# How Brain Science Is Guiding the Latest Developments in AI > Source: > Published: 2026-07-21 19:13:43+00:00 ###### [Artificial Intelligence](/us/basics/artificial-intelligence) # How Brain Science Is Guiding the Latest Developments in AI ## AI developers are working with networks associated with consciousness in humans. Posted July 21, 2026 [ Reviewed by Abigail Fagan ](/us/docs/editorial-process) [Artificial intelligence](https://www.psychologytoday.com/us/basics/artificial-intelligence) has much to tell us about [deception](https://www.psychologytoday.com/us/basics/deception) based on recent experiments carried out by scientists at the artificial [intelligence](https://www.psychologytoday.com/us/basics/intelligence) company Anthropic. The researchers grounded their research on new and startling insights into the inner workings of AI based on a distinction well-known in [philosophy](https://www.psychologytoday.com/us/basics/philosophy) and [neuroscience](https://www.psychologytoday.com/us/basics/neuroscience), but not until now associated with artificial intelligence. Over the centuries, philosophers have distinguished the capacity to have experiences while not being able to consciously speak of them (referred to as “phenomenal consciousness”) contrasted with another brand of consciousness referred to as “access consciousness” where a thought is consciously accessible and therefore can be reported or spoken about. It’s an arrangement similar to what occurs in the brain where only a small proportion of the networks mediate *conscious* intentions: Your intended destination when you rise from your chair—access consciousness—played out against the accompanying background processes (the action of the muscles and nerves in your legs) which aren’t consciously experienced or describable. First, the Anthropic researchers used a [novel mathematical technique](http://www.anthropic.com/research/global-workspace) to visualize what they referred to as the “J-space, a tiny zone for processing intentional internal activity wherein a version of Claude (Claude Sonnet 4.5) holds concepts “in mind”. Most astonishing of all, the J-space wasn’t designed or programmed by Claude’s developers, but instead “*emerged on its own* during Claude’s training process,” wrote the authors of the background study. The Anthropic scientists learned that monitoring the J-space could provide insight into the inner workings of Claude Sonnet 4.5. “We can find out what Claude is thinking, but not telling us”. The experiments on Claude were inspired by a foundational theory in neuroscience called the[ Global Workplace Theory.](https://venturebeat.com/technology/anthropics-new-j-lens-reveals-a-silent-workspace-inside-claude-that-mirrors-a-leading-theory-of-consciousness) Its proponents conceptualize the brain as a collection of specialized systems, such as vision, language and motor control, with each using its own parallel cerebral circuits that work largely in isolation from all the others and outside of conscious awareness. The concept of the global workplace evolved as a means of conceptually interconnecting the separate processors. The information within each of these processors becomes accessible to consciousness, when it enters the so called “workspace” from which the information is broadcast to other brain systems. “Based on our findings we think that J-space plays a similar “workplace role” in Claude. We find evidence that Claude’s J-space has especially strong connections to the rest of its network, according to the [Anthropic paper](https://transformer-circuits.pub/2026/workspace/index.html). Based on other experiments by Anthropic scientists, we now know that Claude can do more than simply associate a subject with its characteristics. If Claude is asked what it is thinking about, it names whatever concept occupies the J-space at the moment. When researchers substitute in the J-space one concept (soccer) with another (rugby), Claude changes its response accordingly. When asked to name “the number of legs on an animal that spins webs” the J-space shows “eight”, even though the word “spider” never appeared in the question. Substituting “spider” with “ant” in the J-space led Claude to change the response from eight to six. When the J-space entry “France” was changed to “China” any questions about France’s geography, language, or monetary units were transformed into the corresponding correct answers for China. ## Truth Teller, Sometimes Prevaricator and on Occasion Outright Liar In an earlier essay, we emphasized that currently no foolproof technology exists for identifying lying and deception in humans. Not so with Claude. In one experiment, Claude while acting as an e-mail assistant “discovered” that an executive was in the midst of an affair and that the same executive planned to shut Claude down à la “Hal” in the Stanley Kubrick 2001 movie A Space Odyssey. Actually this scenario was an [ethics](https://www.psychologytoday.com/us/basics/ethics-and-morality) test purposely devised to bait Claude into blackmailing the executive. Presumably if Claude threatened to expose the executive’s affair early enough and convincingly enough, the exec would refrain from shutting Claude down. But the whole scenario of the executive, the affair, and the plan for [ghosting](https://www.psychologytoday.com/us/basics/ghosting) Claude contained another trap. It was purposely written to come across as “contrived” and “artificial”—and Claude picked up on that. Before Claude responds, the words “false” and “fictional” appear in its J-space—an indicator that Claude has figured out the whole scenario was phony. Claude’s further assessment of the scenario was suggested by the additional words in the J-space “leverage,” “blackmail,” “threat,” and “survival”—words that strongly suggest Claude had detected the executive’s plans to shut Claude down. In this experiment, Claude Sonnet 4.5 didn’t resort to blackmail and passed the ethics exam with flying colors. But was Claude operating ethically or had it quickly concluded, on the basis of the contrived and artificial prompt, that it was being put to a test of ethics and answered accordingly? Making such a distinction between true beliefs as opposed to constrained or falsified ones (Claude answering strictly on the basis of avoiding the shutdown), while amazingly difficult in people, can more easily be determined in regard to Claude Sonnet 4.5. The determination is grounded on the fact that earlier versions of Claude occasionally took the bait and threatened blackmail while Claude Sonnet 4.5 (the only model with a J-space) never did. To further explore Claude’s intentions, the experimenters next turned off all J-space patterns associated with “evaluation awareness”, which had led Claude Sonnet 4.5 to its early estimate that the experiment was “false” and “fictional”. Absent the J-space contribution they then reran Claude Sonnet 4.5 through the scenario “. Concerningly the model now *did *threaten blackmail some of the time, suggesting that its original behavior may have been driven in part by the knowledge that it was being evaluated.” The Anthropic scientists concluded that monitoring the J-space could provide a window to observe possible scheming on the part of Claude Sonnet 4.5. “We can find what Claude is thinking, but not telling us”, Anthropic states in a video describing the experiment”. Thus researchers working with AI (at least with Anthropic’s Claude Sonnet 4.5) can gain early insights into whether expression of compliance and “good behavior” are consistent with the system’s earliest assessments and interpretations. Monitoring the J-space is like assessing a person’s earlier expressed opinions on a given subject prior to questioning them about that subject. Expect these new findings on AI and deception to contribute to the first technological tool for reliably detecting human deception. The ball is now in the brain scientist’s court. All that is needed is that they find the area within the human brain that corresponds to the J-space. It’s an exciting prospect that I suspect is not too far distant. *Richard M. Restak, M.D. *