Abstract image of a whistleblower amid a see of faces.

Congress Can Get AI Whistleblower Protection Right On Its First Try

Nine small images with schematic representations of differently shaped neural networks, a human hand making a different gesture is placed behind each network.

AI Audits Need a Power Test, Not Just a Fairness Score

The Mathare River is one of the rivers that flows into Nairobi River. Photo by Queen Asali/Wiki For Sustainable Futures 2026/CC by 4.0

The Mathare River is one of the rivers that flows into Nairobi River. Photo by 

Taxing Tech in Age of AI Could Fund Human Rights

Composite. Left, Supreme Court Justice Samuel Alito attends a meeting in Rome, Sept. 20, 2025. (AP Photo/Andrew Medichini) Right, Meta CEO Mark Zuckerberg, CEO of Meta, is seen in the US Capitol on Thursday, March 26, 2026. (Tom Williams/CQ Roll Call via AP Images)

How the Supreme Court and Big Tech Both Weaponize 'Race-Neutrality'

Children’s Digital Rights Need More Than a False Promise of Safety

In an aerial view, the Meta El Paso Data Center is seen on August 13, 2026 in El Paso, Texas. Meta continues construction on its more than $10 billion data center, occupying approximately 1,000 acres, and has expanded the project to 1 gigawatt of computing capacity, allowing the facility to support a growing range of AI workloads and the training and deployment of AI models and technologies. (Photo by Brandon Bell/Getty Images)

Zuckerberg’s AI Manifesto Lost it at Compute

Flock Safety camera on Haverhill Drive near Phillips Avenue in Toledo, Ohio. (Photo by Stephen Zenner / SOPA Images/Sipa USA)(Sipa via AP Images)

Unless Flock is Reined In, Mass Video Surveillance Will Grow Even More Pervasive

Protesters rally against data centers in Hillsboro, Ore., on May 16, 2026.  (Photo by Alex Milan Tracy/Sipa via AP Images)

How to Use Your Community’s Data Center Moratorium

Chinese President Xi Jinping, center, waves as he arrives at the opening ceremony for the World AI Conference in Shanghai, Friday, July 17, 2026. (AP Photo/Ng Han Guan, Pool)

Will China Crack Down on Open-Weight Models?

Brazilian presidential candidate Ronaldo Caiado, former governor of Goias state, presents his public security policy in Sao Paulo, Brazil August 10, 2026. REUTERS/Jean Carniel

How a Brazilian Governor Turned Rare Earths Into a Campaign—and Washington Into a Partner

CEO of Meta Platforms Mark Zuckerberg attends the Allen & Company Sun Valley Conference at the Sun Valley Lodge on July 9, 2026 in Sun Valley, Idaho. (Photo by Kevin Dietsch/Getty Images)

Four States Take Meta to Trial in California, the Latest Front in Sprawling Multidistrict Fight

Vietnam Tightens Online Speech Controls as Cyber Law Takes Effect

Cemetery for people on the move in Soufli, Evros, Greece. Photo by Petra Molnar.

The Racial Necrotechnics That Lead to Death By Design at Automated Borders

AI Meets the US Midterm Elections

Michigan residents turn out to vote in the midterm elections at Brookside Christian Reformed Church in Grand Rapids, Mich., on Tuesday, Aug. 4, 2026. (Isaac Ritchey/The Grand Rapids Press via AP)

How AI Is Reshaping Election Information Ahead of the 2026 US Midterms

A brightly coloured illustration which can be viewed in any direction. It has several scenes within it: people in front of computers seeming stressed, a number of faces overlaid over each other, squashed emojis and other motifs.

Here Come the Synthetic Voters

A person passes next to an Aadhaar Seva Kendra in Kolkata, India on August 15, 2023. (Photo by Sudipta Das/NurPhoto via AP)

Aadhaar's Success Story Deserves a Closer Look

The Discord logo is displayed on a mobile phone screen in this photo illustration in Brussels, Belgium, on February 28, 2026. (Photo by Jonathan Raa/NurPhoto via AP)

Brazil Suspends Discord's Livestreams in First Enforcement Test of Child Protection Law

Mark Zuckerberg, CEO of Meta, is seen in the US Capitol after a meeting in the office of Senate Majority Leader John Thune, R-S.D., on Thursday, March 26, 2026. (Tom Williams/CQ Roll Call via AP Images)

When an AI Manifesto Runs Into Reality

Richmond County Courthouse sign in Augusta, Georgia. Shutterstock

Richmond County Courthouse sign in Augusta, Georgia. 

Why Public Defense Should Incorporate AI Carefully

In 2019, researchers examined a commercial algorithm used by US health systems to identify patients who should receive additional care. The system predicted future health care costs rather than medical need. Because less money was historically spent on Black patients with comparable illnesses, Black patients assigned the same risk score were considerably sicker than white patients.

Replacing cost with a closer measure of need would have increased the proportion of Black patients selected for additional support from 17.7 percent to 46.5 percent, according to the 

The model was not simply inaccurate. It was competently predicting the wrong thing.

That distinction matters as audits become a central instrument of AI regulation. 

 for certain automated employment decision tools. The 

 provides for conformity assessments and fundamental-rights impact assessments for high-risk systems. The 

US National Institute of Standards and Technology’s (NIST) framework

 encourages organizations to govern, map, measure, and manage AI risks.

These mechanisms are necessary. But an audit that begins after an institution has chosen the system’s objective may certify an unequal policy with impressive technical documentation. AI accountability therefore needs a 

A fair result can still serve the wrong objective

Most technical audits ask whether a system performs consistently, whether error rates differ across groups, and whether documented controls are working. These are important questions. They are not the same as asking whether the system’s target is legitimate.

The health care case shows the difference. Predicting expenditure may appear administratively sensible because cost data are available and measurable. Yet expenditure reflected unequal access to care. Once cost was treated as a proxy for need, a social inequality became part of the model’s objective.

 has long warned against isolating a model from the institution in which it operates. A statistically balanced system can still support a punitive welfare policy, an exclusionary hiring process, or a surveillance practice that should not have been automated in the first place.

 can occur. The term does not imply that auditors intentionally conceal discrimination. It describes a narrower institutional risk: political choices about what should be predicted are translated into technical targets, assessed through compliance metrics, and returned to the public with the authority of an audit.

Existing frameworks contain the right pieces, but not a complete test

Some governance frameworks already move beyond model performance. The 

 asks organizations to document intended purposes, consult relevant external actors, and make a go-or-no-go decision about whether an AI system is appropriate.

 similarly asks whether AI adoption is justified, identifies relevant stakeholders, and examines positive and negative consequences.

The EU AI Act’s fundamental-rights impact assessment requires covered deployers to describe the context of use, identify affected groups, assess risks, and establish mitigation and redress arrangements. However, participation by affected groups is encouraged “

” rather than made a uniform requirement.

Under the current political agreement on the EU’s implementation timetable, the relevant high-risk requirements have also been postponed while standards and guidance are completed. This creates additional time to improve the substance of the assessment process rather than treating the delay only as a concession to compliance concerns.

The building blocks therefore exist, but they remain fragmented. Some are voluntary. Some apply only to particular organizations or categories of systems. Others require documentation without clearly establishing who has authority to challenge the purpose of the system or the contractual arrangements behind it.

The emerging audit regime needs a shared power test.

Four questions every consequential AI audit should answer

An audit should identify who selected the objective, which alternatives were considered, why the AI system was chosen as the solution, and what would happen without automation. It should name the proxies used and explain why they represent the underlying social goal. This requirement would have made the difference between health care cost and medical need impossible to treat as a minor modeling choice.

Audit documentation should disclose who owns or controls the relevant data, model, computing infrastructure, and intellectual property. For public-sector systems, procurement contracts should guarantee access to documentation, change logs, incident reports, and independent testing.

Contracts should also specify whether an institution can switch providers or decommission a system without losing access to essential data or services. Vendor ownership does not make a system inherently harmful, but it determines where and how accountability can be exercised.

3. Who receives the benefits and bears the errors?

Group-level error rates remain important, but audits should also examine how benefits, burdens, and savings are distributed. A system may reduce administrative costs while transferring investigation, delay, documentation, or appeal costs to applicants, workers, patients, or welfare recipients. These effects should be treated as part of system performance rather than as externalities.

Notice and explanation are insufficient without practical routes to correction. Audits should verify that affected people can obtain relevant information, challenge data and classifications, reach a responsible human decision-maker, and receive timely remedies.

Affected communities should also have standing to influence system objectives before deployment, not merely report harms afterward.

Make the power test part of existing governance

A power test does not require a new regulatory agency or an additional audit industry. It can be incorporated into existing impact assessments, procurement reviews, conformity procedures, and public transparency records.

Algorithmic Transparency Recording Standard

 already asks public bodies to publish information about system ownership, rationale, deployment context, data, risks, and accountability.

The EU’s fundamental-rights assessment template could make affected-group participation mandatory for consequential public uses and add questions about vendor dependence, benefit distribution, and the no-AI alternative. Regulators could require a confidential annex where genuine security or trade-secret concerns prevent full publication, while still publishing a meaningful public summary.

The requirements should be proportionate. A power test is most justified for systems that materially shape employment, education, health, welfare, credit, migration, policing, or access to public services. Standardized templates could reduce compliance costs, particularly for smaller organizations.

Auditor independence also matters. Providers should not be able to define the benchmark, select the evidence, and determine what counts as a successful result without external scrutiny. Previous 

 has identified unresolved questions concerning audit objects, credentialing, and declining audit quality.

Regulators can establish minimum methodological standards, approve qualified auditors, require disclosure of financial relationships, and trigger reassessment after material system changes or serious incidents.

Auditing power is not the same as politicizing technical review

A predictable objection is that auditors should assess systems, not decide public policy. That is correct. Auditors should not replace legislators, regulators, courts, workers, or affected communities.

But the objective of an AI system is already a policy choice. Treating it as a fixed technical specification does not make the choice neutral. It merely shields it from scrutiny.

The purpose of the power test is to make these decisions visible and assign responsibility for them.

Technical auditing remains indispensable. Regulators need evidence about accuracy, robustness, discrimination, security, and compliance. Yet these measures answer whether a system works according to its specification. Democratic accountability also requires asking who wrote the specification, whose interests it serves, and who can refuse its consequences.

An AI system can be accurate, documented, and compliant while still allocating care, work, welfare, or scrutiny on an unjust premise. As audit regimes harden into governance infrastructure, regulators should ensure that the audit object does not stop at the model. The institution, objective, contractual dependencies, and distribution of power must also be examined.

A fairness score asks whether a system treats groups consistently. A power test asks who has the authority to define consistency, and who has the power to challenge it.

Without that second test, compliance can certify inequality. With it, an audit can become a democratic checkpoint rather than a technical seal of approval.

, faced a choice no one should have to make, between nearly $2 million in vested equity or his right to warn the public about the technology he was helping build. He chose the warning and gambled the money because he would not sign a lifelong non-disparagement agreement. The company backed down in the face of media coverage; the law offered 

. Kokotajlo was not an isolated case; he is one of 13 current and former employees of OpenAI and Google DeepMind who, in June 2024, 

 titled “A Right to Warn about Advanced Artificial Intelligence.” They complained that broad confidentiality agreements “block us from voicing our concerns,” and ordinary whistleblower laws, which protect those who report violations of the law, didn’t help because the risks they feared weren’t illegal yet.

The pattern continues. In January, a senior OpenAI’s 

 within weeks of objecting to a risky product launch and raising child-safety concerns. The company said her departure had nothing to do with what she’d raised and pinned it on “sexual discrimination against a male employee,”which she flatly denies. The trouble is that almost no one can prove her firing was retaliation, and that is its own message to everyone still inside the company. Employees like these see AI’s risks first, and they are telling us that they cannot speak freely.

Good whistleblower law succeeds when it is tailored to the specific ways an industry silences insiders. That has long been the pattern of whistleblower laws in the financial sector, but it has always been a step behind. The financial sector’s legislation was written as a crisis response, introduced after a catastrophe had already happened, and no one had spoken up. For instance, Enron and the “

” gave us Sarbanes-Oxley. The 2008 financial crisis gave us Dodd-Frank and, through subsequent SEC rulemaking, a framework that finally 

targeted the tactics companies used to buy silence

. Neither reform was perfectly responsive from the outset; they evolved over time. 

The AI industry is different. Lawmakers already have ample evidence of how frontier AI companies can suppress reporting. They must not wait for a defining catastrophe before designing protections that reflect those realities. The test for the current AI bills is whether they answer how this industry, specifically, keeps its insiders quiet. Judged against that standard, the current proposals are an encouraging and important first step.

, introduced in May 2025, represents Congress's first serious attempt to establish AI-specific whistleblower protections. In June, Sens. Grassley and Chris Coons (D-Del.) introduced 

nearly identical protections as an amendment

 to the must-pass National Defense Authorization Act (NDAA), giving the proposal its best chance of reaching a vote. More recently, a sweeping bipartisan draft, the 

, also carries its own whistleblower protections. That means now is the time to get the details right, because in whistleblower law, the details are often the difference between a statute that works and one that merely looks reassuring.

The strongest proposal remains Grassley's AI Whistleblower Protection Act (AWPA). It would provide the first federal whistleblower protections written specifically for AI employees, and the most robust drafted to date. Its 

 is that it protects employees who disclose a “substantial and specific” danger to public health, public safety, or national security, not merely those who report a violation of law. That distinction is critical because many of the most serious frontier AI risks are not yet illegal.

For example, recently an OpenAI model undergoing cybersecurity evaluations broke out of its test environment, exploiting a previously unknown vulnerability and reaching Hugging Face’s systems. Suppose an employee had identified the risk beforehand and believed that the system posed a “substantial and specific” danger to the public and disclosed this to a government body. Under the current whistleblowing framework, it is unclear whether they would be protected. There is no single federal whistleblower law. The federal Whistleblower Protection Act uses the “substantial and specific danger” standard, but covers only government employees; protections for private-sector workers are generally tied to specific forms of wrongdoing that AI-safety warnings generally fall outside of. State law is similarly patchy and most protect only reports of legal violations, except New York. An employee who reasonably believes that a model poses a grave national security or public safety risk should not have to identify a statutory violation before safely alerting the government.

Importantly, “substantial and specific” danger is not a novel or unusually expansive standard. Federal whistleblower law has protected disclosures of 

 that threshold as requiring catastrophe before an employee may speak up. Indeed, the standard compares favorably with existing AI legislation. California's 

, for example, protects safety disclosures only where the risk reaches a "catastrophic" threshold of more than 50 deaths or injuries or $1 billion in damage. The federal approach is considerably broader and arguably better suited to the uncertain nature of frontier AI risks. The NDAA version strengthens it further by extending protection to contractors as well as employees and by providing meaningful remedies against retaliation.

Grassley’s AI Whistleblower Protection Act, the NDAA, and to an extent, the Great American AI Act represent substantial steps forward and contain many strengths. Measured against whistleblower best practices, they get most of things right, such as holding workers to a reasonable-belief standard, protecting reporting both internally and externally and barring the pre-dispute arbitration clauses companies use to keep disputes out of view. Most importantly, they provide the people building frontier AI with protections written specifically for them, rather than having to rely on the existing patchwork. However, they do not yet capture the full range of pressures that discourage AI insiders from speaking up. Three significant gaps remain, and each can be closed using tools that have already been employed in other areas of federal whistleblower law.

First, the bills help only after someone has already spoken.

 The current AI whistleblower bills make a company’s NDAs unenforceable against a whistleblower and keep their retaliation claim out of forced arbitration. This is a very useful protection, but only once a whistleblower has come forward and is defending against NDA enforcement. They do nothing for the researcher still deciding whether to make the call, weighing whether the company will sue them into the ground even if they’d eventually win.

NDAs have a chilling effect on whistleblowers regardless of whether they can be enforced. A 

 in March found that sweeping NDAs chill reporting even when they’re unenforceable, because employees don’t know they’re unenforceable. That's why, under 

, which makes the mere imposition of a silencing agreement its own violation, enforceable by the SEC whether or not anyone is ever actually silenced.

The SEC brought a record number of cases under Rule 21F-17 in fiscal year 2024, including one resulting in an 

 penalty. But companies keep writing the unenforceable NDAs because they keep working. Good whistleblower law exists not only to protect whistleblowers but to make whistleblowing possible. If the law waits until someone has already taken the risk, it arrives too late. Congress has already recognized that principle in financial regulation, and AI legislation should do no less.

The second gap is that current legislation only reaches paid staff and some of the people best placed to catch AI’s risks often aren’t on the payroll at all.

 Much of the serious testing is done in-house by employees, but a large proportion is done by outsiders, and the labs rely on it. Anthropic 

 companies to facilitate independent third-party red teaming, acknowledging that outside researchers play an essential role in identifying AI risks. These researchers depend on company-controlled access that can be revoked through terms of service or account suspensions. Invited red-teamers can have their access revoked through terms of service or account suspensions and independent researchers who probe these systems on their own 

 for doing it, with no company offering them real protection. While breaking terms of service can be fair grounds for being cut off, the concern is that the very rules that let a lab grant access can be used to revoke it when a researcher finds something inconvenient for the lab. Yet the bills reach only the people a company pays. The red-teamers and unpaid academics are left out, a regressive move in comparison to the 

, which has rewarded outsiders for years. The public’s safety shouldn’t depend on whether a lab happens to be cutting the tester a check.

The third gap concerns a compensation structure that gives frontier AI companies an unusual source of leverage over employees.

equity constitutes a substantial share of compensation

, but because that equity is typically illiquid, employees often depend on 

 offers to realize its value. That hands a company enormous leverage (like forcing an employee to sign broad and restrictive NDAs), which is exactly why the threat that hit Kokotajlo wasn’t merely being fired but also losing millions in vested equity.

Whether the bills cover this is unclear just from the text. A court might read their language broadly enough to reach an equity clawback 

 and treat valuation of equity as speculative, especially where the company is private. In an industry where equity is both the paycheck and the leash, leaving the central question to a years-long court fight is a strange thing to do on purpose. One sentence making equity clawbacks an express form of unlawful retaliation would remove that uncertainty altogether.

These bills represent a significant step forward, but the people inside these companies who notice trouble first shouldn’t have to choose between their conscience and their savings or gamble that a judge will side with them years later. Congress has already done the difficult work of identifying the right framework. The remaining changes are modest, well-tested, and already found elsewhere in federal law. Lawmakers should not waste the opportunity to get AI whistleblower protections right on their first try.

Sarah Saadoun is a senior economic inequality adviser at Human Rights Watch.

 about the economic fallout of artificial intelligence have brought tax policy into the spotlight. Governments could 

find themselves desperate for more revenue

 at exactly the same time that they are losing the taxes those vanished jobs would have generated.

There is no magic bullet governments can use to resolve this problem. One approach with 

 AI profits. The idea is that governments could use that revenue to help manage the human rights impacts of rapid, AI-driven change. This approach has support in some surprising quarters. 

 have both floated proposals that would increase taxes on companies benefiting from AI as part of their case for widespread adoption of AI.

The current international tax system doesn’t allow states much room to do this effectively, however. A century-old rule common in bilateral tax treaties requiring physical presence 

 of the world’s governments, which was a growing problem even before the advent of AI. A change requires a much larger, collective solution to the larger failings of our global tax system. Luckily, it is already taking shape.

 that could finally make it easier to tax tech companies, among other badly needed global tax reforms. Not only could this treaty better equip governments to weather AI-driven uncertainty, but it could give them essential tools to better protect and realize human rights across the board.

Rules for cross-border taxation set out in the 1920s and fine-tuned by the Organization of Economic Cooperation and Development (OECD) form the basis of thousands of bilateral tax agreements between countries. These treaties generally let governments tax a foreign company’s profits only if it has a “permanent establishment”—a physical presence—in the country.

This means tech giants like Meta or Uber, which often operate entirely remotely, can earn fortunes in a country while paying tax on profits somewhere else entirely. For instance, Facebook has some 

 in Bangladesh, more than one-third of the population, yet has 

 of the “Silicon Six” found that these tech giants earned nearly half their $11 trillion revenues overseas between 2015 and 2024, yet only 30 percent of what they reported in estimated taxes was booked to foreign governments—disproportionately tax havens. For Global South governments especially, this increasingly adds to larger problems with an international tax system that hampers their ability to adequately fund human rights imperatives, such as education, health, and an effective justice system.

 of the global GDP and is growing fast. If even modest 

 of AI’s transformative impact pan out, the cost of the old tax rules to government revenues—and to rights—will significantly grow.

Attempts to get around these rules, including by opting out of bilateral tax treaties that impose the constraints, typically yield relatively 

 because, on their own, governments struggle with 

. Unilateral measures also make governments vulnerable to 

 by other governments, particularly from the 

This is one reason why many governments have banded together to push for a first-ever UN tax treaty. The US has walked out of negotiations, but dozens of other governments are hammering out an agreement that could be adopted in 2027. One of its key issues is how digital services are taxed. African countries, for example, have 

 allowing governments to also tax companies where payments and market engagement take place. This could include, for instance, social media advertisements, digital subscriptions, and digital labor platforms.

The treaty also aims to tackle other obstacles to fairly taxing companies, like profit-shifting to tax havens and tax systems that struggle to cooperate across borders.

Kenya’s uphill battle to tax tech illustrates why this matters. Although the country’s economy has

 over the past decade, between 2019 and 2022 the number of Kenyans living below the national poverty line increased by approximately 4 million. At the same time, government spending on public 

 fell well below best-practice global benchmarks. In fact, education spending as a share of GDP 

In Mathare, one of the largest informal housing settlements in Nairobi, residents have fought for years to get even a single public elementary school built in an area with 100,000 people. Jack Owuor, one of these residents, explained to me that he had no choice but to pay for private schools for his children, like a 

. Another resident said she has been denied medicine for her asthma because she couldn’t pay.

 raising revenues, although other issues are also to blame. One way Kenya has tried to increase its revenues is by taxing its digital economy, a fast-growing sector that is 

 to contribute $5 billion to GDP by 2028. This doesn’t distinguish between national and foreign firms, but Kenya struggles to effectively tax those based overseas.

Kenya does not have a tax treaty with the US, where most tech giants are headquartered. In 2021, the government put in place a 1.5 percent 

 on non-resident providers like streaming platforms, ride-hailing apps, and online marketplaces. Annual revenues from the tax have gradually increased, but in 2024 it still only brought in

. In December 2024, it replaced this tax with an effective 3 percent levy on the gross digital revenues of foreign providers. Figures from Kenya’s revenue authority indicated that these taxes generated a combined $12 million in the fiscal year that ended in June.

These are positive steps given the limitations, but a drop in the bucket relative to the scale of the need—or the sector’s size. Kenya also faces 

 similar taxes. A country acting alone can only do so much. It would take a global fix to enable governments to tax global tech firms in a way that reflects the scale of their domestic economic activity and profits.

The difficulty of taxing tech and other company profits means higher taxes on regular people, especially since an IMF program that ran between 2021 and 2025 required the government to hit revenue targets. In 2023, Alfredo Akeyo, who also lives in Mathare, 

 that price increases for public transportation following a new fuel tax meant that he was sometimes unable to send his children to school. The following year, a 

 that would have raised taxes on things like cooking oil and sanitary pads sparked nationwide protests.

Kenya’s situation is not unique. Even in a decade marred by a pandemic, wars, and inflation, nearly all governments saw their 

, many significantly. But these gains often have not translated into better fulfilment of people’s rights in part because an outdated 

 makes it difficult for governments, particularly in the Global South, to raise enough revenue. According to the International Monetary Fund, 71 governments have tax-to-GDP ratios 

, which it considers the “tipping point” for funding essential services.

People have paid the price through their governments’ inability to fund key services their human rights depend on. Between 2019 and 2022, 37 governments 

 their spending on health care per capita in real terms; four of five people worldwide now live in countries where public health spending is below the widely accepted benchmark of 5 percent of GDP. According to the 

, education spending per child has either decreased or stagnated globally. And about 

 the world’s population lacks any form of social security benefit, contributing to an estimated 

—equivalent to the population of the US—experiencing food insecurity than before the pandemic.

The UN treaty could help turn the tide. It is a landmark opportunity for governments to come together to decide on a better global tax system, one that also adequately deals with multinational tech companies.

But if policy debates about taxing AI continue to narrowly focus on the US and other rich countries, they will leave out governments already struggling with tax rules that will leave them at a profound disadvantage in this new policy arena. It is worth heeding their hard-learned lessons, and these states deserve a solution that includes them.

In an aerial view, the 1,000 acre Meta El Paso Data Center as of August 13, 2026 in El Paso, Texas. (Photo by Brandon Bell/Getty Images)

 is this: “superintelligence-for-all” empowerment will never happen if Meta and other hyperscalers monopolize compute. Control over compute is a tool of domination as powerful as any of the frontier model secret weights that Zuckerberg slags.

 observes wryly that in the manifesto, “a man who single-handedly commands a $1.5 trillion tech giant comes out as a critic of corporate power.” Platformer’s 

 notes that while Zuckerberg “celebrates individual voices and democracy,” his company “is working to warp the democratic process through secretive data center deals.” Through those deals, Meta is trying to catch up with the other four hyperscalers (Amazon, Google, Microsoft, and Oracle), who together 

 of the world’s AI compute. Zuckerberg doesn’t talk about 

 source of power. Instead, he focuses further up the stack on models and applications. Here, he says, “[r]ather than centralizing superintelligence, we should distribute it widely and give every person the ability to direct it.”

 the physical infrastructure needed to produce that superintelligence? Infrastructure power 

, in its 2024 report on open weight models (p.30) noted that access to compute is a significant chokepoint for innovating with such models. As 

, “the computational resources needed to build new AI models and use existing ones at scale… are scarce, extremely expensive, and concentrated [and] this significant resource asymmetry undermines any claims to democratization that the availability of ‘open’ AI models might be used to support.” (p.8)

In other words, frothing about AI democracy and liberty is useless without addressing compute. Zuckerberg knows this. That’s why he acknowledges compute as a pinch point in the flow of freedoms he champions. “No matter how intelligent AI becomes, there will always be a finite amount of compute and therefore an opportunity cost for how we use it,” he writes. His response to this scarcity is for the tech giants to build ever more compute so that there’s enough for training “recursive self-improvement” while also “committing the significant majority towards people's individual goals.” Here, as elsewhere in the manifesto, governance choices are the tech companies’ to make. 

 will own the means through which people self-actualize. Every “exceptionally capable personal agent” and every “powerful tool to build businesses” and all the other compute-driven wonders the manifesto teases will depend on the few companies controlling compute.

For the conveniently libertarian political theory of Zuckerberg’s manifesto, this concentrated private power is unimportant. The only real threat to AI-enabled individual liberty is government tyranny. “There is a risk that an imbalance of power between individuals and government leads to a loss of freedom or totalitarian state,” Zuckerberg writes. “This is a constant tension in democracy.” What’s also a constant tension in democracy is how to constrain overweening private power. Technology policy in the American democracy has never been libertarian. Zuckerberg’s only recognition of private compute power is his assurance that Meta (and other benign titans) will distribute compute widely and focus heavily on alignment. In other words, “trust us.” Given Meta’s history on the trust front, as it appeals a pathbreaking New Mexico court judgment that it is a public nuisance, this tack is laughable. And it has been loudly laughed at 

. Obviously, binding and transparent regulation is a more effective check. The manifesto will have none of it ,and Meta has been instrumental in making sure the federal government enacts none.

What might a compute policy look like that took self-empowerment and democracy seriously? A person writing a version of Zuckerberg’s screed with an eye to network studies might have cited Ithiel de Sola Pool’s classic 1984 work, 

. Pool wrote about how new technologies (like “electronic publishing”) might usher in new freedoms through “distributed intelligence” (p. 251) if only government could abstain from centralizing control. This fits well with Zuckerberg’s paean to decentralized open weight models and preference for policy that amounts to little more than facilitating Meta’s business model. But here’s the thing. Pool understood that freedom at the application layer might require regulation at the infrastructure layer in the form of common carriage and access requirements.

For compute, and more generally for cloud services that mediate access to compute, we can look to the same family of tools long used in US regulation of concentrated networks, platforms, and utilities.

Remove chokeholds on access to compute where necessary through some combination of nondiscrimination, interconnection, and common carrier obligations. Jai Vipra and Sarah Myers West of the 

 have a nice report on this from 2023 when the FTC and the Department of Treasury were beginning to sound the alarm on undue concentration and points of failure in the compute and cloud services markets. Last year, 

 of the Vanderbilt Policy Accelerator published a blueprint for federal legislation for the AI stack, including compute, that he described as pro-competition and pro-national security, and which might actually create the balance of power Zuckerberg’s manifesto praises.

Ensure that there is some meaningful supply of “public compute” that is not controlled by any company. The scarcity of compute for scientific research has led to nascent federal efforts to amalgamate commercial and government compute for research in the 

 with new compute resources, again mostly for research. One might imagine a federated network of state compute forming a sovereign compute infrastructure of the kind 

 is trying to build. A new idea would be to regularize public compute set-asides as part of data center builds. What if hyperscalers had to set aside a certain amount of  “

”, like a public park, for the autonomous use of nonprofits, local governments, and educational institutions? Cable companies had to set aside PEG channels when building out their networks and the FCC reserved noncommercial broadcast channels for public service.

Require transparency. One could talk about transparency at each of the layers in the stack, but let’s limit it here to compute. Here too, the libertarian Pool acknowledged that to have freedom at the top layers – the only ones Zuckerberg really addresses – we need to know how companies are managing the infrastructure choke points. In the case of compute, this might include the amount of compute capacity a firm actually controls, to whom that capacity is allocated (including to the owner itself), and pricing terms.

Latest

Podcast

Recent

Our content delivered to your inbox

Thank you!

Around the globe