchy one goes, the more diverse a taxon becomes. If a taxon becomes too large, then saying that the taxon possesses some feature ceases to be informative. If 50 different arthropods each possess one (and only one) of the features we investigated, and each species possesses a different feature, a database with the category “arthropod” would give the misleading impression that arthropods definitely have the capacity for valenced experience.

On the other hand, there is also pressure to examine each type of animal at the highest taxonomic rank that remains informative. To a certain extent, we want our results to generalize. There is little point in learning that a particular subspecies of jumping spider experiences pain and pleasure if we have no idea what this fact tells us about other arachnids. Another, more practical, source of upward pressure was the amount of research which had been conducted at each taxonomic rank. The higher up the taxonomic hierarchy one goes, the more research becomes available for us to analyze. For all but the most commonly studied invertebrates (e.g., Drosophila melanogaster and Caenorhabditis elegans), it would have been difficult for us to fill in the database at the level of species.

Ultimately, our decisions were guided by a balance of these considerations. With a few notable exceptions, we opted for taxa that were relatively homogeneous, in hopes that information about individual species within the taxon would generalize to other members of the taxon.^[26] However, where the quantity of available research was limited, we pushed upward the taxonomic hierarchy, sometimes up to the level of order, for example with spiders (order Araneae). This is higher up the taxonomic hierarchy than we otherwise would like to go, but given the relevant practical limits, it was the best we could do. As more research on invertebrate sentience becomes available, we hope to be able to offer a more fine-grained analysis.

Project Limitations

In this section we describe nine limitations of our database.

Morgan’s Canon vs. the Precautionary Principle

Our database alone cannot be used to draw any definitive conclusions about the distribution of valenced experience among invertebrates (or any other part of the biological world). The information in the database must be suitably interpreted according to one’s favored epistemic and pragmatic principles.

According to Morgan’s Canon,^[27] animal behavior should not be explained in terms of “higher” psychological processes (in our case, that means conscious experiences) if the behavior can be fairly explained in terms of “lower” psychological processes (unconscious reflexes and instincts). Morgan’s Canon does not preclude attributing valenced experience to nonhuman animals, so long as it can be shown that similar explanations which do not invoke valenced experience are inadequate. Morgan’s Canon does, however, place the burden of proof squarely on those who would attribute valenced experience to nonhuman animals. Morgan’s Canon is an epistemic principle. It provides a criterion for interpreting animal behavior. It aims to shape our beliefs about animal consciousness.

The precautionary principle, on the other hand, is a a pragmatic principle and thus aims to guide our actions. According to the precautionary principle, when the evidence that an animal possesses the capacity for valenced experience is mixed, we should err on the side of caution when interacting with the animal. When formulating animal welfare strategies, the precautionary principle tells us that, so long as there is some reasonable chance that an animal has the capacity for valenced experience, we ought to afford the animal the protections and moral concerns that would be warranted if the animal did possess the capacity for valenced experience. According to the precautionary principle, even if we believe an animal lacks the capacity for valenced experience, so long as there is a reasonable chance we are wrong, we ought to give the animal the benefit of the doubt by acting as if it did have the capacity for valenced experience.

Morgan’s Canon and the precautionary principle need not be in tension. Morgan’s Canon tells us how we ought to set our credences with respect to animal consciousness. The precautionary principle tells us where we ought to set the evidential bar for treating animals as if they don’t possess the capacity for valenced experience. Taken together, both Morgan’s Canon and the precautionary principle can play a role in a well-informed cost-benefit analysis of the importance of invertebrate welfare. One might conclude that the odds that invertebrates have the capacity for valenced experience are low, but that the gains to be had from addressing invertebrate welfare are potentially sufficiently high enough as to make it a cause area worth prioritizing.

The Selection of Features Influences the Results

The number and composition of features selected for a project such as ours necessarily affect the broad contours of the investigation. This fact complicates interpretation of our results. It would be convenient if there were some function that took us from the percentage of features that a taxon satisfies to a probability that animals in that taxon are conscious. Unfortunately, without knowing the necessary and sufficient conditions on consciousness (assuming there are such conditions), there is no way to know which set of features would justify this inference over all potentially sentient organisms. The type of evidence for sentience we do have available is not always applicable to all lifeforms. For this project, we were explicitly aiming to evaluate the evidence for sentience relevant to invertebrates. Hence, we did not want to consider features that would obviously only be applicable to higher mammals.^[28] So, for instance, we did not include ‘ability to process language’ or ‘ability to report experiences to scientists’ because invertebrates are plainly not able to satisfy these features.^[29] For this reason, we advise against thinking about our results solely in terms of what percentage of features a taxon satisfies. For practical reasons, the selection of features in a project such as ours will reflect the interests and emphases of the researchers. Somewhat arbitrary decisions will then come to affect what percentage of features a given taxon satisfies.

Not All Features Are Relevant for All Species

There is probably no set of features which is universally relevant for the detection of valenced experience. Owing to anatomical, environmental, and social differences among animals, valenced experiences are apt to be expressed in behaviorally and neurobiologically distinct ways. Different species of animal, after all, are different. To take a trivial example: most of the time, when humans are in pain, they grimace. But the hard exoskeleton of an insect does not allow for grimacing. Is the fact that insects don’t grimace (a small piece of) evidence that they don’t feel feel pain? Presumably not. Insects can’t grimace, so the fact that they don’t grimace is irrelevant. More subtly, we might expect animals in different environments to be sensitive to different types of noxious stimuli. Aquatic animals, for example, might not have nociceptors that register extreme differences in temperature because they never have to deal with the threat of burning.^[30] But just because aquatic animals can’t experience one type of pain doesn’t mean that they can’t experience pain at all. When constructing similarity arguments either for or against a particular species’ capacity for valenced experience, one should keep in mind the environment in which the creature evolved and how different evolutionary pressures might lead valenced experiences to be expressed differently.^[31]

Anthropomorphism

Another worry is that the lifestyle and environment of phylogenetically distant animals may be so alien that their behavior cannot even be described without resorting to problematic anthropomorphizing. Honey bees, for instance, last shared a common ancestor with humans about 797 million years ago. Honey bees have four distinct life stages: egg, larva, pupa, and adult. The typical worker bee does not usually survive longer than two or three months. Honey bees communicate with each other via complicated dances; they use pheromones to signal alarm, orientation, and colony recognition; they rely on olfaction to locate novel food sources. Given these radical differences, is it possible to detect honey bee emotions? According to some researchers, yes. Agitated honey bees allegedly exhibit pessimistic cognitive biases, a hallmark of human anxiety. Hence, we have evidence for negative emotional states in honey bees.

Such studies should not be dismissed out of hand. Honey bees differ dramatically from humans, but we should not make the mistake of supposing that emotional similarity, at least to a certain degree, is impossible. Still, it may be helpful to regard such studies somewhat skeptically. Terms like ‘anxious’ and ‘tense’ come loaded with anthropocentric connotations. Applying these terms to invertebrates may inevitably be misleading.

The Literature Might Be Biased Toward Surprising Results

The investigation of valenced experience in invertebrates is very much in its infancy. Because there may be a general bias toward non-null experimental results in the sciences, especially for relatively immature fields, we should be cautious about the conclusions of any one study. There is reason to think that academic journals favor papers with surprising results over papers which merely confirm the expected. In our case that might mean that studies which purport to demonstrate novel behavior in invertebrates are overrepresented in the literature. Replication studies are, in general, under-rewarded in academia, so correcting for this overrepresentation may take years or even decades. Because much of the literature we canvassed is so recent, some of the conclusions we reached about invertebrate behavior will undoubtedly be proven mistaken in the future.^[32]

We Don’t Consider How Widespread a Feature Is Within a Taxon

According to our methodology, so long as a single species within a given taxon possesses the feature in question, the entire taxon counts as possessing the feature. We did not systematically investigate how widespread a feature is within each taxon. Such an investigation would have been either extremely time-intensive (if species-level data within a taxon are abundant) or, more likely, inconclusive (if species-level data within a taxon are sparse). For relatively small, relatively homogeneous taxa, this methodology is unproblematic. However, in some cases available research constraints pushed us to consider taxa at a higher rank than we otherwise would have liked. (For example, we investigated spiders at the rank of order. Order Araneae contains approximately 45,000 species.) In these cases, generalizing from a single species to the rest of the taxon is potentially problematic. At the extreme, a small number of unrepresentative species may bias the entire taxon. The case of plants (kingdom Plantae) is a good example. For instance, very few plants are capable of movement rapid enough to be detected with the naked eye, but these species are disproportionately represented in our database.^[33]

Most Features Come in Degrees

One of the most important limitations of our database is that many (perhaps most) of the features we investigated come in degrees. For example, because centralization is not an all-or-nothing phenomenon, the question of whether some animal possesses a centralized nervous system is better answered with a scalar value than a “yes” or “no.” Unfortunately, our database is not equipped to provide such information. Our four credence buckets (0-.25, .25-.50, .50-.75, .75-1.00) represent our position regarding whether an animal possesses a feature; they do not necessarily represent the extent to which the animal possesses the feature. For example, we might believe with high credence (.8) that a given animal possesses a nervous system only 40% centralized relative to some relevant standard. On the other hand, we might believe with low credence (.4) that a given animal possesses a nervous system 80% centralized relative to some relevant standard. These different epistemic positions might call for subtly different reactions. Unfortunately, to properly account for such differences, we would need to include a second axis for each feature, an inclusion which would have made the database significantly more complicated. We opted for the simplicity of a single axis system over the added nuance of a two-axis system.^[34]

It’s Sometimes Difficult to Ascertain Whether a Creature Possesses a Feature

Even when a feature doesn’t come in degrees, it is often difficult to determine with high confidence if an animal possesses the feature under investigation. Sometimes this uncertainty stems from a dearth of scientific data. (In these cases, the relevant cells have been marked ‘unknown.’) More often, a fair amount of research is available, but the data are inconclusive for one or more reasons. Self-recognition, for instance, is plausibly an important feature to investigate. (If self-recognition implies self-awareness, then on some theories of mind, self-recognition is pretty direct evidence of consciousness.) Self-recognition in nonhuman animals is generally measured via the mirror test. The test assesses self-recognition by determining whether an animal can recognize its own reflection in a mirror. This is accomplished by secretly marking the animal with a small dot that is only visible by looking in the mirror. If the animal sees the dot in the mirror then touches the dot on its body (indicating it understands the relationship between itself and the creature in the mirror), it passes the test.

The mirror test has several well-known methodological shortcomings. For example, because the test uses visual perception, creatures which perceive the world primarily through other sense modalities are at a disadvantage. (Dogs, who have repeatedly failed the classic mirror test, navigate the world mostly by a combination of olfaction and audition.) Creatures for whom eye contact is a sign of aggression also seem to be at a disadvantage because they will either refuse to directly investigate their reflection, or, if they do make protracted eye contact, will move to counteract the perceived aggression of the ‘foreign’ creature before they have an opportunity to notice the dot. (Gorillas, who are mostly reported to fail the mirror test, fit this profile.)

These flaws lead to false negatives. But false positives are also possible. The mirror test is an imperfect measure of self-recognition. But even if the mirror test were a perfect measure, it would still be difficult in some cases to definitively say whether some animal passes the test. For instance, according to a recent study, a species of bony fish (the bluestreak cleaner wrasse, Labroides dimidiatus) passes the mirror test, the first animal outside birds and mammals to do so. From our non-specialist perspective, it is unclear how to evaluate such a new, controversial finding, especially given the small sample size of the study. Thus, for this feature and many others, we have been forced to make some difficult judgment calls. Again, we have tried to indicate as much in comments attached to the relevant cells.

It’s Unclear How Much Evidential Weight to Assign the Features

Not all features provide equal evidence for the capacity for valenced experience, and the range in evidential power among the features is probably fairly extreme. (Indeed, some features might not provide any evidence for the capacity for valenced experience.) Knowing the evidential power of the features would make the construction of sound similarity arguments much easier. Nonetheless, we have made no attempt to assign evidential weights to the features. The reason we have not done so is because such a task is too ambitious to complete within the time constraints we imposed on this project. Assigning evidential weights is a difficult task, requiring more scientific investigation and philosophical reflection than we are currently prepared to undertake. Even after such an undertaking, chances are that the weights we could confidently assign the features would come in wide ranges, undermining their usefulness. In this post we have briefly explained why we investigated the features that we did. These justifications offer hints as to our tentative views on the importance of various features. These tentative views, however, are too hazy and incomplete to easily systematize in a way that would allow for fruitful cross-comparison of features. Assigning evidential weights to these features is an endeavor we leave to other researchers (including, potentially, our future selves).

Conclusion and Forthcoming Work

This post summarizes Rethink Priorities’ effort to better understand the issue of invertebrate sentience. We created a comprehensive database of the extant scientific literature relevant to invertebrate sentience, covering 53 features across 18 biological taxa. This database will help others gauge the distribution of sentience outside the subphylum Vertebrata. In forthcoming work, we draw on this database (as well as a variety of other inputs) to analyze whether invertebrate welfare is a promising cause area.

Credits

This essay is a project of Rethink Priorities. It was written by Jason Schukraft with contributions from Daniela R. Waldhorn, Marcus A. Davis, Peter Hurford, and Max Carpendale. Kim Cuddington, Marcus A. Davis, Neil Dullaghan, Peter Hurford, Tegan McCaslin, David Moss, and Daniela R. Waldhorn provided helpful comments on this essay. If you like our work, please consider subscribing to our newsletter. You can see all our work to date here.

Notes