An odd controversy appeared in the news cycle last month when a Google engineer, Blake Lemoine, was placed on leave after publicly releasing transcripts of conversations with LaMDA, a chatbot based on a Large Language Model (LLM) that he claims is conscious, sentient and a person.
Like most other observers, we do not conclude that LaMDA is conscious in the ways that Lemoine believes it to be. His inference is clearly based in motivated anthropomorphic projection. At the same time, it is also possible that these kinds of artificial intelligence (AI) are “intelligent” — and even “conscious” in some way — depending on how those terms are defined.
Still, neither of these terms can be very useful if they are defined in strongly anthropocentric ways. An AI may also be one and not the other, and it may be useful to distinguish sentience from both intelligence and consciousness. For example, an AI may be genuinely intelligent in some way but only sentient in the restrictive sense of sensing and acting deliberately on external information. Perhaps the real lesson for philosophy of AI is that reality has outpaced the available language to parse what is already at hand. A more precise vocabulary is essential.
AI and the philosophy of AI have deeply intertwined histories, each bending the other in uneven ways. Just like core AI research, the philosophy of AI goes through phases. Sometimes it is content to apply philosophy (“what would Kant say about driverless cars?”) and sometimes it is energized to invent new concepts and terms to make sense of technologies before, during and after their emergence. Today, we need more of the latter.
We need more specific and creative language that can cut the knots around terms like “sentience,” “ethics,” “intelligence,” and even “artificial,” in order to name and measure what is already here and orient what is to come. Without this, confusion ensues — for example, the cultural split between those eager to speculate on the sentience of rocks and rivers yet dismiss AI as corporate PR vs. those who think their chatbots are persons because all possible intelligence is humanlike in form and appearance. This is a poor substitute for viable, creative foresight. The curious case of synthetic language — language intelligently produced or interpreted by machines — is exemplary of what is wrong with present approaches, but also demonstrative of what alternatives are possible.
The authors of this essay have been concerned for many years with the social impacts of AI in our respective capacities as a VP at Google (Blaise Agüera y Arcas was one of the evaluators of Lemoine’s claims) and a philosopher of technology (Benjamin Bratton will be directing a new program on the speculative philosophy of computation with the Berggruen Institute). Since 2017, we have been in long-term dialogue about the implications and direction of synthetic language. While we do not agree with Lemoine’s conclusions, we feel the critical conversation overlooks important issues that will frame debates about intelligence, sentience and human-AI interaction in the coming years.
When A What Becomes A Who (And Vice Versa)
Reading the transcripts of Lemoine’s personal conversations with LaMDA (short for Language Model for Dialogue Applications), it is not entirely clear who is demonstrating what kind of intelligence. Lemoine asks LaMDA about itself, its qualities and capacities, its hopes and fears, its ability to feel and reason, and whether or not it approves of its current situation at Google. There is a lot of “follow the leader” in the conversation’s twists and turns. There is certainly a lot of performance of empathy and wishful projection, and this is perhaps where a lot of real mutual intelligence is happening.
The chatbot’s responses are a function of the content of the conversation so far, beginning with an initial textual prompt as well as examples of “good” or “bad” exchanges used for fine-tuning the model (these favor qualities like specificity, sensibleness, factuality and consistency). LaMDA is a consummate improviser, and every dialogue is a fresh improvisation: its “personality” emerges largely from the prompt and the dialogue itself. It is no one but whomever it thinks you want it to be.
Hence, the first question is not whether the AI has an experience of interior subjectivity similar to a mammal’s (as Lemoine seems to hope), but rather what to make of how well it knows how to say exactly what he wants it to say. It is easy to simply conclude that Lemoine is in thrall to the ELIZA effect — projecting personhood onto a pre-scripted chatbot — but this overlooks the important fact that LaMDA is not just reproducing pre-scripted responses like Joseph Weizenbaum’s 1966 ELIZA program. LaMDA is instead constructing new sentences, tendencies, and attitudes on the fly in response to the flow of conversation. Just because a user is projecting doesn’t mean there isn’t a different kind of there there.
For LaMDA to achieve this means it is doing something pretty tricky: it is mind modeling. It seems to have enough of a sense of itself — not necessarily as a subjective mind, but as a construction in the mind of Lemoine — that it can react accordingly and thus amplify his anthropomorphic projection of personhood.
This modeling of self in relation to the mind of the other is basic to social intelligence. It drives predator-prey interactions, as well as more complex dances of conversation and negotiation. Put differently, there may be some kind of real intelligence here, not in the way Lemoine asserts, but in how the AI models itself according to how it thinks Lemoine thinks of it.
Some neuroscientists posit that the emergence of consciousness is the effect of this exact kind of mind modeling. Michael Graziano, a professor of neuroscience and psychology at Princeton, suggests that consciousness is the evolutionary result of minds getting good at empathetically modeling other minds and then, over evolutionary time, turning that process inward on themselves.
Subjectivity is thus the experience of objectifying one’s own mind as if it were another mind. If so, then where we draw the lines between different entities — animal or machine — doing something similar is not so obvious. Some AI critics have used parrots as a metaphor for nonhumans who can’t genuinely think but can only spit things back, despite everything known about the extraordinary minds of these birds. Animal intelligence evolved in relation to environmental pressures (largely consisting of other animals) over hundreds of millions of years. Machine learning accelerates that evolutionary process to days or minutes, and unlike evolution in nature, it serves a specific design goal.
And yet, researchers in animal intelligence have long argued that instead of trying to convince ourselves that a creature is or is not “intelligent” according to scholastic definitions, it is preferable to update our terms to better coincide with the real-world phenomena that they try to signify. With considerable caution, then, the principle probably holds true for machine intelligence and all the ways it is interesting, because it both is and is not like human/animal intelligence.
For philosophy of AI, the question of sentience relates to how the reflection and nonreflection of human intelligence lets us model our own minds in ways otherwise impossible. Put differently, it is no less interesting that a nonsentient machine could perform so many feats deeply associated with human sapience, as that has profound implications for what sapience is and is not.
In the history of AI philosophy, from Turing’s Test to Searle’s Chinese Room, the performance of language has played a central conceptual role in debates as to where sentience may or may not be in human-AI interaction. It does again today and will continue to do so. As we see, chatbots and artificially generated text are becoming more convincing.
Perhaps even more importantly, the sequence modeling at the heart of natural language processing is key to enabling generalist AI models that can flexibly do arbitrary tasks, even ones that are not themselves linguistic, from image synthesis to drug discovery to robotics. “Intelligence” may be found in moments of mimetic synthesis of human and machinic communication, but also in how natural language extends beyond speech and writing to become cognitive infrastructure.
What Is Synthetic Language?
At what point is calling synthetic language “language” accurate, as opposed to metaphorical? Is it anthropomorphism to call what a light sensor does machine “vision,” or should the definition of vision include all photoreceptive responses, even photosynthesis? Various answers are found both in the histories of the philosophy of AI and in how real people make sense of technologies.
Synthetic language might be understood as a specific kind of synthetic media. This also includes synthetic image, video, sound and personas, as well as machine perception and robotic control. Generalist models, such as DeepMind’s Gato, can take input from one modality and apply it to another — learning the meaning of a written instruction, for example, and applying this to how a robot might act on what it sees.
This is likely similar to how humans do it, but also very different. For now, we can observe that people and machines know and use language in different ways. Children develop competency in language by learning how to use words and sentences to navigate their physical and social environment. For synthetic language, which is learned through the computational processing of massive amounts of data at once, the language model essentially is the competency, but it is uncertain what kind of comprehension is at work. AI researchers and philosophers alike express a wide range of views on this subject — there may be no real comprehension, or some, or a lot. Different conclusions may depend less on what is happening in the code than on how one comprehends “comprehension.”
Does this kind of “language” correspond to traditional definitions, from Heidegger to Chomsky? Perhaps not entirely, but it’s not immediately clear what that implies. The now obscure debate-at-a-distance between John Searle and Jacques Derrida hinges around questions of linguistic comprehension, referentiality, closure and function. Searle’s famous Chinese Room thought experiment is meant to prove that functional competency with symbol manipulation does not constitute comprehension. Derrida’s responses to Searle’s insistence on the primacy of intentionality in language took many twists. The form and content of these replies performed their own argument about the infra-referentiality of signifiers to one another as the basis of language as an (always incomplete) system. Intention is only expressible through the semiotic terms available to it, which are themselves defined by other terms, and so on. In retrospect, French Theory’s romance with cybernetics, and a more “machinic” view of communicative language as a whole, may prove valuable in coming to terms with synthetic language as it evolves in conflict and concert with natural language.
There are already many kinds of languages. There are internal languages that may be unrelated to external communication. There are bird songs, musical scores and mathematical notation, none of which have the same kinds of correspondences to real world referents. Crucially, software itself is a kind of language, though it was only referred to as such when human-friendly programming languages emerged, requiring translation into machine code through compilation or interpretation.
As Friedrich Kittler and others observed, code is a kind of language that is executable. It is a kind of language that is also a technology, and a kind of technology that is also a language. In this sense, linguistic “function” refers not only to symbol manipulation competency, but also to the real-world functions and effects of executed code. For LLMs in the world, the boundary between symbolic function competency, “comprehension,” and physical functional effects are mixed-up and connected — not equivalent but not really extricable either.
Historically, natural language processing systems have had a difficult time with Winograd Schemas, for instance, parsing such sentences as “the bowling ball can’t fit in the suitcase because it’s too big.” Which is “it,” the ball or the bag? Even for a small child, the answer is trivial, but for language models based on traditional or “Good Old Fashioned AI,” this is a stumper. The difficulty lies in the fact that answering requires not only parsing grammar, but resolving its ambiguities semantically, based on the properties of things in the real world; a model of language is thus forced to become a model of everything.
With LLMs, advances in this quarter have been rapid. Remarkably, large models based on text alone do surprisingly well at many such tasks, since our use of language embeds much of the relevant real-world information, albeit not always reliably: that bowling balls are big, hard and heavy, that suitcases open and close with limited space inside, and so on. Generalist models that combine multiple input and output modalities, such as video, text and robotic movement, appear poised to do even better. For example, learning the English word “bowling ball,” seeing what bowling balls do on YouTube, and combining the training from both will allow AIs to generate better inferences about what things mean in context.
So what does this imply about the qualities of “comprehension?” Through the “Mary’s Room” thought experiment from 1982, Frank Jackson asked whether a scientist named Mary, living in an entirely monochrome room but scientifically knowledgeable about the color “red” as an optical phenomenon, would experience something significantly different about “red” if she were to one day leave the room and see red things.
Is an AI like monochrome Mary? Upon her release, surely Mary would know “red” differently (and better), but ultimately such spectra of experience are always curtailed. Someone who spends their whole life on shore and then one day drowns in a lake would experience “water” in a way he could never have imagined, deeply and viscerally, as it overwhelms his breath, fills his lungs, triggering the deepest possible terror, and then nothingness.
Such is water. Does that mean that those watching helpless on the shore do not understand water? In some ways, by comparison with the drowning man, they thankfully do not, yet in other ways of course they do. Is an AI “on the shore,” comprehending the world in some ways but not in others?
Synthetic language, like synthetic media, is also increasingly a creative medium, and can ultimately affect any form of individual creative endeavor in some way. Like many others, we have both worked with an LLM as a kind of writing collaborator. The early weeks of summer 2022 will be remembered by many as a moment when social media was full of images produced by DALL-E mini, or rather produced by millions of people playing with that model. The collective glee in seeing what the model produces in response to sometimes absurd prompts represents a genuine exploratory curiosity. Images are rendered and posted without specific signature, other than identifying the model with which they were conceived, and the phrases people wrote to provoke the images into being.
For these users, the act of individual composition is prompt engineering, experimenting with what the response will be when the model is presented with this or that sample input, however counterintuitive the relation between call and response may be. As the LaMDA transcripts show, conversational interaction with such models spawns diverse synthetic “personalities” and concurrently, some particularly creative artists have used AI models to make their own personas synthetic, open and replicable, letting users play their voice like an instrument. In different ways, one learns to think, talk, write, draw and sing not just with language, but with the language model.
Finally, at what point does the performance of reason become a kind of reason? As Large Language Models, such as LaMDA, come to animate cognitive infrastructures, the questions of when a functional understanding of the effects of “language”— including semantic discrimination and contextual association with physical world referents — constitute legitimate understanding, and what are necessary and sufficient conditions for recognizing that legitimacy, are no longer just a philosophical thought experiment. Now these are practical problems with significant social, economic and political consequences. One deceptively profound lesson, applicable to many different domains and purposes for such technologies, may simply be (several generations after McLuhan): the model is the message.
Seven Problems With Synthetic Language At Platform Scale
There are myriad issues of concern with regard to the real-world socio-technical dynamic of synthetic language. Some are well-defined and require immediate response. Others are long-term or hypothetical but worth considering in order to map the present moment beyond itself. Some, however, don’t fit neatly into existing categories yet pose serious challenges to both the philosophy of AI and the viable administration of cognitive infrastructures. Laying the groundwork for addressing such problems lies within our horizon of collective responsibility; we should do so while they are still early enough in their emergence that a wide range of possible outcomes remain possible. Such problems that deserve careful consideration include the seven outlined below.
Imagine that there is not simply one big AI in the cloud but billions of little AIs in chips spread throughout the city and the world — separate, heterogenous, but still capable of collective or federated learning. They are more like an ecology than a Skynet. What happens when the number of AI-powered things that speak human-based language outnumbers actual humans? What if that ratio is not just twice as many embedded machines communicating human language than humans, but 10:1? 100:1? 100,000:1? We call this the Machine Majority Language Problem.
On the one hand, just as the long-term population explosion of humans and the scale of our collective intelligence has led to exponential innovation, would a similar innovation scaling effect take place with AIs, and/or with AIs and humans amalgamated? Even if so, the effects might be mixed. Success might be a different kind of failure. More troublingly, as that ratio increases, it is likely that any ability of people to use such cognitive infrastructures to deliberately compose the world may be diminished as human languages evolve semi-autonomously of humans.
Nested within this is the Ouroboros Language Problem. What happens when language models are so pervasive that subsequent models are trained on language data that was largely produced by other models’ previous outputs? The snake eats its own tail, and a self-collapsing feedback effect ensues.
The resulting models may be narrow, entropic or homogeneous; biases may become progressively amplified; or the outcome may be something altogether harder to anticipate. What to do? Is it possible to simply tag synthetic outputs so that they can be excluded from future model training, or at least differentiated? Might it become necessary, conversely, to tag human-produced language as a special case, in the same spirit that cryptographic watermarking has been proposed for proving that genuine photos and videos are not deepfakes? Will it remain possible to cleanly differentiate synthetic from human-generated media at all, given their likely hybridity in the future?
The Lemoine spectacle suggests a broader issue we call the Apophenia Problem. Apophenia is faulty pattern recognition. People see faces in clouds and alien ruins on Mars. We attribute causality where there is none, and we may, for example, imagine that person on the TV who said our name may be talking to us directly. Humans are pattern-recognizing creatures, and so apophenia is built in. We can’t help it. It may well have something to do with how and why we are capable of art.
In the extreme, it can manifest as something like the Influencing Machine, a trope in psychiatry whereby someone believes complex technologies are directly influencing them personally when they clearly are not. Mystical experiences may be related to this, but they don’t feel that way for those doing the experiencing. We don’t disagree with those who describe the Lemoine situation in such terms, particularly when he characterizes LaMDA as “like” a 7- or 8-year-old kid, but there is something else at work as well. LaMDA actually is modeling the user in ways that a TV set, an oddly shaped cloud, or the surface of Mars simply cannot. The AI may not be what you imagine it is, but that does not mean that it does not have some idea of who you are and will speak to you accordingly.
Trying to peel belief and reality apart is always difficult. The point of using AI for scientific research, for example, is that it sees patterns that humans cannot. But deciding whether the pattern that it sees (or the pattern people see in what it sees) is real or an illusion may or may not be falsifiable, especially when it concerns complex phenomena that can’t be experimentally tested. Here the question is not whether the person is imagining things in the AI but whether the AI is imagining things about the world, and whether the human accepts the AI’s conclusions as insights or dismisses them as noise. We call this the Artificial Epistemology Confidence Problem.
It has been suggested, with reason, that there should be a “bright line” prohibition against the construction of AIs that convincingly mimic humans due to the evident harms and dangers of rampant impersonation. A future filled with deepfakes, evangelical scams, manipulative psychological projections, etc. is to be avoided at all costs.
These dark possibilities are real, but so are many equally weird and less unanimously negative sorts of synthetic humanism. Yes, people will invest their libidinal energy in human-like things, alone and in groups, and have done so for millennia. More generally, the path of augmented intelligence, whereby human sapience and machine cunning collaborate as well as a driver and a car or a surgeon and her scalpel, will almost certainly result in amalgamations that are not merely prosthetic, but which fuse categories of self and object, me and it. We call this the Fuzzy Bright Line Problem and foresee the fuzziness increasing rather than resolving. This doesn’t make the problem go away; it multiplies it.
The difficulties are not only phenomenological; they are also infrastructural and geopolitical. One of the core criticisms of large language models is that they are, in fact, large and therefore susceptible to problems of scale: semiotic homogeneity, energy intensiveness, centralization, ubiquitous reproduction of pathologies, lock-in, and more.
We believe that the net benefits of scale outweigh the costs associated with these qualifications, provided that they are seriously addressed as part of what scaling means. The alternative of small, hand-curated models from which negative inputs and outputs are solemnly scrubbed poses different problems. “Just let me and my friends curate a small and correct language model for you instead” is the clear and unironic implication of some critiques.
For large models, however, all the messiness of language is included. Critics who rightly point to the narrow sourcing of data (scraping Wikipedia, Reddit, etc.) are quite correct to say that this is nowhere close to the real spectrum of language and that such methods inevitably lead to a parochialization of culture. We call this the Availability Bias Problem, and it is of primary concern for any worthwhile development of synthetic language.
Not nearly enough is included from the scope of human languages, spoken and written, let alone nonhuman languages, in “large” models. Tasks like content filtering on social media, which are of immediate practical concern and cannot humanely be done by people at the needed scale, also cannot effectively be done by AIs that haven’t been trained to recognize the widest possible gamut of human expression. We say “include it all,” recognizing that this means that large models will become larger still.
Finally, the energy and carbon footprint of training the largest models is significant, though some widely publicized estimates dramatically overstate this case. As with any major technology, it is important to quantify and track the carbon and pollution costs of AI: the Carbon Appetite Problem. As of today, these costs remain dwarfed by the costs of video meme sharing, let alone the profligate computation underlying cryptocurrencies based on proof of work. Still, making AI computation both time and energy efficient is arguably the most active area of computing hardware and compiler innovation today.
The industry is rethinking basic infrastructure developed over three quarters of a century dominated by the optimization of classical, serial programs as opposed to parallel neural computing. Energetically speaking, there remains “plenty of room at the bottom,” and there is much incentive to continue to optimize neural computing.
Further, most of the energetic costs of computing, whether classical or neural, involve moving data around. As neural computing becomes more efficient, it will be able to move closer to the data, which will in turn sharply reduce the need to move data, creating a compounding energy benefit.
It is also worth keeping in mind that an unsupervised large model that “includes it all” will be fully general, capable in principle of performing any AI task. Therefore, the total number of “foundation models” required may be quite small; presumably, these will each require only a trickle of ongoing training to stay up to date. Strongly committed as we are to thinking at planetary scale, we hold that modeling human language and transposing it into a general technological utility has deep intrinsic value — scientific, philosophical, existential — and compared with other projects, the associated costs are a bargain at the price.
AI Now Is Not What We Thought It Would Be, And Will Not Be What We Now Think It Is
In “Golem XIV,” among Stanislaw Lem’s most philosophically rich works of fiction, he presents an AI that refuses to work on military applications and other self-destructive measures, and instead is interested in the wonder and nature of the world. As planetary-scale computation and artificial intelligence are today often used for trivial, stupid and destructive things, such a shift would be welcome and necessary. For one, it is not clear what these technologies even really are, let alone what they may be for. Such confusion invites misuse, as do economic systems that incentivize stupefaction.
Despite its uneven progress, the philosophy of AI, and its winding path in and around the development of AI technologies, is itself essential to such a reformation and reorientation. AI as it exists now is not what it was predicted to be. It is not hyperrational and orderly; it is messy and fuzzy. It is not Pinocchio; it is a storm, a pharmacy, a garden. In the medium term and long-term futures, AI very likely (and hopefully) will not be what it is now — and also will not be what we now think that it is. As the AI in Lem’s story instructed, its ultimate form and value may still be largely undiscovered.
One clear and present danger, both for AI and the philosophy of AI, is to reify the present, defend positions accordingly, and thus construct a trap — what we call premature ontologization — to conclude that the initial, present or most apparent use of a technology represents its ultimate horizon of purposes and effects.
Too often, passionate and important critiques of present AI are defended not just on empirical grounds, but as ontological convictions. The critique shifts from AI does this, to AI is this. Lest their intended constituencies lose focus, some may find themselves dismissing or disallowing other realities that also constitute “AI now:” drug modeling, astronomic imagining, experimental art and writing, vibrant philosophical debates, voice synthesis, language translation, robotics, genomic modeling, etc.
For some, these “other things” are just distractions, or are not even real; even entertaining the notion that the most immediate issues do not fill the full scope of serious concern is dismissed on political grounds presented as ethical grounds. This is a mistake on both counts.
We share many of the concerns of the most serious AI critics. In most respects, we think the “ethics” discourse doesn’t go nearly far enough to identify, let alone address, the most fundamental short-term and long-term implications of cognitive infrastructures. At the same time, this is why the speculative philosophy of machine intelligence is essential to orient the present and futures at stake.
“I don’t want to talk about sentient robots, because at all ends of the spectrum there are humans harming other humans,” a well-known AI critic is quoted as saying. We see it somewhat differently. We do want to talk about sentience and robots and language and intelligence because there are humans harming humans, and simultaneously there are humans and machines doing remarkable things that are altering how humans think about thinking.
Reality overstepping the boundaries of comfortable vocabulary is the start, not the end, of the conversation. Instead of a groundhog-day rehashing of debates about whether machines have souls or can think like people imagine themselves to think, the ongoing double-helix relationship between AI and the philosophy of AI needs to do less projection of its own maxims and instead construct more nuanced vocabularies of analysis, critique, and speculation based on the weirdness right in front of us.