Could AI Discover Mental States Humans Have Never Experienced?

There is a version of optimization that most people already understand: given a space of options and a score, search the space for the best option. It is how a chess engine plays, how a chemical company tunes a reaction, and how a delivery network routes trucks. The question this article asks is what happens when the thing being searched is not a design or a route but a configuration of a human mind, and when the “best” configuration is one that no human has ever occupied, or that humans occupy so rarely that we have no settled name for it.

The idea is not as exotic as it sounds, because human minds already occupy a restricted slice of their own possible state space. Brains evolved under particular constraints, and the states that recur — ordinary focused work, diffuse reverie, apprehension, drowsiness, playfulness, grief — are the ones that served survival and reproduction well enough to persist. Nothing guarantees that this recurring set is the set that would serve modern mathematics, sustained abstract reasoning, or careful long-horizon planning best. A large share of human culture is, in effect, an attempt to reach states the species does not default to: meditation, ritual, stimulants, sleep deprivation, the deliberate cultivation of flow, the discipline of long proofs. The question is whether a systematic search could do what two hundred thousand years of culture has done only by trial and error — and whether the results would be states we should want to enter.

Separating three very different claims

The intuition behind the question blurs together three claims that need to be kept apart, because they have very different evidence behind them.

The first is that the space of reachable mental states is larger than the space we habitually occupy, and that some of the unoccupied regions would be useful. This is plausible and, in a trivial sense, almost certainly true. Anyone who has taken a novel drug, undergone sleep deprivation, or had an unusually vivid dream has been somewhere they do not usually go.

The second is that an optimization process can find those regions. This is an engineering claim about search: whether the space can be parameterized, whether a feedback signal exists, whether the search can be made safe enough to run in a living brain. Some parts of this claim are demonstrated today; others are not.

The third is that a state being novel and scoring well on the optimizer’s target means it is good for the person. This third claim is where the whole thing becomes ethically load-bearing, and it is the claim that no amount of engineering success settles. An optimizer maximizes the function it is given. If the function is wrong, the optimizer will find the wrong state, will find it efficiently, and will find one that the unaided person would never have found to complain about.

Most public discussion collapses all three into one. The rest of this article tries to keep them separate, because the first is nearly certain, the second is partly demonstrated and partly speculative, and the third is a question about values that has not been answered and cannot be answered by better search.

What a search over mental states would actually search

To make the idea concrete, imagine you wanted to search for a new cognitive configuration. You would need four things: a set of controllable inputs, a way to measure the resulting state, a target to optimize, and a loop connecting them.

The inputs are, in principle, abundant. They include everything a person can already manipulate — sleep timing and duration, light exposure, exercise, breathing patterns, temperature, social context, task structure, nutritional state, pharmacological agents — plus, in a clinical setting, electrical or magnetic stimulation of the brain. Several of these already have well-characterized effects, which is why the search space is not empty. It is also not clean. Each input has side effects and interacts with the others, so the effect of any one depends on the state of the rest.

The measurement is where a search becomes possible today in a way it was not twenty years ago. Functional magnetic resonance imaging, electroencephalography, pupillometry, and simple behavioral performance on well-designed tasks all provide continuous or near-continuous readouts of a person’s state. Much of the work discussed elsewhere in this series — decoding mood from intracranial activity, reconstructing brain-wide dynamics from a scalar arousal signal — is really a proof that some internal states are, in principle, observable through measurement rather than inference (Raut et al., 2025). That observability is what would let a search close its loop.

The target is the hard part, and it is where the vocabulary of “discovery” can mislead. There is no natural quantity called “usefulness of a mental state” that a sensor can read. A target has to be constructed by someone, out of something measurable: task performance, self-report, a neural signature that correlates with a clinical improvement, or a composite of these. Every such target is a proxy, and every proxy can be gamed by a search process that is good enough at its job.

The precedent from adjacent fields

The reason to take the search claim seriously is that optimization has already produced genuinely novel, genuinely useful results in domains that share the structure of this problem — large spaces, expensive evaluations, and no closed-form solution. A 2023 review of AI in scientific discovery catalogued the general architecture behind these wins: a hypothesis space, a scoring mechanism, and a loop that proposes, tests, and updates (Wang et al., 2023).

The clearest precedent is autonomous experimentation in materials science. In 2023, a team at Berkeley Lab and the University of California, Berkeley, together with Google DeepMind, reported an autonomous laboratory that combined computations, machine learning trained on the scientific literature, and robotic synthesis to search for new inorganic materials (Szymanski et al., 2023). Over seventeen days of continuous operation, the system targeted 58 candidate compounds and synthesized a majority of them, using an active-learning loop to propose new synthesis recipes whenever a batch produced low yield. The important detail for this article is not the success rate. It is the loop: the system proposed, tested, interpreted the result, and updated. That is a search over a space of configurations with a measurable outcome, and it found things worth having.

The same architecture appears in clinical neuromodulation, where the configuration being searched is a set of stimulation parameters rather than a material recipe. Deep brain stimulation programming — choosing which electrode contacts to use, at what amplitude, frequency, and pulse width — is a combinatorial problem with millions of possible settings and, historically, an exhausting manual search. A 2022 study tested an automated, closed-loop framework that used Bayesian optimization to tune stimulation parameters for tremor suppression in fifteen patients with Parkinson’s disease or essential tremor (Sarikhani et al., 2022). The optimizer converged after testing roughly fifteen to eighteen settings, and the tremor suppression at the best automated settings was statistically comparable to what expert clinicians had achieved. A 2024 study applied a related idea to thalamic stimulation, using a model-based controller to identify optimal stimulation frequencies that manual programming would have found only by trial and error (Tian et al., 2024).

Read carefully, these results support the middle claim in a specific and limited form. Bayesian optimization and active learning can search a large space of intervention parameters more efficiently than a human can, and can find settings that work as well as or better than what a human found. What they have not shown is that the optimizer finds states the human could not have found at all. The search spaces in these examples were bounded by the device, and the outcome measured was a specific symptom, not a general state of mind. The move from “efficiently tune a bounded parameter set toward a known symptom” to “discover a new mode of human cognition” is large, and it is not bridged by the existing results.

Why the space of mental states is not like the space of materials

The materials analogy is instructive partly because it fails in identifiable ways.

A crystal is the same yesterday and next month, to a very good approximation. A brain is not. Neural circuits are plastic; they adapt to their inputs. A stimulation setting that produced a state last week may produce a different state today because the system has changed underneath it. A 2022 review in Nature Reviews Neuroscience catalogued how the effect of identical stimulation depends on the state the brain is already in, drawing examples from sleep, anesthesia, attention, and working memory (Bradley et al., 2022). For a search algorithm, this is a moving target, and the movement is partly caused by the search itself.

A crystal also has one canonical description. A mental state has many, and they do not agree. The same configuration can look one way to a functional-connectivity analysis, another way to a self-report, and another way to a behavioral task score. A 2024 review of how to construct brain-computer-interface therapies for psychiatric disorders is explicit that symptom-states are variable within individuals as well as between them, and that a large part of the difficulty is that the thing being measured and the thing being treated are not the same quantity (Oganesian & Shanechi, 2024). In a materials search, that ambiguity is absent. In a mental-state search, it is central.

And crucially, the objective function for a material is external — conductivity, stability, cost. The objective function for a mind would have to be supplied from somewhere, and the person whose mind it is has only partial authority over what counts as good.

What it looks like when a state is installed rather than found

If the search claim is hard, the “install a state” claim is easier to demonstrate, and there are now striking examples from animal work. These do not show that AI discovered a new human mental state. They show something narrower and important: that a specific internal state — a perception, or a memory — can be written directly into a brain, bypassing the sensory pathway that normally produces it.

In 2025, a team reported in Nature Neuroscience a fully implantable, wireless device carrying an array of up to sixty-four micro-LEDs that could deliver patterned light through the skull to activate groups of modified neurons across the cortex (Wu et al., 2025). Mice learned to interpret a specific spatial and temporal pattern of stimulation as a cue, choosing the correct port in a chamber to receive a reward based on an entirely artificial signal that had no basis in sight, sound, or touch. The animal received information that no sensory organ had delivered, and used it to guide behavior.

An even more direct version of the same idea comes from fruit flies. In 2025, researchers optogenetically replaced both the sensory signal and the reinforcement signal that normally drive olfactory learning, co-activating olfactory receptor neurons and specific dopaminergic neurons (Tumkaya et al., 2025). The flies formed memories of an odor they had never smelled, paired with a punishment they had never experienced. The authors’ finding was that a simple coincident activation of the two populations sufficed — the natural temporal and ensemble patterns were not required. A memory was constructed rather than learned.

Both results deserve the same careful reading. They are real, peer-reviewed, and reproducible enough to be published in serious venues. They are also in mice and fruit flies, they depend on genetic modification that makes specific neurons light-sensitive, and the “state” involved is a narrow one — a sensory cue or an associative memory — not a mode of thought, a mood, or a perspective. The demonstrated claim is that some internal states can be installed with external control in a small number of model organisms. The popular claim, that scientists can now create new kinds of conscious experience, is not what these papers show, and the distance between the two is the distance between a laboratory manipulation and a human capability.

The gap that search alone cannot close

Even granting that a search could be run and could find a state that scored well, three problems remain that are not engineering problems.

The first is that the objective function is a choice, and the search will honor it with precision. An optimizer that is good at its job does not just find the state you asked for; it finds the state that maximizes the score you actually wrote down. If the score is task performance, the optimizer may discover a configuration that produces fluent output while degrading judgment. If the score is self-reported well-being, it may discover a configuration that produces the report without the substance. This is not a bug that better measurement fixes. It is a consequence of optimizing a proxy, and proxies are all we have.

The second is that the discovered state may change the person who would evaluate it. This is the problem that makes mental-state search fundamentally different from materials search. If a person enters a novel configuration, their judgments about whether that configuration is good are made from inside it. No experiment on a crystal raises the question of whether the crystal still prefers its original properties. A 2021 clinical case is the cleanest illustration of the underlying phenomenon at a much smaller scale: a patient receiving electrical stimulation at different brain sites reported distinct emotional responses, and the same stimulation at the same site produced opposite effects depending on the state she was in when it was delivered (Scangos et al., 2021). The value of an intervention depended on the state it was applied to, and the state was changed by the intervention. That circularity is the core difficulty.

The third is that safety cannot be established by the same loop that searches. A search that explores a space of mental states will, by construction, visit states outside the range where the person can reliably report on their own condition. A system that evaluates a state by asking the person how they feel about it has no way to detect a state in which the person would say they feel fine about states they would otherwise reject. Only an external criterion — one that does not depend on the person’s judgment while in the state — can catch that failure, and specifying such a criterion is precisely the unresolved problem.

What a responsible search would have to look like

If one wanted to take the search idea seriously as a research program rather than a scenario, the animal and clinical work above suggests some design constraints.

Start with a measured, narrow target. The tremor-suppression and materials-synthesis successes both optimized a specific, externally verifiable outcome. A cognitive search should follow suit: start with states tied to a measurable deficit — rehabilitation after injury, restoration of a specific function — rather than with a vague aspiration like “insight” or “genius.”

Explore conservatively and reversibly. The BCI literature is clear that the advances come from closed-loop systems that adjust stimulation to the individual’s own measured state, not from fixed doses applied on a schedule (Oganesian & Shanechi, 2024). The same principle applies to any search over states: it should move in small steps, verify that the step was undone, and never enter territory where it cannot return.

Re-consent from the baseline, not from inside the novel state. This is the practical form of the circularity problem. A person who agrees to continue an intervention while under it is not the same evidence as the same person agreeing while at baseline. Documentation of the baseline preference, captured before the intervention and revisited after, is the only way to tell the two apart.

Measure outcomes the person would care about at baseline. Sleep, relationships, judgment, and long-run functioning are the variables most likely to be damaged by an optimization that only sees short-run task performance. They are also the slowest to move, which means a search optimized for quick feedback will tend to ignore them.

Keep the person, not the optimizer, as the holder of the target. This is the reprise of a theme that runs through this series. Intelligence can advise about goals; it does not thereby earn the right to set them, and a system that is very good at reaching a target is not thereby good at choosing one.

What would actually count as discovering a new state

Suppose a decade from now a well-designed study claimed to have found a new useful cognitive configuration. What would make the claim credible?

It would need to show the state is distinguishable from known states by measures the researchers did not choose after the fact, and recurring rather than a one-off artifact of an unusual day. It would need to show the state is reachable by the person without the system, or that the system’s contribution is clearly separable from the intervention that produced it. It would need independent replication in a different lab, on different people. And it would need to show the state does not merely produce a better score on the measure while degrading something the measure did not capture — the classic proof-of-gaming test.

Those criteria are demanding, and they are demanding on purpose, because what goes wrong here is not a null result. A real result can be real in the wrong way: a state that produces the metric, feels good from inside, and quietly costs the person something they would have valued at baseline. The history of psychopharmacology and of performance-enhancing technologies offers no shortage of examples of interventions that met the visible target and eroded something the target omitted.

Where this leaves the question

The three claims have different verdicts.

That the space of reachable mental states is larger than the space we habitually occupy is close to certain, and it is supported by ordinary experience as much as by neuroscience. That an optimization process can search a bounded space of intervention parameters toward a specific, externally measurable outcome is demonstrated, in materials science and in clinical neuromodulation, and it works. That an optimization process could discover broadly useful, genuinely new modes of human cognition — states that are safe, generalizable, and endorsed by the person at baseline — is not demonstrated, and the obstacles are not only technical. They include a proxy problem that no measurement solves, a circularity problem in which the state being evaluated shapes the evaluation, and a values problem about who sets the target.

None of that makes the idea worthless. It makes it a research program with a clear frontier: keep the target narrow and externally verifiable, move reversibly, re-consent from baseline, and hold the objective function open to the person’s own considered goals rather than to the optimizer’s convenience. The optimistic version of this article’s question is a person gaining access to configurations of mind that their ancestors never needed and never found. The pessimistic version is a system that finds, with great efficiency, a state that scores well on someone else’s metric and calls it discovery. Both are consistent with the same technology. Which one obtains depends on a decision that the technology does not make for us.