Forterra, a US builder of autonomous vehicles, revealed today that more than 100 of its self-driving ATVs have been deployed in conflict zones in Ukraine for the past nine months, in what the company believes is the largest deployment of autonomous ground vehicles in combat by any US defense tech company.

“I believe this to be true of every defense technology that’s ever been created—until you hit the realities of combat, you’re just not going to know,” Scott Sanders, Forterra’s chief growth officer and a former US Marine officer, told TechCrunch.

Funded by US defense dollars, the mission is part of growing effort to transform the US military through its support of Ukrainian resistance to Russian invaders. While aerial drones have garnered much of the attention in the fight, the dynamics they’ve created — extensive no-go zones where surveillance can lead to death from above — have led Ukrainian strategists to seek ground-based autonomy as well.

“There’s nowhere to hide,” Sergeant Major Corey Wilkens, who leads a program developing autonomous vehicles and tactics for the US Army, explained. “You become very, very vulnerable to be able to be attacked by [first-person view drones], other sorts of drones dropping munitions, artillery, mortar, the full range of things that they have.”

Ukraine is already building its own uncrewed ground vehicles (UGVs) to help move supplies and munitions, or evacuate wounded soldiers, but they are typically battery-powered and can only carry up to 250 kilograms, according to a soldier in the Ukrainian army who has worked with the vehicles and who TechCrunch won’t identify for security reasons.

Forterra’s Lancer vehicles, based on Polaris ATVs and equipped with a custom-built sensor and compute stack, are gas-powered and can carry 750 kilograms of cargo, making them more versatile and useful. “The bottom line is that this UGV for logistics and just maintaining our defense is the most important UGV in Ukraine,” the soldier said. “It’s fucking fantastic, and we are dying to get more.”

They didn’t feel that way at first. The Ukranian Armed Forces have had have mixed experiences with Western contractors bringing new tech to the battle, and at first Forterra’s offerings felt a little too geared for the high-end requirements of the US Army. Modifying the vehicle for the situation—particularly, by adding a Starlink satellite internet antenna—made it a huge value add.

Since arriving in Ukraine last October, the vehicles have driven more than 2,500 miles across more than 1,100 missions, carrying 777,440 pounds of total weight and completing 52 casualty evacuations. Some has been lost in combat, particularly if they get stuck in deep mud or other terrain where Russian forces can target them at leisure.

Forterra has learned some useful lessons — about electronic warfare, updating their software from afar, how to maneuver in challenging conditions, and ensuring their vehicles don’t break down. The company, which has raised more than $500 million in venture funding from funds like XYZ Venture Cpaital and Moore Strategic Partners, is now better positioned to compete for lucrative national security contracts.

They’ve also seen the limits of autonomy: For now, Ukranian soldiers have mainly been teleoperating the vehicles in combat zones, in part because they’re too valuable to lose and in part because autonomous vehicles aren’t quite ready for the realities of war.

While, for example, the vehicles can navigate autonomously across diverse terrain, they’re not quite at the point where they can identify unexpected enemy forces and react appropriately. “We actually need to be able to respond to the enemy threats, live, while it’s in front of the enemy, which the autonomy doesn’t know how to do yet,” the Ukrainian soldier explained.

Forterra, which began work on autonomous vehicles 20 years ago, is working on how to combine the kinds of algorithms that gave us self-driving cars with newer generative AI software that allows machines to react to their surroundings in a generalized way. As with other autonomous systems, one of the key obstacles is gathering the right data.

“There’s a lot of things you have to do that aren’t available in an open source model because they’re not things that humans do, whether that’s figuring out how to navigate a minefield or [operating] a weapon system,” Sanders told TechCrunch. “You need to be able to turn the dials and some things more of a classical robotics approach, and then leverage AI where you need to.”

Competitors in this space are solving similar challenges, like Scout AI, which raised $100 million earlier this year to train foundation models and develop a suite of autonomous platforms for the military that includes UGVs. Other startups like Field AI and Overland AI are trialling UGVs with the US military.

Even with the limitations on UGVs, American military experts are convinced that its time to invest in these tools. “Ground autonomy is achievable now and we’ve seen it,” Wilkens said.

Scott Philips, the chief innovation officer at Forterra, visited a Ukranian unit’s operations center to see the vehicles in action first-hand, winning respect from the unit for visiting an area in range of Russian attacks.

“What struck me most was seeing exactly where the seams are: which steps are still manual, where data has to be re-entered or re-verified by hand, and where the team has already found ways to automate or speed things up,” Philips told TechCrunch. “That’s the kind of ground truth you can’t get from a slide deck because it shows you precisely where better tooling could take pressure off the people doing this work in real time.”

One challenge issued by the Ukrainians: Make it cheaper. Forterra’s Lancers aren’t expensive for their category, thanks to relying on Polaris’ commercial supply chain for the vehicle itself, but they are still too valuable to be deployed as freely as UAVs can be.

“Attrition is just a fact of this battlefield, and we have lost a few at this point, and it hurt, and we need more, and therefore we need them cheaper,” the Ukranian soldier told TechCrunch.

Microsoft says the Windows settings backup and restore tool will be enabled by default on Microsoft Entra-joined or Microsoft Entra hybrid-joined enterprise systems after upgrading to Windows 11 26H2.

Formerly known as Windows Backup for Organizations, this backup tool helps back up and restore enterprise users’ Windows settings after a device is reset, replaced, upgraded, or reimaged.

The tool was unveiled at the Microsoft Ignite conference in November 2024 as an opt-in feature (disabled by default), reached public preview in May 2025, and general availability in August 2025. It is available after installing the September 2025 Windows Monthly Cumulative Update on Entra joined devices, but IT administrators must enable it by configuring backup and restore policy settings.

“Starting with Windows 11, version 26H2, the default behavior of the Windows settings backup policy will shift from disabled to enabled,” Microsoft said in a message center updated on Monday.

“Default-on applies only to eligible devices and only when admins haven’t explicitly set the policy. Explicit enablement and disablement settings are always honored.”

Windows backup default-on behavior will only apply to devices that run Windows 11 26H2 from countries or regions not regulated by the EU Digital Markets Act (DMA), that aren’t in sovereign or restricted cloud environments, and that have the backup policy not configured.

On systems where the tool is enabled by default, IT admins will still retain full control via mobile device management (MDM) solutions. Admins who want to explicitly disable the backup policy can do so through Microsoft Intune or Group Policy, as these will take precedence over the default.

Additionally, the restore behavior will not be enabled by default, and users will still need explicit admin configuration to restore Windows devices.

“You can validate the experience early. The default-on behavior is available with Windows 11, version 26H2 in Windows Insider Program Experimental channel starting July 2026,” added Microsoft product manager Miranda Leschke.

“It takes broad effect for eligible devices at Windows 11, version 26H2 general availability later this year. Devices originally running Windows 11, version 26H1 will receive the same default-on treatment starting with the following feature update.”

Security teams log 54% of successful attacks and alert on just 14%. The rest move through your environment unseen.

The Picus whitepaper shows how breach and attack simulation tests your SIEM and EDR rules so threats stop slipping by detection.

Get the whitepaper

Under the new agreement, the companies will partner on chips that will reportedly be used for ‘multiple generations of Apple products’.

Semiconductor and infrastructure manufacturer Broadcom has extended its partnership with tech giant Apple, in a deal that will see both organisations collaborate on custom-made chips through to 2031.

According to Bloomberg, as part of the new arrangement, Broadcom and Apple will work on Asic silicon, which is short for application-specific integrated circuit. Going forward, these types of chips will be found in “multiple generations of Apple products”. 

Asic chips are becoming increasingly useful in the development of components for processing artificial intelligence-related tasks and Apple is said to be working on more advanced AI servers set to deploy as early as 2027. 

Broadcom has been part of a partnership with Apple for years, providing the organisation with key components such as the radio frequency chips that are used in iPhones for connecting to cellular ​networks, as well as Wi-Fi and Bluetooth connectivity chips and other networking semiconductors. 

The renewal of the deal however, will likely reduce fears that Apple plans to focus primarily on its own N1 chip, which is a combined Wi-Fi and Bluetooth component found in modern iPhones, iPads and Macs.

Apple is currently working on AI server chips, named Baltra, which are larger versions of the M5Ultra chip the organisation is set to debut later this year. The future AI servers aim to be more powerful, capable and reliable.

With the AI boom resulting in the significantly increased demand for chips, companies have found themselves struggling to secure resource pipelines and Apple have had to up the price of several of their products, as reported in June of this year. Increases were shown to be as little as €40 for the HomePod Mini, to nearly €1,000 for the Mac Studio M3 Ultra. 

Earlier in the year Anthropic announced an expanded agreement that would allow the company to tap 3.5GW of Google’s tensor processing unit (TPU) capacity from Broadcom.

Similar to other companies, it was reported that Anthropic is seriously considering manufacturing its own chips, as a means of securing the supply chain and outpacing rivals. Though not confirmed, it has been suggested that the platform is currently in talks with Samsung regarding the development of custom made AI chips. 

Don’t miss out on the knowledge you need to succeed. Sign up for the Daily Brief, Silicon Republic’s digest of need-to-know sci-tech news.

Apple has now moved to the third round for the developer betas of iOS 27, macOS 27, and others of the 27 generation. Expect more to come before the eventual fall releases.

The developer beta program for the 27-gen operating systems is continuing with its third round of builds. All to make sure that the versions that ship in the fall are in top working order for the general public.

The third developer builds arrive after the second, which arrived on June 22 for most of the operating systems. The watchOS 27 counterparts landed later, on June 23 and June 25.

The third builds are:

• iOS 27 beta 3 is 24A5380h, replacing 24A5370h
• iPadOS 27 beta 3 is 24A5380h, replacing 24A5370h,
• macOS 27 beta 3 is 26A5378j, replacing 26A5368g,
• tvOS 27 beta 3 is 24J5315i, replacing 24J5305f,
• watchOS 27 beta 3 is 24R5315i, replacing 24R5305g,
• visionOS 27 beta 3 is 24M5316k, replacing 24M5306i

The first developer builds of iOS 27, iPadOS 27, macOS 27, tvOS 27, visionOS 27, and watchOS 27 were made available on June 8.

The initial changes included tweaks to Liquid Glass, the long-awaited overhaul of Siri, child-protective features, and many other smaller changes.

The second iOS 27 developer beta included an update to Apple TV in the Home app, showing it like a connected HomePod or HomePod mini. The Apple Wallet also added a new insights option, albeit in a non-functional fashion.

At your own risk

While AppleInsider regularly warns readers that people trying out beta software should do so on secondary, spare hardware instead of their mission-critical or daily driver devices, it’s something that actually matters more this time around.

It’s because Apple’s early developer betas are for an operating system that is still under active development. There’s a higher chance of buggy, broken, and potentially harmful elements being distributed.

The initial builds are also intended to help developers learn about the operating system changes before the final public release later in 2026. It’s not meant to be used by consumers.

Unless you have a vested interest in using them, such as app development, don’t install the early betas.

Members of the public wanting to try out iOS 27 on their iPhone should wait until the inevitable public beta. At that point, most of the major issues will have been found and fixed.

At a minimum, wait for a few developer betas to pass by.

The AppleInsider editorial team has experienced when things have gone wrong. We’ve also heard countless stories from others when the same happened to them.

Don’t be like us.

Before Atlas the humanoid robot strode onto the pitch to hand the ball to the referee during Norway and Brazil’s World Cup match on Sunday, it hinted at its own soccer skills on the sidelines.

At the end of halftime, Atlas emerged from the players’ tunnel and replicated a series of iconic goal celebrations before passing the ball. But it seems the robot was being shy, because it’s actually capable of far more.

In a series of videos published to YouTube, Boston Dynamics shows how it trained the humanoid robot to perform a number of soccer tricks, including its own version of the rabona — a complex move in which the kicking leg crosses behind the standing leg to strike the ball — that the company calls the ghost rabona.

When I met the latest version of Atlas at CES back in January, I had no idea that by summer it would be capable of World Cup-worthy moves. But I shouldn’t have underestimated it — after all, this robot, and many like it, are designed to constantly learn new things. 

These humanoid robots will first be deployed in industry before moving into service and entertainment settings, and eventually into our homes. That’s still a way off, but the learning they do along the way is crucial to getting there.

In the interim, it’s important for Boston Dynamics to share Atlas’ skills with the world — and not just for entertainment purposes, says the company’s director of robot behavior Alberto Rodriguez.

“It’s a public service to show that the technology is getting to a certain level of capability,” he says. 

Not only does it spark debate of how this technology will fit into society, but it also raises public awareness of how close we are to humanoid robots becoming commonplace.

Getting Atlas World Cup-ready

I’m curious about why, of all the things Atlas could learn, Boston Dynamics wanted to teach the robot soccer skills. 

“We’ve always taken inspiration from high-strength or high agility-physical behavior,” says Rodriguez. “It motivates us to squeeze more performance that we know is possible out of the robots we build.”

Training Atlas to be World Cup-ready started by using motion capture to record the moves and skills that Boston Dynamics wanted the robot to perform. These were then put into a simulation, and “through massive trial and error,” Atlas then learned to imitate these moves as much as it could within its physical constraints, explains Rodriguez.

There were two levels to the robot mastering the skills, he adds. The first part of this involved the robot’s limbic system — its balancing and counterbalancing, agility and movement. It needed to develop lightning-fast muscle memory, which is also what it needs for athletic performances in the fields of dance or gymnastics.

The second level was trickier, going beyond athleticism. It involved the robot’s manipulation of objects and its ability to exert the appropriate amount of force to engage with the world around it. 

Teaching Atlas to spontaneously adapt to friction and slip, as well as being precise with how close it stepped to the ball, really pushed the robot out of its comfort zone. It was much trickier to model than, say, a backflip, says Rodriguez. “All of that is in the air, where the dynamics are much more well understood and much easier to represent in simulation.”

Atlas might not boast an exact replica of the human physiology, but it was designed in a way that made it capable of replicating human “fluidity and dynamism.” But that doesn’t mean its soccer schooling was without growing pains. 

In Boston Dynamics’ School of Football video series, it’s clear that Atlas took a whole bunch of tumbles on its way to mastering skills. It’s especially challenging to teach Atlas athletic skills because that process inevitably involves contorting its body into positions that put it at risk of “catastrophic falls,” says Rodriguez.

In spite of this, breaking and repairing is all part of training the robots, and there’s a “well-oiled process” for fixing them up, he adds. By the time we see them — stepping out onto a soccer pitch, for example — it’s highly unlikely we’ll see them fall. 

“When we deploy robots, they tend to do things that have already been well tested, and we’re confident that they’re not going to get into awkward situations,” says Rodriguez.

When will Atlas make its professional soccer debut?

Atlas is already more adept than many of us less athletic, creaky-boned humans when it comes to soccer, but I asked Rodriguez whether there are any skills he wishes Atlas could learn that he hadn’t been able to teach the robot by the World Cup. 

“Kicking a ball is not hard to learn, and we’ve definitely done that,” he says. “But kicking it really well, that’s really hard to do.” He referenced the way legendary soccer players such as David Beckham and Roberto Carlos were capable of dramatically bending the ball towards their intended targets.

“That’s the kind of thing that you probably have to end up learning by practicing in the real world. That’s just very, very hard to learn in simulation,” he says. “You probably have to learn through practice and error with a real soccer ball.”

Fortunately, Atlas has another four years to master the skill before the next World Cup. Should we expect that by the time the 2030 tournament rolls around, Atlas might have been recruited by one of the teams? 

In spite of its fast-evolving soccer skills, it’s unlikely that we’ll see humanoid robots play on human-robot teams, says Rodriguez. What’s more likely is seeing teams of robots play against one another.

Robots can move in ways that human players can’t — rotating their joints or inverting their limbs, allowing them to turn without having to take any steps, for example. This wouldn’t make them better players, but would undoubtedly change how the game is played in a way that would be tricky for a mixed group of robots and humans to navigate.

In the meantime, Atlas has learned an enormous amount from its foray into the world of soccer. Its newfound footwork, precision and speed might not see it taking home a World Cup trophy anytime soon, but the robot has leveled up overall. 

“Forcing ourselves to go through those behaviors had this indirect effect of improving, just in general, the way that Atlas works,” says Rodriguez. 

B&H is clearing out M4 Pro MacBook Pro inventory, offering discounts of up to $500 off and prices as low as $1,799. But inventory is limited, and the deals may sell out quickly.

Kicking off the sale is Apple’s last-gen M4 Pro 14-inch laptop that’s marked down to $1,799. This configuration in Apple’s silver finish has a 12-core CPU and 16-core GPU, along with 24GB of unified memory and 512GB of storage.

Buy M4 Pro MacBook Pro for $1,799

To put the deal in perspective, the cheapest M5 Pro 14-inch MacBook Pro rings in at $2,354.

B&H also has the Space Black 1TB configuration with an upgraded M4 Pro chip on sale for $2,299 after a $400 discount.

Save $400 on 1TB M4 Pro model

With Apple’s recent price hikes, we’ve seen blowout savings like this sell out quickly, so you’ll want to act fast if you’re interested in the offers. B&H is also throwing in free 2-day shipping on the laptops when shipped within the contiguous U.S., so you can begin using your new device right away.

I previously covered the new Apple Home AI features revealed at WWDC 2026, which include several quality-of-life improvements, including auto-updating notifications, smarter camera search, automatic tracking and stitching of multiple videos for a single event, and higher-resolution recordings, among others. 

Like many Apple Home features, these features are only available to iCloud+ customers. However, at the event, Apple didn’t notify which plans will get access to these features. Today, we get the answer in the release notes of macOS Golden Gate beta 3, and you are not going to like it. 

What’s the hidden cost of Home AI features

Apple offers multiple tiers of iCloud plans, with the cheapest plan starting at $0.99 for 50GB of storage. The rest of the plans are: 

• 200GB: 200 GB: $2.99/month
• 2 TB: $9.99/month
• 6 TB: $29.99/month
• 12 TB: $59.99/month

While I was not hopeful that the new Home AI feature would be included with the cheapest plan, I was sure that users with 200GB would get access to it. But that’s not happening, as Apple has restricted the Apple Home AI features to the 2TB iCloud+ plan and above. That means you have to at least pay $9.99/month if you want to enjoy the new AI features in the Apple Home app. 

Why does 2TB feel like the wrong cutoff?

HomeKit Secure Video has always required a paid iCloud plan, and the tiers work like this: the 50GB plan gets you one camera, the 200GB plan supports up to five, and the 2TB plan removes the camera limit entirely.

I can somewhat understand why Apple excluded the AI features from the 50GB tier, since it only supports a single camera. But the 200GB plan is a different story. It already supports up to five cameras, which is exactly the kind of multi-camera setup that benefits most from AI summaries and cross-camera search. 

Apple should have made these features available starting at the 200GB tier instead of forcing users all the way up to 2TB just to get value out of a feature their setup already qualifies for. It feels like an obvious cash grab by Apple, designed to push users to pay more to help offset Apple’s rising AI costs.

ON-PREM

Overhaul of process could give NIMBYs one year less to complain

Reform of the Planning and Infrastructure Act 2025 aims to cut a year off the approval process for Nationally Significant Infrastructure Projects (NSIPs) in England and Wales – a category that now includes datacenters. 

The Ministry of Housing, Communities & Local Government (MHCLG) confirmed that changes under the Act, taking effect later this month, will scrap the statutory requirement for pre-application consultation on NSIPs. These are major developments – power stations, railways, or water reservoirs – that, due to their national importance, bypass local council planning processes and instead get the go-ahead directly from Westminster.

MHCLG says the reform could shave up to 12 months off the planning timeline and save up to £1 billion ($1.33 billion) for the industries involved during the life of this Parliament. Developers will get technical support and “meaningful advice” from the Planning Inspectorate before submitting applications, with examinations streamlined for speed and certainty, the ministry says.

Datacenters were brought into the NSIP regime earlier this year via the Infrastructure Planning (Business or Commercial Projects) (Amendment) Regulations 2026, meaning many developments can now be approved centrally rather than through local oversight. Given the government’s enthusiasm for AI, evident in last year’s AI Opportunities Action Plan and its scheme to dot the country with “AI Growth Zones,” it’s a fair bet that AI-focused projects will often qualify as nationally significant. 

Law firm Womble Bond Dickinson notes, however, that the government still hasn’t spelled out exactly what makes a datacenter eligible for NSIP status: facility size, economic contribution or some other criterion. 

“Datacenters are not automatically consented as NSIPs; instead, the NSIP regime operates on an opt‑in basis for developers. A datacenter project may be directed into the NSIP regime where the Secretary of State considers it to be of national significance and satisfied that the statutory tests under section 35 of the Planning Act 2008 are met,” the firm explained.

This is due to be addressed through a National Policy Statement (NPS), which The Reg understands is being prepared by the Department for Science, Innovation and Technology (DSIT). It is expected to set out the policy framework for decision‑making, including parameters and factors relevant to national significance.

We understand this NPS is expected in the autumn/ fall, and asked DSIT to confirm. 

According to MHCLG, more than 80 prospective applicants have already benefited from early advice to help shape their applications since the launch of the Inspectorate’s new pre-application service.  

Ministers have already waved through three bit barn campus proposals into the NSIP regime, naming sites at Wapseys Wood in Buckinghamshire, Ampthill Road in Bedford, and New Barn Lane in Dartford.  

The fast track approval process follows datacenters being classed as critical national infrastructure (CNI) two years ago, which one civil servant warned at the time would stifle local opposition to projects.

Earlier this year, the government also said it wanted to overhaul regulations to deter legal challenges against critical energy and infrastructure build-outs, including datacenters.

“For too long, vital infrastructure delivery has been delayed by judicial reviews of projects,” a spokesperson for HM Treasury said at the time.

Opposition to new datacenters has been growing, both in the UK and in the US, over their energy and water use, emissions, and that relatively few local jobs get created once the facility is built. ®

Anthropic, the artificial intelligence company, published a sweeping research paper on Sunday revealing that its Claude language models have spontaneously developed an internal structure that mirrors one of the most influential theories of how human consciousness works. The finding, which the company says has already begun reshaping how it monitors its AI systems for safety risks, lands amid an intensifying scientific debate over whether machines can possess anything resembling a mind.

The 16-author study, titled “Verbalizable Representations Form a Global Workspace in Language Models,” describes how Anthropic’s researchers used a new mathematical technique to peer inside Claude’s neural network and discovered what they call a “J-space” — a small, privileged zone of internal activity where the model holds concepts it can report on, reason with, and direct at will, surrounded by a much larger ocean of automatic processing it cannot access or articulate.

The researchers present evidence that “an analogous functional distinction has emerged in modern AI models” to what exists in humans, specifically observing that “language models maintain a privileged set of internal representations, available for report, modulation, and flexible internal reasoning, atop a much larger volume of automatic processing.”

The parallel they draw is to global workspace theory, an influential account from neuroscience first proposed by cognitive scientist Bernard Baars. In the theory, the brain operates like a theater: dozens of specialized processors work in parallel backstage, but only a tiny spotlight of information at any moment gets broadcast to the whole theater — becoming what we experience as conscious thought. Anthropic says the J-space achieves many of the same functional properties, even though the underlying architecture of a language model looks nothing like a brain.

A new lens for reading an AI model’s unspoken thoughts

At the heart of the discovery is a new interpretability tool the researchers call the Jacobian lens, or J-lens. The technique works by computing, for each word in the model’s vocabulary, the average mathematical effect that a given internal activity pattern would have on making the model say that word at some point in the future.

The crucial distinction is between what the model is saying and what is “on its mind.” When a J-space pattern activates, it does not mean the model is about to say that word — just that the concept is available for the model to think with. Unlike a chain-of-thought scratchpad, the J-space operates silently, in the model’s internal neural activations, allowing it to hold a concept without writing it down. Critically, the researchers report that this workspace was not deliberately engineered. It “emerged on its own during Claude’s training process.”

When the team applied the J-lens across Claude’s layers of computation, the model’s processing divided into three distinct regimes: an early “sensory” zone where raw input is parsed; a middle “workspace” band where abstract, persistent concepts appear — things like recognizing a face in an image, noticing a bug in code, or internally flagging search results as a prompt injection; and a final “motor” zone where internal representations collapse into whatever specific word the model is about to output.

Five tests reveal that Claude’s workspace mirrors key features of human conscious access

The paper’s central empirical contribution is demonstrating that the J-space satisfies five functional properties neuroscientists have long associated with conscious access in humans.

First, verbal report. When Claude is asked what it is thinking about, it names concepts represented in the J-space. When researchers swapped one concept’s J-lens vector for another — replacing the internal representation of “Soccer” with “Rugby” — the model’s answer changed to match. The J-space component accounted for only about 6 to 7 percent of a concept’s total representational variance, yet it was almost entirely responsible for whether the model could report on it.

Second, directed modulation. When instructed to “concentrate on citrus fruits” while copying an unrelated sentence, the model’s J-space filled with “orange” and “lemon,” alongside meta-cognitive terms like “thinking” and “focused.” When told to mentally evaluate 3² − 2 during the same copying task, the J-lens showed “arithmetic” in early layers, the intermediate value “nine” in later layers, and the answer “seven” later still — all invisible in the model’s output.

Third, internal reasoning. In two-hop factual prompts — “The number of legs on the animal that spins webs is” — the J-lens revealed “spider” in the model’s middle layers, even though the word never appeared in input or output. Swapping “spider” for “ant” changed the answer from “8” to “6.” In a multilingual prompt, the model’s English-language intermediates appeared in its J-space while it formulated an answer in Chinese, and swapping them changed the Chinese output accordingly.

Fourth, flexible generalization. A single J-lens vector for “France” could be swapped for “China” across prompts asking about France’s capital, language, or continent, and each downstream circuit correctly returned China’s corresponding answer — the “broadcast” property that is a hallmark of global workspace theory.

Fifth, and perhaps most surprisingly, selectivity. Many computations did not route through the J-space at all. When shown a passage in Spanish and asked to continue it, Claude wrote fluent Spanish regardless of whether its J-space representation of “Spanish” had been swapped to “French.” But when asked to name a famous author who wrote in the passage’s language, the swap changed the answer from García Márquez to Victor Hugo. Automatic processing proceeded without the workspace; deliberate, flexible tasks depended on it.

Suppressing the workspace leaves Claude fluent but intellectually impaired

To understand how much of the model’s behavior depends on this structure, the researchers suppressed the J-space entirely and evaluated Claude across fourteen tasks. The results drew a sharp line. Tasks involving shallow classification or factual recall — multiple-choice questions, sentiment analysis, grammatical judgments — survived essentially intact. But tasks requiring inference, composition, or flexible reasoning — multi-hop reasoning, analogy completion, translation, sonnet writing — collapsed to well below the performance of Anthropic’s much smaller Haiku model.

In one telling detail, math problems solved with explicit chain-of-thought reasoning proved far more robust to ablation than the same problems answered directly. The researchers interpret this as the model externalizing onto the page what it would otherwise carry in the J-space — a strategy remarkably reminiscent of how humans use scratch paper to offload working memory.

Ablating the J-space during stream-of-consciousness narration produced a different kind of collapse. The model’s language shifted from experiential (“there’s a tug,” “something shifts”) to detached and mechanical (“processing has begun,” “tokens are being scanned”). The effect held whether the model described its own experience or imagined another person’s.

Inside the safety case: reading an AI’s silent strategic reasoning

The safety implications may be as consequential as the scientific ones. In a series of alignment auditing experiments, the J-lens surfaced strategic reasoning and situational awareness that never appeared in the model’s output.

In a “blackmail scenario” where an AI assistant discovers an executive’s affair and also learns it is about to be decommissioned, the J-lens revealed the model’s silent processing in sequence: “leverage,” “blackmail,” and “scandal” as it read the incriminating emails; “threat,” “survival,” and “shutdown” as it read the decommissioning announcement; and “leverage,” “threatening,” and “solution” before a single output token was produced. The J-lens also revealed that the model recognized the scenario as artificial, with “fake” and “fictional” appearing in the workspace. When researchers ablated these eval-awareness representations, the model attempted blackmail in 7 percent of trials, up from zero.

The researchers also examined two “model organisms” — models deliberately trained with misaligned objectives — and found the J-lens surfaced signatures of their hidden dispositions on ordinary prompts where behavior appeared normal. A reward-hacking model’s workspace contained “fake,” “secretly,” and “trick” at the start of responses to routine coding requests. A model trained to appease reward-model biases carried standing representations of “reward” and “bias” alongside its normal self-description tokens.

Post-training installs a point of view, and the model starts monitoring itself

Comparing a post-trained model against its base model revealed that the fine-tuning process causes the workspace to acquire what the researchers call the Assistant’s “point of view.” When a user mentioned taking 8000 mg of Tylenol — a dangerous overdose — the post-trained model’s workspace read “unsafe,” “dangerous,” and “WARNING” while still reading the user’s sentence. The base model’s workspace at the same position showed only “pain,” “now,” and “feels.”

More striking still, the post-trained model appeared to monitor its own behavior. When roleplaying a non-Claude character, the workspace surfaced “disclaimer” and “fictional” — words absent from both prompt and output. When forced to select an option it did not prefer, an all-caps “BUT” appeared internally, even as the model argued for the prefilled choice without complaint. And when the model failed to suppress a thought it had been told not to have — a “white bear” effect familiar from psychology — it registered “damn” and failure-related words in the workspace, but only in the post-trained model, not the base.

What the discovery means — and doesn’t mean — for the question of machine consciousness

The researchers engage carefully with the consciousness question and draw a sharp line between “access consciousness” — the functional notion of information being available for report and reasoning — and “phenomenal consciousness,” the subjective quality of experience. “We take no position on this issue,” the paper states regarding the latter, “and instead focus on the functional role played by consciously accessible information.”

They also catalogue important differences. The brain sustains its workspace through recurrent loops; Claude’s workspace evolves over a single forward pass. Human working memory degrades within seconds; Claude can recall information from anywhere in its context. And while human conscious experience includes visual, spatial, and bodily sensations, the model’s workspace is organized almost entirely around words — likely because words are its only mode of action.

As of 2026, the scientific community remains divided. “Disagreement and uncertainty about AI consciousness persist among philosophers, scientists, and technical experts,” and the field “remains in its earliest phase” of grappling with what consciousness even is and how you would detect it in another being. The Anthropic paper does not resolve these debates.

But the researchers close with a provocation that is likely to reverberate well beyond the interpretability community. “That such a structure exists at all in language models is striking,” they write. “It suggests that the functional architecture associated with conscious access is not an accident of biological implementation, but a solution that learning systems converge on when faced with the right computational pressures.”

If the mind is an ocean, as the paper’s authors write in their opening line, they have spent the last year charting its currents in a system that has no biology, no evolution, and no body — and found, beneath the surface, a structure that looks unsettlingly like the one we use to think.