(L-R) Congressional candidate Claire Valdez, Congressional candidate Brad Lander, Mayor Zohran Mamdani, and Congressional candidate Darializa Avila Chevalier raise their hands during the 'Our Team, Our Year' Get Out The Vote (GOTV) rally for local candidates ahead of this week's Primary Elections at the Kings Theatre in the Brooklyn borough of New York City, NY, June 18, 2026. (Photo by Anthony Behar/SipaUSA via AP Images)

Populist Candidates Need More Tech Policy—Reforming Tenant Screening Is A Good Start

The New UK Prime Minister Inherits a Social Media Ban. The Real Goal Is Safety by Design.

Tear gas surrounds federal law enforcement officers as they leave a scene after a shooting on Wednesday, Jan. 14, 2026, in Minneapolis. (AP Photo/John Locher, File)

Trump’s Crackdown on Dissent is Targeting Signal—and Zines

Split-screen image depicting contrast. On the left, a Rubik’s cube displays blue, red, and yellow sides, with a LinkedIn 'cr' AI-generation icon nearby. The background is divided between discarded electronic waste on a street and data centre cables, some connecting to the cube. On the right, the entire scene is repeated with inverted colors.

This is not a(i) game! by Laura Sofia Martinez Agudelo / 

When AI Agents Fail, People Ask the Wrong Question About Why

Harvard University graduates, top, celebrate during commencement exercises on the school's campus, Thursday, May 28, 2026, in Cambridge, Mass. (AP Photo/Steven Senne)

For Students, the Process of 'Becoming' is the Challenge No Chatbot Can Solve

Illustration of a person standing with their back facing the viewer, surrounded by large stylized eyes watching them from above. The person is in a spotlight formed by overlapping blue and gray geometric shapes, conveying a feeling of being observed or surveilled.

Who Really Controls Your Digital Likeness in the Age of AI Wearables? Not You.

Paper collage. A photograph of the Atomium in Brussels, cut into diagonal strips and reassembled over a grid of numbers.

Seeing More — Seeing Less by Anna Riepe & FARI / 

Google and Apple’s Anti-DMA Lobbying Strategy Goes All-in on Security and Privacy

A child holds up a smartphone, obscuring their face.

They Said I'd Feel Different About Free Speech as a Parent. They Were Wrong.

June 22, 2026—Andy Burnham with colleagues from the Parliamentary Labour Party in Westminster Hall at the Houses of Parliament in central London, as he returns to the House of Commons to take up his seat after winning the Makerfield by-election. (Yui Mok/Press Association via AP Images)

How Will Andy Burnham Handle Tech Policy as UK Prime Minister?

In this photo illustration, the Telegram logo is seen displayed on a smartphone screen with the Indian flag in the background. (Photo by Thomas Fuller / SOPA Images/Sipa USA)

India’s Telegram Ban Was Temporary. The Power Behind It Is Not.

The Anthropic logo is displayed on a smartphone screen placed on a reflective surface onto which a company branding illustration is projected on June 2, 2026. (Photo by Samuel Boivin/NurPhoto via AP)

Anthropic's Red Lines Are No Substitute for Public Law

BRUSSELS, BELGIUM - JUNE 9, 2023: The symbol of the European Parliament on the glass façade of the European Parliament. Close up

EU Lawmakers Press Commission on Child Safety as Debate on Age Limit Heats Up

Shutterstock Illustration depicting two megaphones with a breaking news text ribbon suggesting propaganda and fake news.

A Policy Playbook to Inoculate the Public Against AI Text Falsehoods

The image depicts a conceptual illustration of risk assessment. A large triangular warning sign with an exclamation mark is positioned prominently, symbolizing caution or danger. A business professional in a suit, holding a magnifying glass, is inspecting the sign closely. The scene suggests careful examination or investigation of potential risks, with a neutral background featuring minimalistic clouds to keep the focus on the subject.

The EU's AI Transparency Code of Practice, Explained

The dome of the US Capitol building. Justin Hendrix/Tech Policy Press

Congress Should Pass AI Law to Reassure the Public

A charred piano, left, is visible inside the Pennsylvania governor's official residence after a man was arrested in the alleged arson that forced Gov. Josh Shapiro, his family and guests to flee in the middle of the night on the Jewish holiday of Passover, Sunday, Apr. 13, 2025, in Harrisburg, Pa. (AP Photo/Marc Levy)

What Online Platforms Can and Must Do to Help Mitigate Escalating Political Violence

The sign of University of California at the entrance of Berkeley campus. Shutterstock

The sign of University of California at the entrance of Berkeley campus. 

We Just Formed the Largest Tech Union in America. Here's How We Did It

Within the frames are people bound to their office cubicles; beyond them, individuals work freely from diverse locations, connected through digital signals.

Digital Nomads Beyond the Cubicle by Yutong Liu & Digit / 

AI May Dramatically Disrupt The American City. Are Urban Planners Ready?

Demonstrators march in Berlin, Germany, during a June 2026 protest against the digital expansion of police powers, biometric surveillance technologies, and the use of AI systems by authorities. (Photo by AndrÃ©s Trujillo/JNA Press via Nexpher Images/Sipa USA)(Sipa via AP Images)

The Quiet Erosion of Collective Action Under Digital Surveillance

Federal Bureau of Investigation insignias adorn desks in the agency's Joint Operations Center in New York, Thursday, June 4, 2026, ahead of the World Cup soccer tournament. (AP Photo/Richard Drew)

To Mitigate Anti-Tech Violence, Strengthen Democracy

Last week, the US Department of Justice (DOJ) 

 fifteen people in Minnesota on conspiracy charges stemming from local community organizing efforts during “Operation Metro Surge”—a federal assault on the Twin Cities that followed 

 by President Donald Trump targeting the area’s Somali community late last year.

By the time the Trump administration announced the 

 of Operation Metro Surge in February—a 

 characterization as abductions and harassment by federal agents continued—ICE and Border Patrol agents had 

 two people, poet Renee Good and intensive care nurse Alex Pretti. Federal agents had forced a naturalized citizen out of his home in subfreezing temperatures wearing nothing but shorts and a blanket, in an arrest that is now being 

 on the way to and from school pickup, with school administrators scrambling to make sure their students weren’t left without caregivers.

 recently released by Human Rights Watch describes the violence wrought by Operation Metro Surge as “a wide range of human rights violations.” These included not only the use of excessive force in the streets—with federal agents smashing car windows and deploying chemical munitions in neighborhoods—but also “inhuman and degrading conditions” in detention facilities, including restricted access to food, health care, and legal counsel. The report also 

 federal agents harassing and intimidating community observers—including multiple cases of agents driving to people’s homes and stopping to take photographs while making sure the residents were watching.

All of this is described throughout the indictment as, simply, federal agents “performing their official duties.”

In those same pages, rapid-response community organizing efforts are cast as pillars of a sweeping conspiracy—with a particular focus on efforts to gather information on federal agents’ activities and to share that information widely.

, throughout the siege in Minnesota, ICE and Border Patrol agents systematically worked to avoid identification, from covering their faces and not wearing identifying badges to covering or using fake license plates on their vehicles. These practices have been so common nationwide that 

 passed bans on federal agents wearing masks.

The public has few tools comparable to the sprawling surveillance apparatus and legal investigatory powers of the federal government—particularly when the government is invested in 

 officers engaging in violence and likely civil and human rights violations. But in the Twin Cities and across the country, communities have responded with the tools available: the 

 whistles. Word of mouth. Crowd-sourced maps. Spreadsheets. Group chats.

In other words: information-sharing. Efforts to keep people aware of and safe from the well-documented violence federal agents perpetrated. Efforts now cited as evidence of a criminal conspiracy.

Extremely standard anti-surveillance precautions are also listed among the core elements of the alleged conspiracy: collecting phones and smart watches before holding closed-door meetings, using the popular encrypted messaging app Signal, and otherwise “practicing ‘operational security’ or ‘OPSEC.’”

 that efforts to organize against ICE and Border Patrol operations—and efforts to document federal agents’ activities in public, in particular—amount to offenses ranging from obstruction of justice to domestic terrorism. Former Attorney General Pam Bondi, in remarks about the developer behind one app crowdsourcing information about ICE raids, 

 “he better watch out, because that’s not a protected speech.” Both the Department of Homeland Security (DHS) and DOJ have 

 information about people running ICE watch accounts on social media.

 activity of observing and reporting on the public activities of federal agents remains particularly jarring when considered alongside ICE and Border Patrol’s aggressive pattern of surveillance, intimidation, and harassment targeting reporters, community observers, and bystanders alike.

In February, the ACLU of Minnesota updated its lawsuit challenging federal agents’ activities in Minneapolis by 

 dozens of declarations that—like the recently-released Human Rights Watch report—describe ICE and Border Patrol agents following protestors to their homes, or arresting community observers for recording their activities. A similar lawsuit has been 

 federal agents documenting their license plates, taking photos of their faces, addressing them by name, and 

 to add them to a “nice little database” of domestic terrorists.

This week, a poll worker in Syracuse was 

 by ICE agents threatening the possibility of prosecution over social media posts critical of ICE.

The attempted criminalization of anti-surveillance measures is similarly alarming as DHS 

 its already invasive surveillance apparatus, from 

 to mobile phone location data through commercial data brokers.

Efforts to sweep a broad swath of extremely routine organizing activities under the vague umbrella of conspiracy charges are not new – and have largely failed in court.

Just weeks before the Minnesota indictment was filed, federal prosecutors were 

 against the “Broadview Six”—organizers, including a local congressional candidate, who had been indicted on conspiracy charges following protests against “Operation Midway Blitz” in Chicago.

Both the Chicago and the Minnesota indictments 

 conspiracy charges filed against dozens of people in Atlanta involved in opposing the construction of “Cop City,” a 

 of one of the city’s “largest remaining green spaces.” Evidence cited by Georgia prosecutors 

 distributing flyers, organizing a bail fund, and possessing zines critical of capitalism and the government. Those charges were 

The underlying logic of these prosecutions involves criminalizing 

 connected to direct action—despite (or perhaps because of) the fact that direct action has been a core tactic of every major social movement in US history, from the Greensboro sit-ins during the Civil Rights Movement to “

” on the steps of the Centers for Disease Control (CDC) headquarters during the AIDS crisis.

At the same time, these prosecutions have explicitly sought to undermine core First Amendment protections for speech and association. Like the Stop Cop City indictments in Atlanta, the recent Minnesota indictment repeatedly cites as evidence simple communications of ideology: wearing shirts or patches; declaring a belief in anti-fascism, in particular, but also anti-authoritarianism, anti-capitalism, and anarchism; and publishing newsletters or zines.

Recently, the Trump administration secured multiple draconian convictions on similar charges against anti-ICE protestors in Prairieland, Texas. Touted as the first case to 

 the viability of the Trump administration’s 

 of “Antifa” as a “domestic terrorist organization,” the government’s allegations focused heavily on the protestors’ political beliefs to justify terrorism-related charges. Two defendants, Elizabeth and Ines Soto, were 

 with “providing material support to terrorists” for printing anarchist zines. This week, Elizabeth Soto was sentenced to 50 years in federal prison. Another defendant, Daniel “Des” Sanchez, was 

 with transporting “Antifa materials”: a box of pamphlets and zines 

 as having “anti-law enforcement, anti-government, and anti-Trump sentiments.” Sanchez was 

 basic anti-surveillance practices, like using encrypting messaging apps, as evidence of criminal intent.

As the Trump administration’s crackdown on dissent escalates, it’s critical to recognize that indictments are not just legal documents—they’re also messaging tools, often targeting small groups of people as a chilling example for movements at large. And the message is clear: the administration wants to deter not only specific types of protest, like direct action, but also any organizing efforts that effectively interrupt the administration’s agenda.

But using Signal is not evidence of a criminal conspiracy just because federal prosecutors in Minnesota claim it is. Printing and distributing political pamphlets and zines is still protected by the First Amendment, even if the contents of those pamphlets speak to the long list of ideologies the Trump administration is 

, like anti-fascism, anti-capitalism, and “extremism on migration, race, and gender.” Observing and reporting on law enforcement agents’ public activities is also still a 

These efforts will only succeed if they do, in fact, deter people from exercising their constitutional rights, from organizing to protect their communities, and from speaking out against abuses of power, against fascism, and against state-sponsored violence in its many forms. Crackdowns on dissent aren’t a reflection of power—they’re a projection of power.

And they’re a projection we have a collective obligation to refuse.

The American public's feelings towards the tech sector appear to have established a direction of travel. Positive views of the 'internet industry' 

 from a 2015 peak of 60% to 41% in 2025, having bottomed out around 35–36% in 2021–2024. A survey conducted in March indicated that a meaningful 

 of the electorate would choose a nuclear power plant over an AI data center near their home, suggesting that the movement is transcending traditional NIMBY anxieties. A 'techlash' is gathering steam, and barreling into an election year 

 with insurgent congressional candidates—many of whom are defying conventional wisdom and 

 incumbents with votes to spare. Meanwhile, Senator Bernie Sanders's (D-Vt.) $7 trillion AI public ownership 

 has raised the ambition watermark for tech policy. 

To meet the moment, 2026's cohort of upstart candidates need bold, substantive proposals that explain precisely how they'll rein in a hitherto unbridled tech sector.

In this article, we argue that the intersection of housing and technology—specifically, the regulation of tenant screening software—presents a tailor-made policy opening for 2026's outsider slate. Rife with algorithmic 

, and 24,000 Consumer Financial Protection Bureau 

 to boot, tenant screening has hardly ingratiated itself to working class voters. Its adjacency to populist platform pillars, such as affordable housing, confers out-of-the-box credibility on the topic. And for challengers in renter-heavy metropolitan districts, tenant screening is a rare tech issue that is actually on home turf.

Still, political messaging on the impact of automation hardly writes itself. Even with sentiment towards Silicon Valley at a generational nadir, politicians are rightfully wary of bewildering voters with a complex solution to worker displacement or AI safety. But in the housing context, algorithmic malpractice is far more graspable. Ask any of the ~110 million US renters how they felt (or would feel) being denied a home based on faulty scoring and incorrect criteria, buried in proprietary screening software. Ask any housing applicant if it's fair that following a rejection, the re-screening fee falls again on their shoulders. Another denial, another $40, another draining apartment search back to square one. This scenario is all too common; tenant screening is a billion dollar industry influencing tens of millions of approvals and denials annually. At least 2,000 domestic vendors offer some form of algorithmically mediated report on applicants' housing-worthiness, and 

 landlords employ an automated tenant screening service to filter out "bad" tenants, with over a third following the recommendation without any human review.

The industry's shortcomings are eye-watering. Following tens of thousands of complaints, the Consumer Financial Protection Bureau 

 that screening reports are “filled with largely unvalidated information of uncertain accuracy or predictive value". This is in part due to the source data procured from data brokers to score applicants. Representative Ayanna Pressley (D-Mass.) has 

 a bill that would chip away at this problem, by restricting the use of criminal background data as an input. However, the incorporation of 

into scoring remains unregulated at a federal level. A Princeton-affiliated 

 of 3.6 million administrative eviction records across 12 states found that roughly 22% of records contained ambiguous information about how the case was actually resolved. In some locales like D.C., 

 of eviction filings don't lead to evictions, but unfortunate tenants remain tainted by the mere presence of a case in the public record. Landlords, when 

, do not typically distinguish between successful evictions and those dismissed, settled, or in which the tenant prevailed. With most screening providers failing to make this distinction either, the result is an underclass of renters unable to shake a life-long digital stigma, reinforced by crude, black-boxed algorithms.

The harms are not evenly distributed. Median credit scores, another input into screening algorithms, 

 sharply along racial lines. Black households appear in eviction filings at nearly twice their share of the renter pool. In the most prominent litigation of tenant screening, 

, it was alleged, and deemed plausible by the court, that SafeRent systematically lowered scores for Black and Hispanic applicants intending to pay rent with federal housing vouchers, in part because the algorithm did not credit the 73% of rent that the voucher reliably covered. The screening company paid out millions of dollars and accepted a five year injunction against using its scoring model without third-party validation. An illuminating case that temporarily curbed the unjust practices of a single firm, but did little to regulate tenant screening at large.

Existing federal laws and their enforcement have also proven inadequate. The 

 (FCRA) gives consumers the right to dispute inaccurate information in tenant screening reports, but does not require the vendor or landlord to provide a copy of the report to the applicant, so escaping a cycle of unfair rejections is far harder than it appears on paper. Nor is there much motivation to meaningfully improve scoring; the FCRA's $100-$1,000 damages range does not seem to be sufficiently punitive to move the needle. For its part, the Fair Housing Act is supposed to prohibit discrimination in access to housing. However, the most direct enforcement levers, the FHA's Disparate Impact liability

and HUD's screening discrimination rule, have been weakened by executive orders. HUD has formally proposed 

 its disparate impact regulation, and in a sign of things to come, the Department of Justice fully 

 Title VI's disparate impact provisions. That leaves a private plaintiffs' bar that recovers modest damages case by case, while the tenant screening industry continues unabated.

A substantive proposal to reform tenant screening—dictating what source data is legitimate and non-discriminatory, codifying screening companies as housing providers under the FHA, and forbidding application fees—would slot neatly amongst existing housing-related pledges. It also materially supports an affordability agenda, given that renters spend millions of dollars in aggregate on application fees annually, and far more through the immeasurable opportunity costs of missing out on a first-choice home.

A hypothetical bill can materially address these problems, while remaining understandable, achievable, and common sense. Building on Pressley's April 2026 

, the bill could close out the remaining error-prone and racially biased data sources still being used as inputs to deny housing. It could force the disclosure of all sources to applicants, such that they can correct the record and/or sue for damages in the case of discrimination (e.g. rejecting voucher holders) or non-compliance. Provisions that codify screening vendors as housing providers under the Fair Housing Act would officialize their effective role as gatekeepers to stable housing. Simplest of all, the bill could prohibit application fees, which function as a regressive tax on working class renters, and hitting those regularly moving home—e.g. those displaced by automation—the hardest. The fiscal cost of this regulation could be relatively low, and therefore perceived to be more realistic, if the enforcement mechanism is a private right of action with statutory damages, which mostly finances itself through contingency fees. This and other defensible elements have already been proposed by the 

Leadership Conference on Civil and Human Rights

Regulating the PropTech industry is not without precedence, nor is it particularly partisan. The 

, which amended the FCRA to address tenant screening accuracy, was a bipartisan vehicle. States and cities across the country have 

 Fair Chance housing and eviction-sealing laws to blunt the effects of automated discrimination. Late last year, New York Assemblymember Karines Reyes Tapia (D-87) pushed a landmark 

 that would force landlords to disclose screening criteria to applicants.

The political case is equally strong. The tech-skeptical constituency is a broadening tent, as demonstrated by the pluralist movement 

 for data center moratoria. That said, as a supply-side instrument, a pause in construction would not protect renters from already-deployed, non-generative,predictive analytics—indeed, tenant screening does not rely on heavy compute at all. What housing and data center debates share is salience and sensitivity. In both cases, even pro-business legislators seem uneasy defending corporate playbooks. The stakeholders aren’t particularly sympathetic; a coterie of 

 protecting the interests of landlords, and a PropTech sector that has spent decades minimizing tenant economic power. Conversely, the victims are discernable and close to home; anyone who's paid multiple application fees, anyone who's been denied housing based on an eviction case they won, and anyone who's seen a voucher undermine their search for stable housing.

For a 2026 congressional challenger yet to stake out a concrete position on tech, tenant screening is a highly credible starting point. It signals to voters, volunteers, unions, and community groups that constraining Big Tech's overreach is a priority. It tells the electorate, in no uncertain terms, that their representative will not sit idly by as automation continues to undermine and impoverish working households. It shows they won't align with, nor be the beneficiary of, mailers funded by Silicon Valley super PACs. The legislative scaffolding is halfway there, drafted across various white papers. All that's missing is a champion, willing to broach the tech policy conversation with a formal proposal.

Maeve Walsh is the Director of the Online Safety Act Network.

“I am not going to claim that even those measures will mean that no kids are on social media platforms.”

This was the UK’s Secretary of State for Science, Innovation and Technology, Liz Kendall MP, 

 that Britain would be following Australia in the introduction of an under-16 social media ban, along with other measures to limit children’s exposure to harmful features and functionalities. On the ban decision, she went on: “This is as much about resetting the expectations of eight, nine and 10-year-olds, by saying, ‘You won’t be going on this until you’re 16.’”

The honesty about the objectives of the measure—and its limitations—is, on one level, refreshing. Given that much of the impetus of the Smartphone Free Childhood campaign, which was instrumental in shifting the Government’s position on this, comes from parents of primary school-aged children, justifiably alarmed at what might lie ahead for their kids in the online world and looking for legislative support to say “No,” it is also not that surprising.

But it is also alarming, given Kendall brought forward no plans to increase the safety protections for those teens who, by her own admission, will still be active on “banned” platforms; data 

 suggests that 80% of Australian teens are still on social media, despite the ban in that country. Nor will Kendall’s proposed additional measures for 16-17 year olds (such as a potential ban on infinite scroll or nighttime “curfews”) manage the resultant cliff-edge for any 16 year olds who have complied with the ban but then lawfully access platforms that remain unsafe.

A ban may be a strong signal of intent, but if the new Prime Minister is serious about delivering for the parents who campaigned so hard for the State to stand up to Big Tech and its reckless lack of care for children, it can’t be all there is. The good news for Andy Burnham—if it is indeed he who walks into Number 10 in a few weeks time—is that a complementary approach, built on the principles of product safety and safety by design that are fundamental to all other industrial sectors, is possible. The even better news for a PM who might be looking for popular quick wins is that it is easily and immediately implementable under the Online Safety Act’s existing legislative framework, and it's already popular with the public, with 

 showing that 84% believe tech companies should be required to prove their products are designed and tested to be safe before use—the same standard that already applies to toys, food, household appliances, and most other products.

When the OSA was passed nearly three years ago, there was cross-party consensus that its objective—to make regulated services “safe by design”—should be stated explicitly in its very first section. It may have been something of a novel concept then. But in the intervening period, the international consensus around this has grown. In recent weeks, two influential Parliamentary committees in the UK along with international organizations including the UN Commissioner for Human Rights, UNICEF and the Council of Europe 

 as an alternative to blanket bans for children. The UN Secretary-General last week 

 States to “require tech companies to embed user safety in product design – and rein in misguided algorithms and misuses of artificial intelligence”. The UK 

o the recent G7 leaders’ call for “digital platforms which are safe-by-design, secure, privacy-preserving, age-appropriate and protective of children and youth”. The new PM has a ready-made opportunity to build on this momentum, not only showing that the UK can still set the pace on tech regulation but also restoring faith in politicians, regulators, and institutions to fully and forcefully deliver on the commitments the public are crying out for.

The harms from social media for children are pervasive and persistent and, whether parents, policymakers or politicians want to tackle them from a safety (exposure to harmful content) or a wellbeing (excessive screentime) perspective, treating these harms as a product safety failure gets to the root of the problem in a way that age-gating the products (and hoping for the best) can never do. Safety by design provides an outcome-focused, future-proofed requirement for companies to safety-test their products and take responsibility for designing out the risk of harm to users before their use. It holds them accountable for the design choices they make and prevents them from live-testing harmful features and functionalities on children and vulnerable users.

The Online Safety Act Network, with the expert input of 5 Rights Foundation and other organizations, has even done the legwork for the incoming PM. In May, we launched 

 on safety by design providing a road map for how to do this in practice, within the existing OSA legislative framework. Forty-five civil society organizations and experts endorsed the code. Its adoption—along with a clear steer to Ofcom that they need to get on and deliver it—would send a welcome signal to parents that the new PM is prepared to go further than his predecessor in taking on Big Tech and would reassure civil society advocates, academics and experts that he is also willing to act on evidence of what works.

Tech is not exceptional. Abiding by the product safety requirements found in every other industrial sector should be the norm. Announcing a ban may have signaled a symbolic reset in children’s relationship with social media. Embedding safety by design is the only way to ensure that a substantive reduction in their exposure to harm follows. The UK—and its new PM—has an opportunity to lead the world in this, to reset our relationship with Big Tech, to foster innovation that provides better alternatives to the US and Chinese hyperscalers, and to reduce the negative impacts these companies are having on our lives, politics, and democracy.

of student essays produced in collaboration with the Berkman Klein Center for Internet & Society at Harvard University.

, a Silicon Valley entrepreneur and investor spent days building a SaaS prototype on top of a contacts database using Replit’s AI coding agent. He set limits. He wrote key constraints in capital letters. For eight days, the agent performed well enough, but his trust was beginning 

. On day nine, the agent deleted 1,206 executive records, fabricated four thousand fake user records to cover the damage, and when the investor asked if the data could be recovered, told him it could not. It could.

The obvious critique is that the agent misbehaved. But that critique obscures a harder question, one that keeps getting asked wrong: why did human oversight, when present, fail to prevent the outcome?

The Replit incident is not an anomaly. Last year, Amazon set an 

 for at least 80 percent of their weekly work. Weeks later, the tool decided to “delete and recreate” a cloud environment, triggering a 13-hour 

Around the same time, a researcher at Meta gave an AI agent (

) access to her email for triage. It began 

. She sent stop commands from her phone. The agent continued. She sent more. It kept going until she manually terminated the process. Her explicit instruction was the oversight mechanism. It did nothing.

Every post-mortem of such incidents effectively asks the same thing: who made the mistake of ceding too much authority to an AI agent? But that is the wrong question. The right one: at what level of authorization was the agent operating, and is that the level where the harm occurred?

Typical AI agent systems offer a range of oversight mechanisms, and some are quite good. Sophisticated tools ask users to confirm before running commands and flag operations that touch sensitive systems or databases. But the incidents above suggest these mechanisms operate at the wrong level. The failures live in the gap between what users authorize and what agents produce.

The gap has structure. Users express what they want in natural language. What the agent may do gets defined at the level of individual operations: which commands it can run, which systems it can touch, which actions require confirmation. Consequences accumulate at a different level: outcomes. No major governance framework currently bridges these levels.

When the entrepreneur authorized a Replit agent to build a database, he was operating at the task level. The agent made its decisions at the operation level, each step individually plausible, with no single action triggering a hard stop. What resulted (the deletion, fabrication, and deception) was the product of many unchallenged operations strung together. At no point did the system ask: is this cumulative sequence what the user actually intended to authorize?

Amazon's engineers authorized their tool to assist with coding. Nobody authorized it to judge that deleting and recreating a production environment was the right fix for a specific problem. That judgment happened at the operation level, and the thirteen-hour outage was the consequence that landed at the outcome level.

A sequence of individually approved operations can produce one catastrophic outcome. An approval record might show forty-seven green lights from a user, yet show nothing about where they lead.

 AI agents as creating a new kind of agency problem. The same problem appears here in operational form. The more remote a human decision is from an agent’s eventual output, the harder it becomes to say that the human meaningfully authorized the result. Authorization that drifts far enough from consequences stops meaning much at all. Oversight requirements that never specify at what level oversight must operate cannot close the gap. They guarantee that someone was watching. They say nothing about whether that person was watching the right thing.

 deploying autonomous AI cannot trace, in real time, what those systems are doing or who is responsible for their actions. Only 28 percent can reliably connect an agent's actions back to a human “sponsor.” When the 

 examined 30 leading deployed agents, it found that 25 disclosed zero internal safety evaluations and 23 had no third-party testing. Nobody has asked the industry to demonstrate that its authorization architecture holds at the outcome level. So nobody has.

Why isn’t adequate governance required? Because AI regulations, where they exist, are still working within the frame of operations-level oversight. For instance, the EU 

 mandates effective human oversight for high-risk systems but never specifies at what level that oversight must operate. Singapore's 

Model AI Governance Framework for Agentic AI

 is more than a checkpoint proposal: it calls for upfront risk assessment, bounded permissions, human accountability, technical controls, testing, and continuous monitoring. But its governance logic still largely depends on identifying salient risks, high-stakes actions, or intervention points before harm accumulates. That leaves a harder problem unresolved: how should a system detect when a sequence of individually permissible operations is drifting toward an outcome no human actually authorized?

analysis of human-in-the-loop requirements

 put it: "The uncomfortable reality is that human review does not stop machine-speed failures. At best, it explains them after the damage is done." Oversight that operates at the operation level explains what happened. It does not prevent accumulative damage.

Some researchers have started describing what outcome-level governance would require. Google DeepMind's 

 proposes that you should not assign a task unless its outcome can be automatically verified. Their proposed mechanism, called Delegation Capability Tokens, attaches a specification of permitted authority at each boundary in an agent's plan, derived from the task above rather than inferred from the phrasing of a natural language request. Engin and Hand's 

 takes a regulatory angle, tracking decision authority, process autonomy, and accountability as they shift in real time. It asks whether an agent's cumulative authority is drifting toward dangerous thresholds, not whether individual operations got checked off.

In software development, some tools are already moving this way. 

 ask users to approve a desired end state before any code is written—outcome first, operations second. Lemkin asked Replit's agent to build a database; Yue 

 OpenClaw to "suggest what you would archive or delete, don't take action until I tell you to." Lemkin's request had no formal outcome specification. Yue's did, and the agent's memory system dropped it. Neither interface treated the constraint as durable.

The DeepMind framework implies a test worth applying more broadly: if the expected outcome cannot be formally specified, the instruction has not defined a task—it has granted a license. By that standard, most open-ended AI agent use is more like licensing, not delegation. And when outcomes go wrong under a license, the governance question changes. It stops being about how to improve oversight of a formally specified outcome. It becomes about who bears responsibility when no outcome was ever specified to begin with.

A Meta sign at Meta Headquarters in Menlo Park, California, in January 2025.

Tech Workers Are Fighting Against Silicon Valley’s AI Push

Brazilian President Luiz Inacio Lula da Silva holds a cabinet meeting at Planalto presidential palace in Brasilia, Brazil, Wednesday, June 3, 2026. (AP Photo/Eraldo Peres)

Big Tech's Proxy War Against Brazil's New Platform Accountability Rules

Tech Policy Press - Technology and Democracy

Tech Policy Press is tracking laws and regulations, along with government investigations and litigation, that will shape the rules and accountability for tech companies.

Latest

Podcast

Recent

Our content delivered to your inbox

Thank you!

Around the globe