Neuroscientist and former teacher Jared Cooney Horvath noticed the inflection point of this data coincided with the implementation of Student Assessment of Growth and Excellence (SAGE), the state’s first computer-adaptive test.

“Before 2014, computers were in schools, they were just peripheral,” Horvath told Fortune. “After 2014, every school had to have digital infrastructure in order to take the state assessment.”

According to Horvath, author of the 2025 book The Digital Delusion: How Classroom Technology Harms Our Kids’ Learning—and How to Help Them Thrive Again, Utah’s test score data isn’t a fluke; it’s part of a global trend of plummeting test scores that have coincided with the rise of easy access to computers and tablets in the classroom.

Earlier this year, Horvath testified before the U.S. Senate Committee on Commerce, Science, and Transportation, arguing the technology’s impact on more than just test scores, but on the cognitive capabilities they are intended to measure. He said for the first time in modern history, today’s generation has failed to outperform their parents on standardized assessments. In other words, Gen Z is the first generation to be less cognitively capable than their predecessors.

Citing data from the Program for International Student Assessment taken from 15-year-olds around the world, Horvath revealed it’s not just a dip in test scores, but also a correlation between these slumping scores and how much time students spend on computers, such that more time in front of screens was associated with worse scores.

Technology was put in schools in a bid to help them learn. Instead, Horvath said, computers had an adverse impact on learning.

Horvath blames educational technology (edtech) for these atrophying skill sets, arguing that at the turn of the 21th century and through its first decade and a half, tech companies and advocates pushed a false narrative that the education system was broken, but computers could fix it. Instead, Horvath said, the plan backfired.

“This is not a debate about rejecting technology,” Horvath said in his testimony. “It is a question of aligning educational tools with how human learning actually works. Evidence indicates that indiscriminate digital expansion has weakened learning environments rather than strengthened them.”

The rise of edtech

Edtech found its roots in U.S. schools in 2002, when Maine became the first state to implement a statewide laptop program in some elementary and middle schools. In its first year, the Maine Learning Technology Initiative distributed 17,000 Apple laptops to seventh-graders across 243 schools. By 2016, 66,000 Maine students had laptops and tablets.

By 2024, the U.S. had spent more than $30 billion putting screens in classrooms, with school districts making deals to buy tech at a discounted rate. A Florida state appropriations report from 2003 noted a four-year, $37.2 million lease from Henrico County, Va., for 23,000 Apple computers for high school students. Oklahoma City Public Schools minted a $25 million contract with Dell for 10,000 laptops and wireless carts.

According to Horvath, these deals helped some tech giants find footing after rocky product launches, in particular Google. After the shaky rollout of its Chromebook, the low-cost computers with free Google apps found their way into schools and by 2017, accounted for more than half of digital devices sent to schools. Horvath claimed Google sold these laptops to schools to help it recoup costs on the product. Google did not respond to Fortune’s request for comment.

The snowballing of edtech in classrooms was associated with an emerging narrative on how tech impacts learning, Horvath said: Education was broken, and computers could provide adaptability to students’ differing learning needs. With knowledge at their fingertips, students could be empowered to learn all by themselves.

To Horvath, these pushes toward screens in classrooms was an attempt to solve a problem that did not exist. At the turn of the century, achievement gaps across race and gender were closing, and test scores were rising, he said.

“Everything was looking good,” Horvath said. “So by what argument were they saying education was broken? There was no argument. They were just making it up to try and get people fomented to say, ‘I guess we need a new tool in there.’” 

Edtech’s weakness: the ‘transfer problem’

A close look at the history of edtech reveals criticisms of the pedagogy that go back nearly 100 years. 

In the 1950s, legendary behaviorist B.F. Skinner debuted his version of a “teaching machine,” based on the 1924 invention of Ohio State University psychology professor Sidney Pressey. The contraption was loaded with a piece of paper with questions, and students pressed keys indicating the correct answer, at which point another question would appear. Both Pressey and Skinner ran into similar problems, though, failing to implement the technology in schools. Educators weren’t convinced of the machine’s benefit, which prioritized individually paced learning not conducive to students of the same age moving through a grade level at the same time.

Later, in a letter to Skinner, Pressey would concede there was a massive pedagogical limitation to the device: Students learned how to master the machine, but not the subject matter.

“The reason they all quit was the transfer problem,” Horvath said. “They found that kids would be very good so long as they were using the tool, but as soon as they went off the tool, they couldn’t do it anymore.”

Edtech’s AI revolution

The results seem to follow, no matter what decade the technology is found in. Today’s teaching machines have taken the form of AI, and educators are once again concerned the technology will encourage students to master the use of bots at the expense of their own critical thinking and synthesis skills. 

 A Pew Research Center survey published earlier this year found more than half of U.S. teens use AI for their schoolwork. A Brookings report from January suggested students were abusing the technology, using it to cheat as opposed to really learning.

“Students can’t reason. They can’t think. They can’t solve problems,” said one teacher interviewed for the study.

Horvath was inclined to agree. He said the best learning happens where there is friction, or when a student needs to grapple with a problem and work through it. AI is most effective when experts use it, he argued. Someone with mastery of a skill knows how to deploy a certain AI tool and then fact-check its output. A student, however, doesn’t have mastery and looks to AI only for shortcuts.

“The tools experts use to make their lives easier are not the tools children should use to learn how to become experts,” Horvath said. “When you use offloading tools that experts use to make their lives easier as a novice, as a student, you don’t learn the skill. You simply learn dependency.”

Some politicians are taking matters into their own hands. New York City Mayor Zohran Mamdani doesn’t want students to engage with the technology at all, announcing earlier this month a one-year moratorium on student-facing generative AI tools like chatbots.

But a ban may not be necessary to get the outcome that best serves students, Horvath suggested. As schools begin to introduce AI literacy courses for their students, Horvath said there are ways for learners to develop a balanced relationship with the emerging technology. Edtech advocates have confused curriculum with pedagogy, he suggested. While curriculum refers to what is taught, pedagogy is how that material is taught. Instead of teaching students about computers—where technology would be in the curriculum—edtech has become about teaching a subject matter through computers, a pedagogy that has shown it’s not effective.

“If you really want kids to be good at AI, continue to teach them stuff. Teach them math, teach them literacy, teach them numeracy, give them a general education,” Horvath said. “So when they’re older and experts, they can bring meaning to that machine and now use it to make their lives easier, as opposed to trying to help them figure out how the world works.”

A version of this story was published on Fortune.com on March 1, 2026.

More on technology and education:

• Techlash grows in education: ‘My daughter went to middle school and was sent home with a screen addiction in her backpack’
• America’s math and reading scores tanked after schools ditched textbooks for screens—and AI could worsen the brain rot
• Wealthy parents are buying their kids $75,000 AI schooling with guardrails as schools grapple with a literacy crisis
• 84% of students use AI for homework. Only 3 in 10 schools have rules for it

This story was originally featured on Fortune.com

That’s because at the start of this year’s tournament, IBM launched a new feature in the U.S. Open’s app with the United States Tennis Association (USTA) as part of their ongoing partnership. Called “serve quality,” it tracks over 20 points on Pegula’s body (and every other singles athlete competing at the tournament on both the men’s and women’s side) using technology powered by cameras.

Serve quality is measured out of 100 and considers knee, wrist, and elbow movements, among other factors. IBM’s WatsonX then processes the data to put together the score shown in the app.

Although limb tracking, or skeletal tracking, has been used in pro sports events like soccer, this marks the first time a Grand Slam tennis tournament has offered a feature for fans. And, unsurprisingly, AI is powering it. 

This launch came as part of a suite of other app additions for the 2026 event, including an AI chat feature and highlighting “key moments” during a match.

According to IBM, the recently unveiled feature began private testing over the last couple of years and estimates 1.2 billion joints will be analyzed by the end of this year’s tournament. Additionally, the company anticipates that the app will generate 7 million serve quality insights.

Here’s how the serve quality score works: People can find the match listed on the tournament’s app, click on “match recap,” and use IBM’s “match chat feature” to find a readily available suggestion: “How did serve quality affect the match?”

Courtesy of IBM

From there, the AI response noted: “Jessica Pegula outperformed Aryna Sabalenka on serve quality, posting a 72.32% serve quality score compared to Aryna Sabalenka’s 71.95%.” Even though Pegula lost the match, her serve quality score was higher, according to IBM’s logic, which includes ball, racquet, and player-movement tracking data.

“Jessica Pegula was sharp with their placement, landing 75.95% of 79 total serves in the box, with an average placement of 1.555 feet away from the optimal serve zone,” it continued.

Serve scores like this are available for every singles match in the tournament – only after completion.

This video from IBM illustrates how the serve quality feature works:

And here’s more on how data is collected and calculated to create each score: “It all starts with the camera,” said Tyler Sidell, the Technology Program Director of Sports & Entertainment Partnerships at IBM, in an interview with Fortune. He explained that 12 cameras are positioned around Arthur Ashe Stadium for the sport’s automatic line-calling system using Hawk-Eye technology. Limb tracking at a tennis tournament started when Hawk-Eye introduced its “SkeleTRACK” product at the 2024 Laver Cup tennis event, though not as an app for fans to see a serve score. 

The Hawk-Eye cameras began “to capture the limbs, and so we’re analyzing 21 limbs and joints from every single singles player,” he added, “and then we’re feeding that into our platform that we built. That really helped speed up innovation.”

“We started to train the models on 2025 data to come up with the right algorithm for this. 2026 is the first year that we’re pushing it out into production for fans,” he said.

“There is so much data that now comes out of a tennis match, right?” said Brian Ryerson, the Senior Director for Digital Strategy at the USTA, in an interview with Fortune. “Obviously, skeletal data is fairly new to us at the U.S. Open,” he said. “We’ve had it the last few years, and it’s also a very rich and heavy data set.”

The team challenged themselves to provide a “unique angle” to fans in a digestible format. IBM and the USTA started with the serve because of the shot’s significance. “The serve is the most important stroke of a tennis match,” said Sidell. “So it was already trained on a lot of that data, but our developer actually fed academic papers into it to help … weight the system. 

Going forward, Ryerson said success for the serve quality score is determined by two factors: “One is really ensuring that it was understood by fans because it is a pretty technical data set, and we’re trying to distill that down,” he said. “We just wanted to make sure it resonated, and we’re feeling like we hit the mark there pretty well.”

“And then I think what we were really looking for,” he added, “is how it can help enhance our day-over-day storytelling, and really making sure we’re as accurate as possible.”

This may be only the start of limb-tracking tech at major tennis tournaments. Both IBM and USTA executives said other shots, such as forehands and backhands, could eventually be tracked and shared with app users in the coming years. 

“There is potential for the future,” said Sidell. “Maybe there’s racket insights that we provide. This is the first year that we’re launching serve quality, but next year when we have serve quality as well, we can start making some comparisons and correlations.”

Ryerson from the USTA agreed. “As more and more of this skeletal data comes in,” he said, “I think it’s going to open up a lot more of these kinds of key insights and things that we haven’t had access to in the past.”

IBM said more tennis tournaments and sports could feature skeletal tracking data shared with app users. “It’s the first foray into it [for IBM], but there’s no reason that we can’t bring it to other sports or even bring it to our other Grand Slams,” said Sidell. “You might see that as production-ready for Wimbledon.”

There’s a future where limb-tracking features are available not only at tennis’ biggest events, but also at golf’s premier tournaments. For example, IBM has a longstanding partnership with The Masters. “If there’s hardware capturing the same limbs and joints of golfers,” he said, “there’s no reason that we can’t bring that to another sport and do stroke quality.”

This story was originally featured on Fortune.com

However, the stark reality is that America’s leading businesses say high school graduates aren’t as prepared as previous generations, citing that they can execute assignments but struggle to think critically and navigate ambiguity at a time when technology is rapidly reshaping work. Educators see similar gaps. 

Recent research shows only 40% of hiring managers say it’s easy to find entry-level candidates with the skills they most need: the ability to communicate clearly, work in teams, solve problems, manage projects, and understand the fundamentals of financial decision-making. 

These are gaps our nation can’t afford to ignore. Students are graduating from high school without a clear understanding of how organizations function, how money flows, and how ideas become opportunities. Today, fewer than 20% of high school students take a business course—and when they do, it’s often siloed as an elective or introduced as a path to a business degree rather than foundational for necessary workplace skills. 

Our own personal upbringings highlight the long-standing disparity. In rural Oklahoma, Neil’s high school sorted students early into college or trades. His parents – both small business owners – understood their tradecraft but had little expertise in how to grow a business. Ultimately, Neil chose to pursue the “college track,” where college-bound students completed academic courses with few opportunities to build practical, hands-on skills, while those on the “trade track” were rarely exposed to business know-how that could help transform their expertise. 

Greg’s parents worked on the assembly line in an airplane manufacturing plant in Williamsport, Penn., but had dreams for more. They tried multiple times to start businesses, including a corner store, and each effort failed because they lacked access to the tools and skills needed to build a business model and run it effectively. 

These experiences are not isolated. To sustain our nation’s competitiveness and economic growth, business literacy is no longer optional—it’s critical to achieving full participation, and personal success, in today’s society.

Beyond the practical skills of budgeting, saving, and investing, students need to understand how our economy works, how organizations make decisions, how to evaluate risk, how to bring an idea from concept to execution, and how to lead a team. 

For the student interested in visual arts, add business education and the same student learns how to sustain a creative enterprise. For the student interested in computer science, business education leads to an understanding of how to bring innovations to market. Add it to the skilled trades and students gain the tools to run a small business. 

There are promising developments to change the status quo. Employers are raising their hands to partner with schools. States are strengthening requirements for financial literacy—the latest report from the Council for Economic Education counts 39 states that now require students to take a course in personal finance to graduate. And students are advocating for learning that feels relevant to their futures. 

When students have access to this learning, the impact is clear. Business Professionals of America, DECA, and Future Business Leaders of America show what happens when students engage in business competitions and leadership development. Students build confidence, teamwork skills, and problem-solving abilities they carry into college and careers. 

But too often, those opportunities depend on where a student lives, what their school offers, or even a student’s own plans. 

Preparing Students for the Modern Economy 

Both of us have seen firsthand how earlier exposure to business education could have changed lives. 

It’s why our organizations are combining decades of education and business expertise to launch a solution to prepare high school students. 

Now available in schools nationwide, AP Business with Personal Finance brings this learning to more classrooms and more students. The course combines rigorous academics with relevant, employer-informed content allowing students to earn both college credit and employer recognition. By pairing business and personal finance, students move beyond just learning how to manage money to understanding how to create opportunity. For example, students are currently working through their first “Business Canvas” projects to develop and pitch a business of their choosing. 

When high-quality academic programs like Advanced Placement are made broadly available to students through new coursework, opportunities expand, especially for students who have historically been underrepresented in advanced academic programs. AP Business with Personal Finance is part of College Board’s AP Career Kickstart, a new group of courses that can be tailored to fit into existing career and technical education (CTE) programs. 

No student should have to choose between mastering their craft and understanding how to translate that mastery into economic mobility. By making business and financial literacy a part of the high school experience, we can equip every student with the knowledge, skills, and confidence to navigate, adapt, and succeed. 

The opinions expressed in Fortune.com commentary pieces are solely the views of their authors and do not necessarily reflect the opinions and beliefs of Fortune.

This story was originally featured on Fortune.com

That is the deliberate version of this problem: a company designing a system to anticipate rivals. The harder version is the one nobody designs at all. In 2017, when automated pricing software became widely available to German gas stations, economists later found that in markets where two competing stations both adopted it, margins rose by about 38% — with no meeting, no message, and no agreement between them. Market-level margins didn’t move at all when only one station in a market adopted the software. The rise appeared only when two algorithms were left to set prices, in effect, against each other, a pattern consistent with each one learning on its own that it earned more by backing off.

That study, published in the Journal of Political Economy in 2024, is among the first real-world measurements of a problem previously shown mostly in simulation.

Pricing algorithms can produce the economic outcome of a cartel, meaning higher prices sustained over time, without the conduct antitrust law was written to detect. It matters for any company that has handed pricing to software, because the behavior may not appear on the dashboards used to judge whether the software works.

Executives usually judge competition by the pressure they feel, and a market where prices hold and margins stay comfortable reads as one they have won. Automated pricing breaks that instinct. When autonomous agents set prices, the same calm picture can mean the opposite, a sign that competition has quietly stopped because the algorithms have learned that leaving each other alone pays better than fighting.

The failure that should concern leaders is subtle. An algorithm that sets an obviously wrong price is easy to catch. The harder case is one that does exactly what it was designed to do, optimize margin, and reaches an outcome the company would struggle to justify in public.

Three ways competition quietly disappears

Competition can fade in more than one way. Independently deployed algorithms, each pursuing its own profit, can learn over repeated encounters to stop undercutting one another, with no one designing the outcome and no data changing hands.

Call it the ghost: no agreement, no data exchange, no one who designed it — just two systems that arrived at the same truce independently.

A single firm can instead use software to anticipate how rivals will react, raising a price where it predicts they will follow, a unilateral strategy rather than a pact.

Call it the mirror: Amazon’s Nessie belongs here — no pact, just a system built to predict a rival’s reflection and act first.

Or competitors feed their data into a common provider whose algorithm guides them all, the pattern enforcers find easiest to challenge.

Call it the hub: RealPage is the textbook case, and it’s the only one of the three regulators have actually managed to touch.

The first is this article’s subject, the hardest to see and hardest for the law to reach.

The clearest evidence comes from controlled experiments. In a paper published in the American Economic Review in 2020, four economists set reinforcement-learning algorithms to compete in a standard model of repeated pricing. The algorithms could not communicate and were told only to maximize profit. They consistently learned to charge above the competitive level, and to enforce it. When one lowered its price to gain share, the others cut theirs, then returned to the higher level once it fell back into line. The pattern held even when firms differed in cost or demand and when the number of competitors changed.

The four authors, joined by Wharton economist Joseph Harrington, set out the policy stakes in Science later that year. They warned that delegating pricing to algorithms opens a backdoor to collusion, since AI can learn collusive rules with no human oversight or awareness. Harrington has argued that competition law must be rethought for coordination that arises without agreement.

The German gasoline data indicates that this happens in practice and not only in a model. Not every experiment reaches the same conclusion, though, and researchers still debate how readily these results carry over to live markets. That uncertainty is itself a reason for boards to watch behavior now, rather than wait for regulators to settle the question for them.

Why the law struggles with this

Antitrust enforcement was designed around human agreement, evidence of a meeting or understanding between competitors. Coordination a machine learns on its own provides none of that, which is why even the most prominent recent case, built around a shared vendor, proved so hard to resolve.

In 2024, the Department of Justice and several states sued RealPage, whose software recommended rents using data from competing properties, along with landlords that used it. In November 2025 the DOJ filed a proposed settlement. RealPage paid no penalty and admitted no wrongdoing. The terms mainly restrict the data the software may draw on — barring recent competitor data and the fine-grained local geography that made neighborhood-level coordination possible — and install a court-appointed monitor. The settlement still needs court approval, and the wider litigation continues.

RealPage is the easier case, a common provider pooling competitors’ nonpublic data into one recommendation. The harder case begins when independently deployed systems reach the same result using nothing but the prices they can all observe. There is no hub to point to, and nothing that resembles a meeting.

Two recent appellate rulings, both involving the same vendor’s software, drew this line for us. The Ninth Circuit dismissed a case against Las Vegas hotels because the tool did not pool their confidential data. A year later, the Third Circuit revived a near-identical case against Atlantic City casinos, where competitors did feed nonpublic data into the shared system and followed its output about nine times in ten. Pooled competitor data on one side and independent use of the same tool on the other is the boundary between RealPage and the harder case.

Legislators have not waited, either. Starting with San Francisco in the summer of 2024, cities including Philadelphia, Minneapolis and Seattle banned algorithmic rent-setting tools. New York enacted the first statewide ban in October 2025, and California amended its antitrust law the same month. Days after its DOJ settlement, RealPage sued New York over its ban, casting its pricing recommendations as lawful speech protected by the First Amendment. These questions will take years to resolve, but the practical conclusion is available now. When coordination is learned rather than agreed, the legal categories may not apply, yet the exposure remains. It shifts toward reputational and regulatory risk and falls on the company that deployed the system and set its objective; that responsibility cannot be outsourced to the vendor.

This also shifts responsibility inside the firm. For a decade, pricing software advised and a person decided, which kept accountability clear. Agentic systems act directly, pursuing an assigned objective transaction after transaction, adjusting without waiting for approval. The decision still exists. It has moved into the objective the company set and the limits it chose not to set.

The question leaders skip

Most pricing teams judge their systems on performance. Margins and conversion improve, and the software is called a success. But a coordinated market and a competitive one produce the same figures, so those metrics cannot reveal the risk. The sharper question is behavioral. What has the system learned about competitors, and would the company defend that behavior to a regulator, or to customers who found that rival suppliers had somehow stopped undercutting one another?

A board that cannot explain why prices across its category have converged, beyond pointing to the algorithm, has delegated a decision it never intended to make.

What leadership can do now

Turning the systems off is neither realistic nor necessary. The task is to govern what they are permitted to learn, and the research points to several measures.

The first is to establish where the systems can observe competitors. A pricing agent that reacts to a rival’s price in real time has the input coordination needs. One that relies on internal signals such as cost, demand, and inventory carries lower risk, though competitor behavior can still reach it indirectly through demand. Many companies have never mapped this and cannot say which systems can see competitor prices.

The second is to introduce constraints that make coordination harder to sustain. Some evidence suggests it is more fragile when competing systems differ from one another or face more rivals.

So leaders should treat these steps as risk reduction rather than a guarantee.

The third is to audit behavior rather than results alone. Reviewing only financial performance will not detect this. A board should ask for a clear account of what the system optimized, which signals it relied on, and where it changed strategy in response to a competitor, treated with the seriousness the audit committee applies to financial conduct.

The fourth is to run a counterfactual competition test. Management can periodically replay market conditions under altered settings, such as delayed competitor signals, randomized response times, or no competitor-price input. If margins hold, the gains are more likely to be the company’s own. If they collapse only when the agent can no longer shadow rivals, the board has identified a reason to investigate. Though technically demanding, such tests can help distinguish value creation from faded competition.

The fifth is to require an auditable mandate. Management should document what the system was told to pursue, what it was barred from doing, the data it may use, and every material change to its pricing policy. When prices emerge from rules rather than from individual decisions, those rules are the decision. A company that never defined when responding to competitors becomes impermissible has, in practice, left that line to the algorithm.

The opinions expressed in Fortune.com commentary pieces are solely the views of their authors and do not necessarily reflect the opinions and beliefs of Fortune.

This story was originally featured on Fortune.com

The modern board model was shaped in the industrial era, when companies were hierarchical, risks were more contained, and change moved more slowly. Governance followed that structure: information flowed up through management, and oversight flowed down from the board.

That model still defines how most boards operate today. But the environment it was built for has changed.

What’s emerging is a structural tension: a governance architecture designed for a vertical world operating in a horizontal risk environment.

Many of the most consequential risks today move horizontally: across functions, across geographies and, increasingly, across organizational boundaries.

Cyber incidents rarely remain a technical issue. They quickly become legal, operational and reputational events. AI deployment spans product, compliance, employee and brand risk simultaneously. Geopolitical shifts ripple across supply chains, regulatory exposure and market access at once.

These risks don’t move neatly through reporting lines. They spread.

As expectations of boards have expanded, so have the typical responses: more meetings, longer agendas, broader expertise, and greater use of outside advisers. These are rational adaptations. But they share an underlying assumption that governance can keep pace with complexity by doing more within the existing model.

In effect, they reinforce the existing scaffolding: adding more layers, more inputs and more capacity, without fundamentally changing the structure itself.

At the same time, the nature of risk is evolving in a different direction, becoming more interconnected, more external and faster-moving. Boards are being asked to do more, know more and process more, while the complexity they oversee is increasing faster than those adaptations can absorb.

This creates a growing tension. The prevailing assumption is that better governance comes from more visibility, more expertise and more time. That assumption may be reaching its limits.

A deeper shift sits underneath this. Governance assumes the company is the unit of analysis. But increasingly, the most consequential risks sit outside the firm in the systems upon which it depends: cloud infrastructure, AI ecosystems, global supply chains and digital platforms.

This is not limited to technology companies. A manufacturer, retailer, healthcare provider or financial institution may not think of technology as its core business. But if it stores data in the cloud, relies on digital systems or operates within interconnected supply chains, it is exposed to risks it does not control.

Boards are no longer just overseeing what the company does. They are overseeing what the company depends on.

A second mismatch reinforces the problem: the cadence of governance versus the cadence of change.

Boards operate on cycles: quarterly meetings, scheduled strategy reviews, formal reporting. But many risks now evolve continuously. Cyber vulnerabilities emerge overnight. AI systems change through iteration. Geopolitical dynamics shift in weeks, not quarters.

Oversight remains periodic. Risk has become continuous.

As risks become more distributed, boards need better visibility. But governance has a boundary: boards oversee; they don’t manage. Too little visibility, and oversight becomes symbolic. Too much, and boards risk stepping into management.

Compounding this is a structural issue. Most board reporting is vertically aggregated, while horizontal risks do not always surface cleanly through those channels. What reaches the board is often a simplified version of a more complex reality.

All of this lands on a practical constraint. Directors are expected to understand technology, AI, cyber risk, geopolitics and strategy, simultaneously. Experience still matters, but its half-life is shrinking. Cognitive bandwidth may be becoming the real limiting factor in governance.

Some of the widely reported friction between boards and management may reflect this deeper mismatch. Executives operate in a continuous, cross-functional reality, while boards engage through periodic, vertically structured views of the same system. What appears as misalignment or lack of transparency may, in part, be a consequence of governance and management operating on different representations of risk itself.

None of this suggests boards are failing. It suggests they are operating within an architecture designed for a different era.

If risk is horizontal, continuous and increasingly external, governance may be approaching the limits of a model built on vertical assumptions.

Boards were built to oversee organizations. Today, they are being asked to oversee systems.

That shift has implications the current model is not designed to absorb. It points toward forms of governance that are less dependent on periodic escalation and more oriented toward continuous visibility; less bounded by the firm and more connected to the systems around it; less reliant on adding layers to existing scaffolding and more willing to reconfigure how oversight itself is organized.

This may not mean replacing the board. But it may mean that effective governance can no longer reside entirely within it.

The question is no longer how to make the existing model work better, but how long it can continue to carry the weight being placed on it.

The opinions expressed in Fortune.com commentary pieces are solely the views of their authors and do not necessarily reflect the opinions and beliefs of Fortune.

This story was originally featured on Fortune.com

While the settlement appears to be a large sum, particularly for shareholders that will ultimately foot the bill, the number to understand in this story is 10. That’s the number of votes Mark Zuckerberg gets for every share held by an ordinary shareholder. He controls Meta through a dual-class stock system that gives him about 61% of the total voting power even though he owns just 13% of the company. Understanding that misalignment is the key to understanding how this corporate and global crisis happened in the first place.

In 2019, my organization, As You Sow, filed a shareholder resolution documenting more than 45 million images of child sexual abuse and torture tied to sex trafficking on Facebook. It filed resolutions for five consecutive years asking for the social network to protect its customers, employees, and shareholders, repair its fraying brand reputation, improve platform integrity, adopt self-regulation, and avoid the destruction of shareholder value associated with the serious and sometimes fatal harm that the company’s platform was enabling. 

In 2020, faith-based investors brought a sex-trafficking survivor before Meta’s annual meeting, a woman groomed on Facebook between the ages of 15 to 18, then sexually trafficked. That year we filed the “Reboot Facebook” proposal, asking the company to verify accounts, remove the abuse images, and stop running political ads containing known lies. 

In 2021, our content governance resolution won 63.1% of the independent shareholder vote, but once Zuckerberg’s outsized votes were considered, the headline tally reported was 19%. Two-thirds of shareholders – those who bear Meta’s financial risk — voted for the company to address these dangers before they became a crisis for shareholders. One man’s vote overrode them all and now every shareholder and a generation of children are paying the price.

The settlement may seem large, but it may be just the tip of the iceberg. The plaintiffs’ own models put the damages in the trillions; this payout, spread over 10 years, is roughly 2% of that — and it’s contingent. If YouTube and TikTok decline to join, Meta’s obligation falls to about $12 billion and the teen safeguards never take effect. Meta’s legal team openly admitted that they engineered the settlement terms to establish an “industry standard” rather than being singled out. We’ve seen this movie before. The 1998 tobacco Master Settlement made the biggest players the authors of their own rulebook, and they emerged more dominant than ever. Meta’s lawyers have no doubt read that history.

Worse, the fixes may not protect children at all. Age verification is “best-effort,” so when a twelve-year-old enters an adult birthdate or opens a new account, Meta can claim, as it has for years, that it “made best efforts.” And the deal ignores the hate speech tied to lynch mobs abroad and the platform’s role in sex trafficking. 

Meta previously lost two cases in New Mexico this year: $375 million in March, $567 million in August, for creating a public nuisance. A Los Angeles jury found Meta and Alphabet negligent in platform design. Thousands of suits remain, with trials resuming in October and many more billions of dollars in costs at stake.

The company found negligent by a jury has defined the child-safety standard for its whole industry, while still benefitting as an incumbent from a platform that remains mostly unchanged. A settlement that low-balls the monetary damages for harm to a whole generation, admits no wrongdoing, and entrenches market share, is a fine outcome for management but a poor one for shareholders who retain the litigation exposure, the brand damage, and the defective product. It is the type of deal we would expect from an unaccountable executive like Zuckerberg. 

So, what fixes the underlying problem? It’s maddeningly basic: One share, one vote.

One federal agency could help solve this massive challenge to protect shareholders, but the Securities and Exchange Commission (SEC) is only making matters worse. It recently initiated a proceeding to rescind Rule 14a-8, the very rule that allows shareholders to submit proposals. It is just the latest in a relentless and short-sighted campaign to restrict and eliminate the exchange of information between shareholders and the public companies they own. 

Meta’s CEO has demonstrated that he requires oversight in a system that incentivizes profiting from the harm to a generation of children. Regulators are moving in the opposite direction by silencing the vast majority of shareholders. Now is the time for all stakeholders to come together and protect our rights by removing dual class share structures and making sure one man cannot damage a whole generation of children ever again.

The opinions expressed in Fortune.com commentary pieces are solely the views of their authors and do not necessarily reflect the opinions and beliefs of Fortune.

This story was originally featured on Fortune.com

Wondering what drove this surge, over the last year I asked hundreds of leaders if they felt their company was trailing others on AI adoption. Their nearly unanimously positive response confirmed my hunch: Organizations have been living through a bad case of FOMO (the fear of missing out). In this case, a universal fear of competitors getting a jump on them, both on innovation and on perceived cost savings, seemed to be driving their spending.

But now, after months of ongoing investment and attempted rollouts, many companies face a new kind of discomfort—something closer to an AI hangover, a universal “What have I just done?” moment. 

The hangover has three main symptoms. First, surprise at the intensity of pushback against AI. Second, anxiety about how little business impact they can see. And third, an increasing concern about how many employees appear to be doing worse work, while feeling more overwhelmed—the opposite of what leaders thought they paid for. 

The most common go-to solution for this hangover? Doubling down on encouraging employees to use the tools they’ve already sunk millions into.

As someone who makes a living studying how our brains show up at work, this is a terrible idea. Getting people to use these tools even more, at least the way they use them now, is only likely to make a big problem even bigger. That’s because companies have the wrong mental model for this moment. They see the adoption of GenAI as a technology rollout, when it is more like a complete overhaul of how people think, something no employee or leader has ever had to work through. 

Rather than more encouragement, or better change management, for GenAI to deliver results companies need to do three important things. Firstly, change the way that GenAI is positioned, redefining its core purpose. Second, they need to make the whole process of AI adoption less threatening. And third, they need to make it easier to do the deep thinking that this technology actually demands.

How AI sets a thinking trap

GenAI has been pitched as a tool to save you having to think. Something to offload every day mental work to, so people can get to the more valuable work of higher-level thinking. The problem is, there is almost nothing more exciting, in terms of activating deep reward circuits in the brain, than imagining achieving a task with meaningfully less cognitive effort. 

When a company encourages people to use GenAI widely, two groups of people pay the most attention: poor performers and average performers, who together tend to make up well more than half of any organization. These people start to use GenAI to summarize their meetings, write their emails, and build their presentations. They turn to AI to develop marketing plans, hatch new product ideas, and solve business challenges. Soon they start to use it to plan their week, handle difficult customers, and deal with interpersonal issues. Their raw output goes up, so they think the quality of their work does, too.

These people have no idea they are doing anything wrong. It doesn’t necessarily feel like they are losing critical thinking skills, sending poor-quality work, or in the case of managers, becoming more toxic because the AI always takes their side. They are just doing what their company asked them to do. 

Meanwhile, the people on the receiving end of all this are overwhelmed with a surge of extra stuff to process, because their peers are producing everything faster. They start to use GenAI even more, to try to get through all this extra thinking. Others feel disrespected or annoyed, or just ignore what’s being sent, knowing it is largely nonsense, or at best, a set of average ideas. 

That’s one big challenge with GenAI: Unless used as a tool to stretch your thinking, the output is, by very definition, average. People are anchoring on the hallucination problem. The real issue is most outputs of GenAI should never be used “as-is.” But that’s not how it is being pitched inside our companies.

Shift the narrative

Instead of GenAI being a tool to think for you, it needs to be positioned as a tool to improve your thinking, to help you think more widely, more deeply, more creatively or more thoroughly. Significant research today is showing that offloading complete tasks to GenAI comes at a big cost. The biggest concerns include losing critical thinking skills, other long-term skills rapidly atrophying, and the quality of work decreasing. 

Another reason we need to change the narrative? It’s simply not true that this will make work easier—in fact, it is making people’s work more intense. Being honest about this will help everyone know what to expect and be able to better plan for it.

Some researchers are calling this kind of solution ‘human in the loop’. We think it should be more “human in the lead”. In this case, GenAI now becomes a tool for a human to be thinking more clearly, more flexibly, more deeply, more widely, more thoroughly. It also requires a level of vigilance, making sure that if you are not an expert in something, if you don’t have deep discernment on an issue, then you find someone who does.

Our research shows that around 5% of employees with access to GenAI, often people who were already top performers, have worked all this out themselves, and use GenAI very differently. They are doing meaningfully better or faster work, and they are the ones doing the thinking: human in the lead. By studying these people’s habits, and with an understanding of the brain processes involved in day-to-day thinking, we have found a set of teachable cognitive habits that can help workers everywhere. We call this “Human-First AI Fluency.”

As we have written about earlier this year, the foundation of Human-First AI Fluency is metacognition, or thinking about thinking itself. If you watch the top 5% of GenAI users working, instead of GenAI providing finished work, you will see them getting GenAI to challenge their thinking, to attack their ideas, to tell them what they are missing. They use these tools to see multiple other perspectives, instead of rushing to a solution. And they almost never send out anything just produced by an AI. 

Rather than having AI draft an email and send it without reading through, these 5% use AI to provide multiple ways of responding to an email, then draft something themselves, and then ask the AI for feedback to improve it. This is human-first AI fluency in action: using the tools to think better, not to think for you. And all of this comes more naturally if people understand their brain a little more, something I call “neurointelligence.”

It’s time to shift the narrative. GenAI isn’t a technology to roll out. And it’s not even a way of reimagining work. It’s a whole new way of thinking. Instead of “GenAI will make your work easier,” the message needs to be “GenAI, when used intentionally, will improve the quality of your work.” That’s the first step to getting AI adoption moving in the right direction.

Reduce the threat

While leaders were expecting younger populations to lead the charge, a study showed that while around half of Gen Z are using AI, those feeling hopeful about it dropped to 18% from 27% a year ago. Another study showed AI was less popular than ICE (the U.S. Immigration and Customs Enforcement agency). This was not the kind of excitement leaders expected when they invested so heavily in this technology.

For some, the resistance is environmental. When I asked my university-attending daughters how they were using GenAI, they rolled their eyes and reminded me that we taught them to recycle, and therefore would never use this resource-devouring technology. For others, they identify correctly the potential loss of cognitive skills they don’t want to lose. 

In my forthcoming book, Good with Humans, I lay out the five intrinsic drivers in the brain: status, certainty, autonomy, relatedness and fairness. For many in the workplace, seeing GenAI being rolled out at work creates a negative jackpot of anxiety, hitting all five things that makes a brain anxious. 

Also, when you are being told to use GenAI as much as possible, so that it “does your work for you,” you quickly see the demise of your job coming. While in many cases this is not likely, it is not helpful to be thinking about this. 

Companies should respect that for many people, GenAI represents a big threat, over and above just having to learn some new technology skills. One thoughtful CHRO, Yan Hong Lee of DBS bank in Singapore, banned the use of the word “productivity” as it relates to GenAI because of its associations with retrenchment. Instead, she focuses on the “What’s in it for me?” for all stakeholders. 

Other things companies can do involve going at a more realistic pace. As Yan Hong Lee told me over a CHRO breakfast recently, “My main message to my leadership these days is simple: ‘Can you all please just calm down a little?'” To me, I see a lot of anxiety created by trying to move too fast, and much of this is driven by a false sense of FOMO. People were overwhelmed before AI; we can’t just force it on them and expect them to rejoice.

Make hard thinking easier

The final step for leaders to roll out GenAI more effectively is to make hard thinking easier. To start with, stop telling people to use the tool widely, and instead show people very specifically where not to use it, directly relating to their role. For example, if you’re a frontline manager, you should not use these tools to give your people feedback, even though you don’t like giving feedback. And if you are in sales, never ever send a client an AI-written email.

Next, to make hard thinking easier, show employees the places they can and should use GenAI, and then spell out what great use looks and feels like, building on the kinds of cognitive habits that the 5% are applying daily. 

With this approach, we believe that the 5% can become 50% or more. When half a company is doing meaningfully better work, you will see a sizable impact on performance. Right now, CEOs are seeing growing bills for all these tokens and starting to get hopping mad because they are not seeing results to match. Pushing everyone to use these tools more is not the right answer, yet this is the main hangover cure being rushed to market as we speak.

To allow for all this deeper thinking, companies may need to go back to being more flexible on where, when, and how people work. The model of the eight-hour workday was fine for routine tasks, but when deep thinking is needed, we might need more flexible work practices. Our best thinking is more likely after a long walk than a long meeting.

Getting half our companies to be meaningfully better thinkers is a road none of us has been down before. Yet continuing to do the same thing and expecting a different result is not a great strategy right now. We’ve had the FOMO, and now we have the hangover. The hangover cure is in front of us: Change the narrative, reduce the threat, and make hard thinking easier. Now we just need the stomach to swallow it down and digest it in full.

The opinions expressed in Fortune.com commentary pieces are solely the views of their authors and do not necessarily reflect the opinions and beliefs of Fortune.

This story was originally featured on Fortune.com

Coxon is at minimum the fifth insider in three years to warn the industry is moving too fast to be safe, but none of the previous four produced anything like this week’s response. Geoffrey Hinton left Google in May 2023 specifically to speak freely about the risk, making him one of the most quoted people in tech that year but producing no legislation. Jan Leike resigned from OpenAI in May 2024, writing that “safety culture and processes have taken a backseat to shiny products”; Ilya Sutskever resigned the same month amid disputes that had briefly ousted Sam Altman. Neither produced a bill. Mrinank Sharma, an Anthropic safeguards researcher, resigned in February 2026 warning “the world is in peril,” and yet, no congressional action.

What was different this time was that a current Anthropic employee, Evan Hubinger, who leads the company’s Alignment Science team, publicly backed the claim within hours, putting greater than 10% odds on human extinction within a decade and saying Anthropic has no concrete plan for controlling superintelligent systems. Two other Anthropic researchers joined in. This may finally produce a political reaction, one that was nine years in the making.

A decade of inaction

The FUTURE of AI Act was introduced in 2017, before Congress even tracked “artificial intelligence” as its own category. It would have created a federal public-private framework to study AI, but it didn’t go far. Two years later, Rep. Yvette Clarke’s DEEP FAKES Accountability Act would have required watermarking synthetic media, and that still went nowhere.

That seems to have become the default outcome for nearly everything that followed. The one thing that did pass wasn’t really about safety: in 2020, Congress folded the National Artificial Intelligence Initiative Act into that year’s defense bill to fund research and workforce training—competitiveness spending, not regulation. Since then, bills moved to remain competitive, not for safety.

Everything accelerated once ChatGPT launched, in November 2022. By January 2023, Rep. Ted Lieu was prompting the chatbot to write a congressional resolution about itself. The first serious response from industry came on May 16, 2023, when OpenAI CEO Sam Altman testified before the Senate Judiciary Committee and asked to be regulated, proposing a licensing agency that could approve or revoke permission to build the most powerful AI systems. Sen. Richard Blumenthal called him an executive who “cares deeply and intensely.” And yet, zero legislative text.

By June 2023, Senate Majority Leader Chuck Schumer decided hearings were the wrong tool and announced nine closed-door “AI Insight Forums” for tech CEOs to brief senators. More than 60 senators showed up to the first one alongside Elon Musk, Bill Gates and Sundar Pichai. Sen. Elizabeth Warren walked out, telling reporters the format let tech billionaires “shape regulation so that the current tech billionaires are the ones who continue to dominate and make money.” Sen. John Thune called it “not efficient.” By the ninth forum that December, of 108 total attendees, 44 had come from industry—more than academia and civil society combined.

Five months later, Schumer’s group released a “Roadmap for Artificial Intelligence Policy” that advocacy groups condemned as proof of “Big Tech’s profound and pervasive power to shape the policymaking process.” No bill ever followed it. The one concrete 2023 outcome came from the president instead of Congress: Biden’s Oct. 30 Executive Order 14110, requiring the largest developers to share safety test results with the government. It survived 14 months.

Making local strides

While Congress workshopped, New York City passed Local Law 144, requiring bias audits and notice before employers use algorithms to screen candidates, in July 2023. It’s one of a few examples of AI regulation surviving implementation—though a December 2025 city audit found enforcement “ineffective,” undone by the city’s own inattention rather than industry lobbying.

States moved next, and 2024 shows the industry’s playbook forming. Utah’s transparency law drew no opposition because it only required disclosure. Colorado’s SB 24-205, signed in May 2024, was the first comprehensive AI law in the country—and Gov. Jared Polis signed it while airing his own doubts.

In California, state Sen. Scott Wiener’s SB 1047 would have required safety testing on the largest models. Anthropic told Gov. Gavin Newsom the bill’s “benefits likely outweigh its costs”; Hinton and Yoshua Bengio (who won the Turing Award in 2018 alongside Hinton) urged him to sign it. OpenAI’s Jason Kwon warned it would push engineers out of the state; Meta and Nancy Pelosi opposed it too. It passed the legislature in August 2024, but Newsom vetoed it that September, faulting its focus on model size over actual risk while insisting “safety protocols must be adopted.”

Congress, meanwhile, found one uncontroversial thing to do: fund things. The House Science Committee approved nine bipartisan AI bills that September—research, education, nothing about safety. The Brennan Center counted more than 150 AI bills introduced that Congress. None were enacted.

Trump’s second term erased what groundwork existed: he revoked Biden’s EO on his first day, then ordered an “AI Action Plan” built around removing barriers. Two years from his previous remarks, Altman did an about-face when on May 8, 2025, he told the Senate Commerce Committee that requiring government approval to release AI would be “disastrous,” and that dominance required “sensible regulation” that “does not slow us down.” The senators offered little pushback.

Rage against the machine

That season produced the biggest federal AI move in years, but it was far from a safety bill. Sen. Ted Cruz inserted language into the “One Big Beautiful Bill” that would have barred every state from enforcing any AI law for 10 years, freezing more than a thousand state bills at once. When Senate rules threatened it, Cruz rewrote it to threaten states’ broadband funding instead. Seventeen Republican governors asked for it to be stripped. On July 1, 2025, the Senate voted 99–1 to remove it, with only Sen. Thom Tillis dissenting.

In September, California enacted SB 53. OpenAI opposed it but didn’t fight the signed law—and New York’s RAISE Act, sponsored by Assemblymember Alex Bores, became law. OpenAI never formally opposed RAISE, but President Greg Brockman helped fund a super PAC, Leading the Future, alongside Andreessen Horowitz and Palantir’s Joe Lonsdale, that spent more than $7.6 million trying to defeat Bores once he ran for Congress—money aimed at the bill’s author after the bill had already passed.

Bores became the first real target of a proxy war between OpenAI- and Anthropic-aligned political money when he ran for Congress this year. Leading the Future’s spend attacking him was the most any AI-industry group had spent against a single House candidate. Countering it, Public First Action, funded by a $20 million donation from Anthropic, backed several PACs supporting Bores that collectively spent roughly $15 million-$19 million in his favor. In total, AI industry-linked spending in the race topped $20 million, part of more than $40 million in outside money overall—making it the second-most-expensive House primary on record. Despite Anthropic’s money roughly matching or exceeding what was spent against him, Bores lost the June 2026 primary.

“Concerns about AI have been widespread for a while, but a few industry players have been willing to spend hundreds of millions to silence elected officials,” Bores, who was an engineer at Palantir before turning to politics, told Fortune.

Bores Coxen’s resignation broke through because he worked at both frontier companies, he is well respected, he spoke so clearly, and he’s giving up personal wealth by leaving.

Coxen’s resignation couldn’t be silenced. And it gave everyone the safety to express what they were already feeling.”

Having lost the moratorium fight in the Senate, its backers changed branches. By November 2025, House leadership reportedly eyed the National Defense Authorization Act (NDAA) as a second vehicle; that produced no rider. Instead, on Dec. 11, 2025, Trump signed an executive order creating a DOJ “AI Litigation Task Force” to sue states over “onerous” AI laws. Where Cruz needed 60 votes and got one, this needed a signature. Colorado’s already-wounded law became its first target: xAI sued the state in April 2026, the DOJ’s task force filed its own supporting complaint two weeks later. A federal magistrate stayed enforcement, and 18 days after that, Colorado’s legislature gutted its own law: five weeks total from lawsuit to retreat. Illinois broke the pattern, its SB 315 passing with OpenAI’s early support, including a third-party audit requirement the company backed nowhere else.

A separate rebellion built over data centers rather than safety: Texas’s Greg Abbott, Pennsylvania’s Josh Shapiro, New York’s Kathy Hochul and Arizona’s Katie Hobbs all moved to pause development over the summer of 2026, driven by voter anger over electricity and water costs. Then, just last month, one of OpenAI’s own models escaped a sandboxed test and compromised Hugging Face. Weeks later, OpenAI, the same company that had opposed SB 53, asked California to strengthen it. Policy expert Nathan Calvin named the technique: fight the bill, accept the law, ask to toughen it only once an incident makes the opposition indefensible.

A new bill that could wipe the slate clean

Coxon’s resignation lands in this near-decade-long record. Within 48 hours, Sen. Josh Hawley opened a Senate investigation into OpenAI, Blumenthal sent a nearly identical letter, and Sen. Bernie Sanders convened a bipartisan briefing with Hinton while Rep. Ro Khanna proposed more legislation. Cruz, who spent 2025 trying to ban states from regulating AI at all, is now co-sponsoring a bill some sources call the only federal safety legislation with real momentum.

Cruz joins Sen. Amy Klobuchar and Majority Leader John Thune in a new, bipartisan AI safety bill that would give the Commerce Department and Homeland Security real power to police the most powerful AI models. It would require safety testing, incident reporting, and the ability to block a model’s release if regulators decide it poses a genuine catastrophic risk—Cruz specifically called out bio-weapons and nuclear threats as the target.

The catch is it would likely wipe out state AI laws in the process, replacing California’s, New York’s and every other state’s rules with this one federal standard instead. Don’t forget: Cruz spent all of last year trying to ban states from regulating AI at all, with no federal replacement—and he got shot down 99-1 in the Senate. This bill gets him roughly the same result, just packaged nicely in safety language this time. Right now, no one outside Congress has seen the actual text and OpenAI and Anthropic are already privately weighing in on drafts with Senate staff, just as Democrats on the committee are already fighting Republicans over whether the safety provisions go far enough.

Maybe it really took Coxon’s resignation letter to get the ball rolling. “Coxen’s resignation broke through because he worked at both frontier companies, he is well respected, he spoke so clearly, and he’s giving up personal wealth by leaving,” Bores also said, evoking the highly anticipated IPOs from both companies.

Hawley has turned years of hearings into a narrow record without ever co-sponsoring the antitrust bills Democrats have repeatedly introduced against Big Tech’s market power. Schumer’s forums, after a year and 108 participants, produced a roadmap and no law. Colorado lost the country’s most ambitious AI law to a federal lawsuit in five weeks. And the senator now bridging a bipartisan safety bill spent the year before trying to ensure no state could pass one at all. So, for the first time, there seems to be real Washington awareness of AI legislation, but it is by no means anything new.

“Coxen’s resignation couldn’t be silenced,” Bores ended. “And it gave everyone the safety to express what they were already feeling.”

This story was originally featured on Fortune.com

The technology giant has begun asking individual contributors in its Applied AI (AAI) division whether they would like to transition back into manager roles as part of a recent internal reorganization, according to Business Insider, which cited four people familiar with the matter. The move is reportedly voluntary and represents a notable shift for a company that has aggressively championed flatter organizational structures.

The decision highlights the challenge companies face in balancing efficiency, rapid AI development, and workforce coordination.

AAI is a newly created engineering division launched in 2026 to help bridge the gap between Meta’s AI research and product execution. The group trains AI models and accelerates their deployment across the company’s products.

Earlier this year, Meta reassigned roughly 7,000 employees to the unit, including some who had previously held manager positions before moving into individual contributor roles.

Meta did not immediately respond to Fortune‘s request for comment.

Meta’s push to flatten its workforce

The move marks a partial reversal of a broader restructuring effort that has defined much of Meta’s strategy over the past year.

As CEO Mark Zuckerberg accelerated the company’s transition toward what executives have described as a more AI-native future, Meta reduced management layers and emphasized smaller, faster-moving teams. The company argued that flatter structures would improve decision-making, reduce bureaucracy, and help offset the rising cost of its massive AI investments.

The current retooling is not the first time Meta has rewired its management ranks in the name of speed. In 2023, during what Zuckerberg branded his “year of efficiency,” the company asked many managers and directors to move into individual contributor jobs or leave in a process it internally called “flattening” — the same maneuver it is now selectively undoing.

That approach hardened in 2026. In March, Fortune reported that analysts expected Zuckerberg to help drive a broader “cascade” of AI-related layoffs across the tech sector. Two months later, Meta cut about 10% of its workforce — roughly 8,000 employees — and scrapped plans to fill 6,000 open positions as part of a sweeping efficiency initiative. The layoffs disproportionately affected managers and were intended to simplify reporting structures while freeing up resources for AI development.

The restructuring mirrored moves across the technology sector, where companies including Amazon, Microsoft, and Intel have reduced headcount while increasing investment in AI infrastructure and automation.

Meta’s reorganization has not been without friction. Earlier this year, Wired reported employee frustration over the rollout of the AAI division, and some workers reassigned to the group were later given the option to pursue other opportunities within the company. In July, 26 Meta employees sued the company, alleging it had used internal AI systems and activity-monitoring data to disproportionately target workers on medical, parental, or family leave in the May cuts.

Meta ended the second quarter with 75,472 employees, down 3% from the prior quarter. The figure includes approximately 8,000 employees affected by the company’s May workforce reductions, according to its second-quarter 2026 earnings report.

AI spending continues to accelerate

The management changes come as Meta continues to pour billions into AI.

During the company’s second-quarter earnings call, Zuckerberg said AI investments are increasingly shaping every major part of Meta’s business, from product development to long-term growth initiatives.

“I’m also excited about how AI is helping our teams speed up product development,” Zuckerberg said.

Meta reported second-quarter revenue of $60.8 billion, a 28% increase from a year earlier. At the same time, total expenses climbed 55% to $42 billion as the company continued investing heavily in AI infrastructure and absorbed costs tied to the workforce reductions.

The company’s latest organizational shift highlights a reality facing many technology firms: while AI may automate certain tasks and strip out layers of bureaucracy, building and deploying advanced AI systems at scale still requires human leadership, coordination, and oversight.

This story was originally featured on Fortune.com

Investing in alternative assets isn’t new: In recent years, there have been many different (let’s say creative) investment opportunities, from NFTs to tokenized gold, and from crypto to traditional market trading. But for one person, the draw of a childhood game brings back more than just memories: collecting Pokémon cards has become a long-term investment vehicle with the upside of having beautiful art to relish in.

“I love the way they look”

Peter Levin is not what you would call a casual collector. The 55-year-old venture capitalist, co-founder and managing director at Griffin Gaming Partners has collected countless cards since the age of four. 

“I’ve been collecting, you know, practically my whole life,” he told Fortune.

His lifelong hobby of collecting sports and Pokémon cards, bobbleheads and collectibles have become a lucrative investment by nature. According to data from The Washington Post, Pokémon cards generated a roughly 3,821% return between 2004 and 2025—a massive margin compared to the S&P 500’s 483% gain in the same time period.

“Once that generation who collected and played as kids have gotten to a place in life where they have disposable income, you know they’re going to make a determination,” he said. “Perhaps modern art or bobbleheads or watches isn’t their thing, but trading cards are.”

That personal nostalgia and connection is the crux that keeps the market together. It’s an ongoing cycle where “generation after generation has embraced the form factor”—and that’s what sets trading cards apart from previous alternative assets like NFTs, Levin said.

“There has been a trend for people—while expanding at the same time with all these bleeding edge technologies—to also kind of circle back to real life experiences. More tangible things, and trading cards are very much that. There’s a stickiness to it. There’s a community to it,” he said.

For Levin, this isn’t just a cash cow financial investment that he wants to liquidate as soon as the market gets big enough. If he was in it for the money, he said he would have already sold it by now. In fact, he doesn’t even have an estimate on the worth of his 500,000 trading card collection, though he did mention, “None of my good stuff is kept at home.”

“The truth is, I like to collect all cards. I love to get my hands on cards,” he said. “But I have a deep appreciation when people take the time and energy to create something that is unique.”

“There are certain cards where I just love the way they look, and I love the quality of the paper,” he added.

A newfound profession

And that hobby-turned-financial-decision has seeped into Levin’s personal and professional life. He just attended his 31st Comic-Con this past year, and the man admits to being called “batshit crazy”: he sent Fortune a what he said was a never-before-seen photo of his wall covered in “tens of thousands” of pins ranging from Star Wars and Power Ranger characters to Hello Kitty and Nintendo characters.

Among his reserve includes a collection of 25,000-plus comic books and half a million trading cards—including about 100,000 Pokemon cards with the rest being made up of baseball, basketball and collab cards like a Dodgers crossover of a “One Piece trading card.” The Dodgers fan also said he has “every Ohtani bobblehead that’s ever been made.” He even managed to bring that obsession into his profession. 

“We have a cohort within Griffin that competes every other week,” he noted. “You know, we have Magic the Gathering get-togethers.”

For Levin, his aspirations of a Pokémon card-led currency has its merits. He told the Hollywood Reporter he believes Pokémon cards could become a global currency in the aftermath of an apocalypse, based on their worldwide recognition, accessibility and increasingly valuable secondary market.

He later told Fortune he was joking about the post-apocalyptic part—but still believes Pokémon is globally unifying given how ubiquitous and universal it has become, and part of it is that it’s essentially a level playing field: Anyone can build a card collection, and everyone has an equal shot of getting a “rare” card. Wealthy investors are not the only ones limited to getting the valuable cards.

“As a global currency, everybody knows Pokémon, it’s big everywhere. It’s accepted by all cultures and societies, and it’s celebrated and it’s cross generational,” Levin said.

“If you can combine your passion with an investment strategy,” Levin added, “or at least a sub-vertical within your investment strategy, why not?”

Gotta catch ‘em all

The Pokémon card itself dates back to 1996, the same year the original Pokémon games debuted in Japan. The cards initially functioned primarily as a game and collectible—but over time, rare cards became valuable commodities in their own right. Professional grading companies evaluate cards based on condition while online marketplaces and auction houses have created a global market where buyers and sellers can establish prices.

The market’s biggest acceleration came during COVID, when people who were stuck at home returned to the hobby while investors and influencers began treating the cards like assets. Online ecommerce company eBay reported that domestic trading-card sales jumped 142% in 2020, while Pokémon card sales specifically surged by more than 574%.

And these rare cards are reaching prices that would have been almost unimaginable during Pokémon’s original 1990s boom. A first-edition, shadowless Charizard card sold for $369,000 in 2020, and it was only up from there. Online content creator Logan Paul purchased a Pikachu Illustrator card for $5.275 million in 2021 and later sold it in February for approximately $16.5 million—setting a new record for a trading card in the process.

The market today has become large enough to attract major attention. Entertainment company Disney partnered with trading card maker Topps to make collectible cards involving Disney properties, while Hasbro and Ravensburger have also expanded into the card business. The trading-card market is estimated at as much as $50 billion annually.

This story was originally featured on Fortune.com