Schrödinger’s A.I. Could Test the Foundations of Reality

Whenever people talk about quantum-physics thought experiments, they hasten to add they’re just that: thought experiments. When Austrian physicist Erwin Schrödinger mused on that alive/dead cat locked in a box, back in the 1930s, he introduced the experiment as "ganz burleske"—highlighting a comedic absurdity—not as a blueprint for a real-world lab test. Some decades later Hungarian physicist Eugene Wigner laid out a related and more complex setup, effectively involving a human watching their human friend watching a cat in a box. Hugh Everett (himself the father of the quantum parallel worlds hypothesis), who was lectured by Wigner, called this ’Wigner’s friend’ scenario "amusing, but extremely hypothetical." Yet, approaching a century after Schrödinger’s cat entered public consciousness, these Wigner’s-friend tests are starting to seem entirely doable, albeit not with cats or humans, but with artificial intelligence, A.I., run on a quantum computer.

You may not even need a fully human-level artificial intelligence algorithm to learn something interesting, which means such tests could be tantalisingly close. "I don’t think it will take that long," says Eric Cavalcanti, a quantum physicist at Griffith University, in Queensland, Australia, who presented the concept at an FQXi-funded meeting at Chapman University, in Orange County, California, in June. Experimental groups around the world are already building rudimentary quantum computers, with IBM’s Eagle currently leading the pack, stringing together 127 quantum bits, or qubits. "Once we have quantum computers that are able to run a few hundred qubits with sufficient fidelity, we will be able to run very simple programs—not very robust, but robust enough," Cavalcanti says. Early quantum A.I. programs will already have a rudimentary form of some of the qualities of an observer, such as being able to measure and process the state of another quantum system. Because it is a computer program, you could manipulate your quantum-computer-based observer as easily as you do a particle.

The holy grail of the program, however, is an experiment with a human-level AI. With the aid of an FQXi grant of over $60,000, Cavalcanti has been laying the groundwork for such an A.I. test with his colleagues Howard Wiseman, a physicist also at Griffith, and Peter Evans, a philosopher at the University of Queensland, in Brisbane. In particular, Wiseman and Cavalcanti and another colleague have just posted specific experimental plans (Howard Wiseman, Eric Cavalcanti & Eleanor Reiffel, arXiv:2209.08491v1 (2022)). Their goal is to pin down some of the slippery concepts that form the backbone of quantum theory.

The standard story is that particles possess multiple conflicting properties at the same time until their properties are measured. Schrödinger’s cat is in a peculiar limbo between alive and dead until you open the box and look at it, when its fate is set. But what counts as an observation and why does it have this seemingly magical power to shift reality? In a separate FQXi-funded project, quantum physicists Veronika Baumann and Časlav Brukner of the University of Vienna, in Austria, are using their almost $110,000 grant to ask what an observer even is. Does the observer have to be conscious? Can a human observer be part of a quantum system observed by their friend? These are all questions that intrigued Wigner back in the 1950s, when he posed the conundrum.

Wigner’s Friend

What makes Wigner’s experiment so beguiling—and so worth doing for real—is that it makes quantum theory’s paradoxes much more vivid. Consider a standard quantum experiment, in which you prepare a particle so that it is in a ’superposition’ encompassing multiple possibilities. In this case, the particle is in several different locations, and then you measure where it is. You will see the particle in only one of those myriad places. Common sense would tell you that the particle was already in that place, before you looked, and that the superposition simply reflected your ignorance about its actual state. In the parlance, the location is a "hidden variable."

In the 1960s, Irish physicist John Bell proposed experiments that probed this hidden-variables explanation (in the most intuitive sense, at least, but more on that later), by performing tests on strings of pairs of particles that had been ’entangled’—prepared in such a way that their properties are twinned—and counting up the degree to which their different measured properties line up, or correlate. Those Bell tests have since become routine. They have shown that, even if particles do have simply defined positions in advance of being measured, that alone cannot explain the observed correlations. For most physicists, that implies that particles do not, in fact, have such positions. The cat in the box really is a peculiar mix of alive and dead; it’s not, unbenownst to you, one or the other.

Now consider Wigner’s more elaborate setup. Here, Wigner’s friend makes the measurement of a particle’s properties inside a sealed lab, while Wigner stands outside. The friend’s object of study is the particle; Wigner’s is the lab—particle, friend, and all. When the friend does a measurement, she finds the particle in some specific place, as before. But Wigner will continue to describe the particle in a superposition until he performs his own measurement, say, by opening the lab door, or calling his friend on the phone. This potentially has a startling implication: Wigner could conduct a Bell experiment, treating the entire lab room as a single quantum system. In so doing Wigner would rule out hidden variables in the lab, and thus he would argue that according to quantum theory, the state of the interior of the lab is indefinite while it’s closed.

That seems fair enough, except this time the "hidden variables" he has ruled out include his friend’s observations of a definite state of the particle and her own mental state. Later, when Wigner meets up with his friend and she tells him that at noon, when the lab door was shut, she saw the particle in one place, Wigner could legitimately argue that, no, at noon, when the lab door was closed, the particle was in a superposition—and so was she! However, Wiseman suspected that we can say more, by not directly following Bell’s formulation. So he and Cavalcanti set out to investigate what consequences such an experiment would have.

It’s one thing to cast doubt on a bloodless abstraction of "hidden variables," quite another to deny a person’s lived experience. "Observations act a bit like hidden variables in these setups, but they become less abstract," says Baumann. "You cannot really argue them away." To prove this point, in 2015, Brukner was the first to merge Bell’s and Wigner’s experiments into one to test whether there was a simple get-out from the paradox: that the friend’s experiment does have a definite result while the lab is closed, and Wigner simply does not know it, until he opens the lab door. Brukner’s set-up has two Wigner/friend pairs, conventionally named Alice/Charlie and Bob/Debbie, who share a pair of entangled particles (see diagram, above). In Wiseman and Cavalcanti’s version, only a single friend is needed. Both proposed experiments reveal that the story is more complicated.

Both experiments further illuminate a basic tension in quantum mechanics, which was also evident in Bell’s set-up: The theory treats some systems as objects, and some as subjects. Objects evolve deterministically, according to an equation developed by Schrödinger, before being observed. Subjects break this deterministic evolution or at least perceive that they do. Because Wigner’s friend is both subject and object, the theory is caught like a deer in headlights.

Photons as Friends?

In 2019, Massimiliano Proietti and his colleagues at Heriot-Watt University in Edinburgh, UK, performed Brukner’s version of Wigner’s experiment, in which a single particle of light, or photon, played the role of each ’friend,’ or more technically the role of an ’observer.’ "It is ridiculous to think of a photon as a friend," notes Cavalcanti, "but it is not as ridiculous to think of it as an observer." These observer-photons change their behavior based on monitoring the state of another quantum system. (I performed a similar implementation using an IBM cloud-based quantum computer.) Then in 2020, Cavalcanti, Wiseman, and their colleagues went a step further by creating their own distinctive theorem, and also implementing a photonic demonstration on a tabletop. "It shows rigorously that reality itself is incompatible with a set of assumptions that are weaker than anyone had hitherto considered in such a theorem," says Wiseman.

For these basic proof-of-principle tests, it was fine to use a single particle or qubit as a stand-in for an observer. Most people would regard the ’observer’ in these experiments as merely an object—just a particle obeying the Schrödinger equation for how particles evolve before measurement. But Wigner’s puzzle is so perplexing because in his set-up the observer is also a subject, having some inner mental life that an elementary particle entirely lacks. "Until we use something more interesting like an A.I., we have not really addressed the core issues raised by the Wigner’s-friend experiment," says Jacques Pienaar, a quantum physicist at the University of Massachusetts, Boston, who is not involved in the new study.

And there’s a lot at stake. In their 2020 paper, Cavalcanti, Wiseman, and colleagues also proved a new theorem, enabling a new way to interpret the experimental results. It suggests further experiments could reveal something new about the nature of quantum theory, helping physicists choose between different interpretations. By now most people accept that something is very weird about the quantum world—they just disagree on what form this weirdness must take, and what aspect of our everyday intuition we have to sacrifice. We may have to give up the absolute reality of our observations—like proponents of Everett’s parallel universes who allow all possible outcomes of any quantum experiment to be instantiated in some world, somewhere. Or we can hold on to a form of hidden variables, if we accept not just the "spooky actions at a distance" that instantly connect objects across vast distances, which quantum physicists are mostly resigned to accepting, but something even stranger. "What our theorem says we would be forced to accept is a much stronger, more confronting form of nonlocality: that the action of an agent causes distant things to change faster than light could get there," says Wiseman. Or we need to allow for retrocausality, as influences from the future travel back in time to muck up our experiments. "The novelty of our 2020 theorem is that it proves that—if quantum mechanics continues to apply at the level of observers—these are essentially the only three options," adds Cavalcanti.

But some skeptics responded that the experiment was not strong enough to force them to make the tough choice among these unpalatable options, because using a photon as a ’friend’—or even identifying it less anthropomorphically as an ’observer’—is not true to the original spirit of Wigner’s thought experiment. "That is precisely the kind of route to responding to our original theorem that we aim to close with our new paper," Cavalcanti says.

So what’s the next step up from a photon as an observer? Physicists are just about able to put a virus or a dormant tardigrade into a quantum state in a controlled way, but it’s not going to be easy to ask a tardigrade to report back about its observations. A cat or human may always be beyond them, but a perfectly adequate substitute is waiting in the wings: artificial general intelligence. An advanced machine could be an observer in exactly the same sense as we are—only a strong exceptionalist about human intelligence would deny that—and is a whole lot easier to do quantum experiments on.

Philosophy of Mind

To close the loophole and provide a version of the Wigner’s-friend test involving an A.I., Wiseman and Cavalcanti have performed a remarkable mashup of physics and philosophy of mind. They note that lurking in the background of all the theorems of quantum physics are unarticulated assumptions about the nature of observers. The team is making those assumptions explicit.

The first assumption in their new theorem is "local agency": choices made in experimental trials cannot instantaneously influence events at a distance, or in the past. The local-agency assumption ensures this by placing several common-sense restrictions on causation. Influences can’t travel backward in time or faster than light, and the choice and outcome don’t have some hidden common cause.

The remaining assumptions come from the philosophy of mind. The second is physicalism: Thoughts have physical correlates. If you see red, your neurons will fire in a certain way; if you see blue, they will fire in some other way. The third assumption, which the authors call "friendliness," supposes that the thoughts of other intelligent agents are as real as yours. This mental egalitarianism is implicit in the original Wigner’s friend experiment, since Wigner assumes his friend is just like him. It would be arrogant to suppose otherwise. With animals and A.I. systems, it’s harder to be sure, but you could be convinced by, for example, running a Turing test.

The fourth and final assumption is what the authors call "ego-absolutism": My thoughts are real in an absolute sense. Of the four assumptions, this one is special because it is a first-person statement. Like Descartes’ "I think, therefore I am," it lets you reason outward from your own experience. The word "absolute" here means that the reality of your thoughts is unconditional. You don’t have to qualify it by saying they are real only in some worlds or from some perspective.

The team’s theorem then follows by working backward through the assumptions. Your thoughts are absolutely real (assumption 4) and correspond to physical events (assumption 2), so other intelligent agents’ thoughts also correspond to absolutely real physical events (assumption 3). That’s crucial because it allows you to label them as definite things in the world, or hidden variables (of at least a restricted sort). Given local agency (assumption 1), you can then formulate a mathematical relationship, in the form of an inequality, that governs measurement outcomes—similar to but not quite the same as the relationship derived by Bell that is now the workhorse of quantum theory. If a future experiment involving intelligent agents violates this inequality—and quantum systems do violate it in prototype lab tests—at least one assumption must be wrong.

So, when the day comes that an A.I. system can play the role of observer in a large quantum computer, theorists will be ready, and wafflers who dither over the implications of such tests for various quantum interpretations will have to come off the fence. "Now, if we find a violation of these inequalities, then it’s no longer an alternative to say, ’Well, but you’re not using the right kind of observer,’" Cavalcanti says. Instead, one of the explicit assumptions must go.

Tough Choices

So what has to give? Wiseman and Cavalcanti’s team have spent much of the past year sorting out where the various quantum interpretations stand. In some cases, the choice is clear. Retrocausality and superdeterminism explicitly abandon local agency.

But other interpretations go after one of the philosophy-of-mind assumptions. For instance, consider Bohmian mechanics, a type of hidden-variables theory. It holds that particles always have definite, if unknowable, positions. Bohmian mechanics involves faster-than-light influences and thus gives up local agency. Less obviously, it sometimes ditches friendliness, too. The reason is that Bohmian mechanics does not treat particles’ properties evenhandedly; it gives primacy to their positions. That means you can no longer assume that what holds for your brain holds for an A.I. system; it depends on which particle properties they use to encode information.

In some interpretations, a novel physical process, called objective collapse, makes the choice among competing options in a superposition. But this, too, violates friendliness, because in a quantum computer the randomness introduced by these theories can be corrected away. For some period of time, the A.I. system has no definite thoughts, and again you can’t extrapolate from your experience to its.

Several interpretations have it in for ego-absolutism. For instance, relational quantum mechanics and QBism both take all facts to be relative. Evans is sympathetic to this relativism. "When we say that something is an objective fact in the world, my view is that it needs to be tied to a particular perspective," he says.

Making an Observer

With this new theorem, physicists have all they’ll need to interpret a full-up Wigner’s A.I.-friend experiment. Now they just have to get cracking with their A.I.—but what form should that take? "The idea would be, if you have a human-level artificial intelligence algorithm, to run it on a quantum computer," says Wiseman. That’s a long way off, but Cavalcanti suggests that they might work up from a simple reflex agent such as a thermostat to a learning agent with a world model. And while the end goal would be to put a human-equivalent test subject into the experiment, in some ways the early-stage A.I. systems are still intellectually interesting. By testing a series of ever more sophisticated A.I. we could perhaps discover the threshold at which a quantum system qualifies as an observer.

Baumann and Brukner have been studying this question in the context of the Wigner’s friend test, and they ask why the outcome of measurement takes the form of classical rather than quantum information. The standard quantum theory of measurement hinges on correlations between two objects, notes Baumann. For instance, when a thermometer takes on the same temperature as the system, or when an LED lights up to indicate that a particle passed through a polarizing filter, they are correlated. But Baumann explains that’s not enough for a full Wigner’s friend test. "If I want something to really be a Wigner’s-friend type of situation, I need the observer system to be able to do something more than just get correlated with the system it’s supposed to observe," Baumann says. To mirror is not to measure.

A true observation requires some additional distillation of information, Baumann explains. Observers have states of perception, which fix the menu of possible outcomes, such as positions in space, from which the measurement will choose. Observers are also able to extract some information and tell it to the world. "The thing that, for me, would signify that this is a Wigner’s-friend experiment is if I have some sort of communication," Baumann says.

Going back to the Wigner’s friend prototype tests in 2019 and 2020, it’s clear that a single photon, acting as the friend, does none of the things that Baumann demands. But even adding just one more particle could make it interesting. This second particle could ’monitor’ the first, providing a rudimentary simulation of self-awareness, or what Fritz London and Edmund Bauer, in the 1939 paper that introduced the very first version of Wigner’s friend, called a "faculty of introspection." "You have your memory, like in a regular Wigner’s-friend setup, and then you have another qubit that registers changes in your memory," Baumann said. It’s not perfect, but it’s already something.

Spin-Offs

So far, so abstract. But beyond the pure-science value of such experiments, the tests may spin off technological innovations, just as Bell’s original proposal to test the nature of quantum reality eventually led to advances in quantum communications and cryptography. Aephraim Steinberg, an experimentalist at the University of Toronto, Ontario, likes that practical aspect of the project. "The act of developing these experiments also serves side goals, pushing us to develop and improve new technologies," he says.

But could the outcome of such tests ever persuade die-hard fans to switch allegiance from one interpretation to another? Perhaps, or at least it will force them to sharpen their models. Pienaar, though a proponent of QBism, says his favored interpretation needs work, and thinks that seeing a Wigner’s friend in action will focus the mind. "I suspect it will push relational quantum mechanics and QBism to be more precise about how agents and observations are defined and applied in concrete situations," he says. "These interpretations stand to gain or lose credibility depending on whether they can give an elegant account of such experiments."

For now, the greatest intellectual benefit is to bridge old intellectual divides. To make sense of quantum measurement, physicists have to reach across to philosophers of mind. "Yeah, it makes a lot of physicists uncomfortable, and for good reason," Wiseman says. "It is very tricky to think about and liable to get you labeled as not actually doing physics. But I guess I’m at the stage in my career where I don’t have to worry about that."