In 2021, Blake Lemoine at Google stirred up a media storm by proclaiming that one of the chatbots he worked on, LaMDA, was sentient—and he subsequently got fired. In fact, most deep learning models are loosely based on the brain’s inner workings. AI agents are increasingly endowed with human-like decision-making algorithms. The idea that machine intelligence could become sentient one day no longer seems like science fiction.
How could we tell if machine have gained sentience?
A preprint paper authored by 19 neuroscientists, philosophers, and computer scientists argues that the neurobiology of consciousness may be our best bet. Rather than simply studying an AI agent’s behavior or responses—for example, during a chat—matching its responses to theories of human consciousness could provide a more objective measure.
It is an unusual theory, but one that makes sense. Speculations of how consciousness emerges in the brain are plenty, with multiple leading candidates still being tested in global head-to-head trials.
The authors don’t subscribe to any single neurobiological theory of consciousness. Instead, they derived a checklist of “indicator properties” of consciousness based on multiple leading ideas. There isn’t a strict cutoff—say, meeting X number of criteria means an AI agent is conscious. Rather, the indicators make up a moving scale: the more criteria met, the more likely a sentient machine mind is.
Using the guidelines to test several recent AI systems, including ChatGPT and other chatbots, the team concluded that for now, “no current AI systems are conscious.”
However, they also said that “there are no obvious technical barriers to building AI systems that satisfy these indicators”. It is possible that conscious AI systems could realistically be built in the near term.
Listening to an Artificial Brain
Since Alan Turing’s famous imitation game in the 1950s, scientists have pondered how to prove whether a machine exhibits intelligence like a human’s. Better known as the Turing test, the theoretical setup has a human judge conversing with a machine and another human—the judge has to decide which participant has an artificial mind. At the heart of the test is the provocative question “Can machines think?” The harder it is to tell the difference between machine and human, the more machines have advanced toward human-like intelligence.
ChatGPT broke the Turing test. An example of a chatbot powered by a large language model (LLM), ChatGPT soaks up internet comments, memes, and other content. It’s extremely adept at emulating human responses—writing essays, passing exams, dishing out recipes, and even doling out life advice.
These advances, which came at a shocking speed, stirred up debate on how to construct other criteria for gauging thinking machines. Most recent attempts have focused on standardized tests for humans: for example, those designed for high school students, the Bar exam for lawyers, or the GRE for entering grad school. OpenAI’s GPT-4, the AI model behind ChatGPT, scored in the top 10 percent of participants. However, it struggled with finding rules for a relatively simple visual puzzle game.
The new benchmarks, while measuring a kind of “intelligence,” don’t necessarily tackle the problem of consciousness. Here’s where neuroscience comes in.
The Checklist for Consciousness
Neurobiological theories of consciousness are many and confusing. But at their heart is neural computation: that is, how neurons connect and process information so it reaches the conscious human mind. In other words, consciousness is the result of the human brain’s computation, although humans don’t yet fully understand the details involved.
This practical look at consciousness makes it possible to translate theories from human consciousness to AI. Called computational functionalism, the hypothesis rests on the idea that computations of the right kind generate consciousness regardless of the medium. According to the team, this suggests that “consciousness in AI is possible in principle”.
Then comes the hard part: how do you probe consciousness in an algorithmic black box? A standard method in humans is to measure electrical pulses in the brain or with functional MRI that captures activity in high definition—but neither method is feasible for evaluating code.
Instead, the team took a “theory-heavy approach,” which was first used to study consciousness in non-human animals. To start, they mined top theories of human consciousness, including the popular Global Workspace Theory (GWT) for indicators of consciousness. For example, GWT stipulates that a conscious mind has multiple specialized systems that work in parallel; we can simultaneously hear and see and process those streams of information. However, there’s a bottleneck in processing, requiring an attention mechanism.
The Recurrent Processing Theory suggests that information needs to feed back onto itself in multiple loops as a path towards consciousness. Other theories emphasize the need for a “body” of sorts that receives feedback from the environment and uses those learnings to better perceive and control responses to a dynamic outside world—something called “embodiment.”
With myriad theories of consciousness to choose from, the team laid out some ground rules. To be included, a theory needs substantial evidence from lab tests, such as studies capturing the brain activity of people in different conscious states. Overall, six theories met the mark. From there, the team developed 14 indicators.
It’s not one-and-done. None of the indicators mark a sentient AI on their own. In fact, standard machine learning methods can build systems that have individual properties from the list. Rather, the list is a scale—the more criteria met, the higher the likelihood an AI system has some kind of consciousness.
How to assess each indicator? We’ll need to look into the “architecture of the system and how the information flows through it”.
In a proof of concept, the team used the checklist on several different AI systems, including the transformer-based large language models that underlie ChatGPT and algorithms that generate images, such as DALL-E 2. The results were hardly conclusive, with some AI systems meeting a portion of the criteria while lacking in others.
However, although not designed with a global workspace in mind, each system “possesses some of the GWT indicator properties,” such as attention, said the team. Meanwhile, Google’s PaLM-E system, which injects observations from robotic sensors, met the criteria for embodiment.
None of the state-of-the-art AI systems checked off more than a few boxes, leading the authors to conclude that we haven’t yet entered the era of sentient AI. They further warned about the dangers of under-attributing consciousness in AI, which may risk allowing “morally significant harms,” and anthropomorphizing AI systems when they’re just code.
The report is far from the final word on the topic. As neuroscience further narrows down correlates of consciousness in the brain, the checklist will likely scrap some criteria and add others. For now, it’s a project in the making, and the authors invite other perspectives from multiple disciplines—neuroscience, philosophy, computer science, cognitive science—to further improve the list.