Illustrations of six data workers, working at computers in isolation from each other. Painted background includes hazy image of cubicles; digital overlay of glass fractures.

Labor Power and the Role of Subcontracting in the AI Economy 

Ambiance at the Global Dialogue on AI Governance, Palexpo, Geneva, Switzerland - 07 July 2026.

Global Dialogue on AI Governance Dialogue on AI Governance

Can the United Nations Bring the World Together on AI?

Illustration of two black-and-white halftone hands reaching toward a stylized street map, with a large red location pin marking a spot on the map.

Breaking Down the Implications of the Landmark Chatrie Ruling

United Nations Secretary General Antonio Guterres speaks about the release of a UN report on artificial intelligence during a news conference at UN headquarters, Wednesday, July 1, 2026. (AP Photo/Jason DeCrow)

Why Fear Matters in Governing Military AI

The image is a conceptual illustration of data analysis. It features a graph with nodes connected by lines on a sheet of paper, held by a hand. Another hand holds a magnifying glass that enlarges a section of the graph. The background includes abstract shapes in yellow and blue, adding a playful, creative element. The composition suggests careful scrutiny or examination of trends or data points.

To Curb Online Harm to Children, Tax Amplification

A police officer checks the list of a polling station during the parliamentary elections in capital Nicosia, Cyprus, Sunday, May 24, 2026. (AP Photo/Petros Karadjias)

EU Political Ad Rules Leave Elections Open to Foreign Manipulation

Palestinians carry food boxes delivered by the World Food Programme (WFP) in Khan Younis, southern Gaza Strip, Sunday, Nov. 2, 2025. (AP Photo/Jehad Alshrafi)

Uncovering the Humanitarian and Nonprofit Sector's AI Governance Crisis

This image depicts clusters of various mushrooms growing on a network of surveillance cameras against a dark green textured background.

India is Building Surveillance Infrastructure on Broken Data and Bad Policing

Image of map and computer language over binary codes on server racks in server room.

Global Digital Policy Roundup: June 2026

European Commission President Ursula von der Leyen speaks during the Defence Industry Forum at the 36th NATO (OTAN) Heads of State and Government Summit in Ankara, Turkiye on July 7, 2026. (Photo by Alexis Jumeau/Sipa via AP Images)

EU Child Safety Panel Tests Von der Leyen's Ban Resolve

In this image taken from video provided by Russian Presidential Press Service on Friday, July 3, 2026. Russian President Vladimir Putin speaks while visiting one of the command posts of the Joint Group of the Russian Forces, in an undisclosed location. (Russian Presidential Press Service via AP)

Why Europe’s Safeguards Against AI Disinformation Won’t Stop Russia’s Next Move

Synthetic or slop? by Marcin Wilkowski / Better Images of AI / CC by 4.0

Synthetic or slop? by Marcin Wilkowski / 

The Next National Security Challenge Is Research Integrity

The United Nations headquarters in New York. Shutterstock

The United Nations headquarters in New York. 

The UN Scientific Panel on AI's Preliminary Report Does Not Establish Its Independence

The Tower of Babel by Lucas van Valckenborch, 1595. Wikimedia Commons

The Tower of Babel by Lucas van Valckenborch, 1595. 

The Pope Found Babel in AI. Here's What Rabbis Saw

Abstract microscopic photography of a Graphics Processing Unit resembling a satellite image of a big city.

India’s Aadhaar Shows Foreign Dependencies Reach Beyond US-China

Illustration of two men inflating a balloon labeled "tech"

Antitrust Enforcement Can Deflate the AI Bubble Before the Public Pays

Anthropic CEO Dario Amodei (not pictured) and Chief Product Officer Mike Krieger unveil Claude 4 during the Code with Claude conference, announcing both Opus 4 and Sonnet 4 AI models to developers and industry leaders on Thursday, May 22, 2025 in San Francisco. (Don Feria/AP Content Services for Anthropic)

What AI Companies Can Learn from the Oversight Board

A data center and its backup diesel generators built by the Markley Group loom over a ballpark and residential neighborhood in Lowell, Mass., on June 30, 2026. (AP Photo/Matt O'Brien)

Governments Can Advance 'Greener AI' Through the Power of Procurement

PICTURE POSED BY A MODEL File photo dated 21/08/14 of a child using an Apple iPhone smartphone. Schools cannot address issues with children's smartphone use and screen time alone, headteachers have warned. Delegates at the annual conference in Belfast for the NAHT school leaders' union voted in favour of a motion calling for national guidance on online safety to help families establish boundaries for their children with screens. Issue date: Saturday May 2, 2026. 84481834 (Press Association via AP Images)

Europe Can Protect Children Online Without Surveillance or Age Bans

A woman is passing by the logo of Europese Commissie - Commission Europeenne - European Commission, at the Berlaymont building, the building of the European Commission Brussels, Friday 17 April 2026. BELGA PHOTO MARIUS BURGELMAN (Photo by Marius Burgelman/Belga/Sipa USA)(Sipa via AP Images)

(Photo by Marius Burgelman/Belga/Sipa USA)(Sipa via AP Images)

Europe's W Social Bet Tests its Vision of Digital Sovereignty

 marks one of the Court’s most important Fourth Amendment rulings in the 21st century. While the Court sent 

 back to the Fourth Circuit Court of Appeals to decide whether this particular geofence warrant was “reasonable” (and thus Fourth Amendment-compliant), it ruled that this form of surveillance is a search that requires a warrant. Specifically, the Court held that individuals retain a reasonable expectation of privacy in their location history even when held by a third party (in this case Google) and when the surveillance only covered a short duration of time.

The Court’s holding will have crucial ramifications for privacy in the digital age, lifting up a shield against new surveillance technologies and techniques far beyond the geofence warrant that was at issue in this case. But on other emerging surveillance issues related to this case, the future for Fourth Amendment rights remains uncertain. Below, we outline some of the key impacts and open questions from 

, and how the decision will influence the broader landscape for digital privacy rights.

 mark the long-awaited death of the third-party Doctrine?

Throughout the digital age the third-party doctrine—which since its recognition in the 1970s has generally held that records and data held by third parties do not retain Fourth Amendment protections—has been an enormous thorn in the side of digital privacy rights. If applied to modern digital tools, this doctrine would prevent the Fourth Amendment from guarding individuals’ emails and text messages from warrantless searches merely because they’re stored on a server by a digital service provider. This situation was not compatible with democratic society, but protections 

 allowed the Supreme Court to dance around this question for years.

But as the amount of digital services we rely on for our daily lives has become practically endless, the need for clear, strong protections has grown urgent. In today’s world most people casually rely on third parties not just to store in the cloud but to digitally streamline a huge array of highly sensitive information: not just our emails and direct messages, but also our location and movement records, calendars, photos, daybooks and journals, health and medical logs, web browsing activities, and much more.

, the Court tosses the third-party doctrine to the dustbin for these types of digital services. Building on logic it laid the foundation for in the 2018 

 decision (where the Court ruled any cellphone tracking of at least a week required a warrant), 

 holds that the third-party doctrine does not apply contingent on a two-pronged test: 1) whether the data shared are highly sensitive, and 2) whether third party’s own control and use of the data was the purpose behind sharing it.

For the first prong, it’s inevitable much ink will be spilled in the future debating whether different types of data clear the hurdle. But it will be a debate over privacy rights held on the most logical terms: How sensitive and revealing is a certain type of data? And how dangerous would it be for the government to access those data without a warrant?

On the second prong, the Court effectively draws out what makes modern digital services so different from the type of entity the third-party doctrine was built for. What matters most for privacy rights is distinguishing whether third-party access was the desired goal. Was data, as the Court puts it, shared “in the normal sense of 

 a third party to see or use it”? (emphasis added). Or, as is often the case for digital services, was “the exposure of that information to [a company] … merely what happens when a user avails himself of one of the services on his cell phone,” whereby “a cell-phone user is not to be viewed as sharing private information with third parties—which then can be freely passed on to the government—just by doing the ordinary things cell-phone users do.”

Are individuals handing over records or data to a third party to be an active and intended sharee, or merely as a steward of those items? Going forward that—rather than a blanket, antiquated rule for all data held by third parties—is the essential distinction.

 will be its holding on how the Fourth Amendment protects electronic location information. When it issued its Carpenter ruling in 2018, the Court said that cellphone tracking for a week or longer required a warrant, but explicitly left open the question of what protections applied for shorter periods of time.

 firmly establishing Fourth Amendment protections for cell phone location tracking of any length. By recognizing that “the durational bounds on the data actually acquired do little to address the Fourth Amendment’s concern about a too permeating police surveillance,” the Court provides clarity for, and more importantly a strong shield against, cell phone tracking that is short in length yet may still be highly invasive in what it reveals.

This is a major boon to privacy rights, although one that is not without caveats and ongoing questions. First, the Court’s ruling rested in part on the logic that geofence warrants allow the government to place in its crosshairs those locations and events that are especially sensitive, a technique that “enables police officers to focus on precisely those sites” like a house of worship, lawyer’s office, medical clinic, protest, or political rally. Does a warrant protection still exist if this target-the-location tactic isn’t being employed, such as checking the cellphone movements for a suspect over an hour of time when a crime occurred? The answer is likely yes, given the Court’s declaration that “where the Fourth Amendment applies, it applies—regardless of ‘the quality or quantity of information’ the government obtains.” But this may be a point of ongoing litigation in the future.

A broader question is how this ruling will apply to other forms of electronic location surveillance than cellphone tracking. If a warrant is required to track someone’s movement via their phone, does that also apply to mapping someone’s movements with automated license plate readers? What about using drones, especially those paired with automated 

? Or what if rather than by using a geofence, police want to identify everyone at a protest using facial recognition?

 ruling did contain one qualifier that could cut against extending the law in some of these situations. In distinguishing this case from a 1983 ruling (

) that tracking cars via a beeper did not require a warrant, the Court emphasized that the fact that “the surveillance there was confined to public roads …. was crucial to the Court’s decision” whereas geofences and cellphone tracking can roam into private residences. Will this be a key distinction, whereby electronic location tracking into private areas requires a warrant but those methods that are limited to public spaces do not?

It’s hard to imagine this will be the case. After all, it would be both illogical and anathema to privacy rights if the Fourth Amendment required a warrant to track everyone while they were inside a church, an abortion clinic, or a gun shop (or host of other sensitive locations), but allowed the government to freely track everyone as they publicly entered and exited those places.

Exactly what types of location tracking technologies will require a warrant in line with 

 will be an important topic for years to come. But for now we have resolution on a matter of great importance: any cellphone location tracking requires a warrant. That is a major boon to privacy, civil rights, and civil liberties.

 mean for geofence warrants in particular? The ramifications of 

 on geofence warrants are, potentially, equally significant and limited. The geofence warrant in this case demanded location data maintained by Google through its Location History feature. That data is now stored locally, meaning Google won’t have responsive information to any geofence warrants issued now. But for other businesses that still hold precise location data, 

 should provide a degree of clarity. If law enforcement attempts to issue a geofence in the future (or employs new investigative techniques that similarly requests location data for a group rather than an identified suspect) companies now have a clearer roadmap to demand a warrant and push back on demands that are overbroad, burdensome, or otherwise inappropriate. However, as Google was the main known recipient of geofence warrants, it’s unclear how broad the practical impact of this will be.

The strength of guardrails around future geofence warrants is another open question. The 

 ruling expressly left for lower courts how to evaluate the proper bounds that would make acting on a geofence warrant a “reasonable” search. One privacy‑protective view would require probable cause that everyone within the geofence committed the crime, a standard that would effectively bar most geofence warrants which often capture innocent bystanders, witnesses, and even unrelated individuals in entirely different buildings or locations who merely happen to be within the fence.

A more permissive approach would be a model suggested by Justice Kavanaugh in oral arguments: to require probable cause that everyone within the geofence would provide evidence as either a suspected perpetrator or a witness to a crime. That standard has legal precedent (in 1978 the Court 

 that a search could occur even if not focused on a suspect so long as there is probable cause it would return evidence) and would provide more flexibility for investigators. But it also risks some abuse in terms of fishing and “surveillance by association” if used in response to crimes that occurred at sensitive events, such as a protest.

Another approach would be to treat each inquiry on a case‑by‑case basis depending on the specific circumstances, the geographic scope of the geofence, its duration, and the degree of judicial oversight at each step. That approach offers greater flexibility to account for specific needs and civil liberties risks, but also increased risk that needed protection won’t be consistently applied. The Court did not choose among these possibilities, and as such the contours of a constitutionally adequate geofence warrant remain unsettled.

By declining to set parameters for geofence warrants—and, in particular, establish what principles and limits should apply to a search focused on a group rather than an individual suspect—the Court left open a broader emerging question about “reverse searches” generally. The reverse search category includes reverse keyword searches, reverse IP address searches, reverse AI prompt searches, and other techniques that identify users based on shared conduct (e.g., the websites they visit, the search terms they use, or the addresses they look up) rather than individualized suspicion.

 makes clear that these types of sensitive user data can receive Fourth Amendment protection even when held by a company. The Court did not decide how probable cause and particularity apply to reverse searches, nor did it offer a roadmap for safeguarding innocent individuals incidentally caught up in this form of surveillance. It did not articulate limits on how much discretion officers may exercise when narrowing anonymized data, or minimization structures more generally. As reverse warrants expand and law enforcement officials begin to use new techniques with increasing frequency, lower courts, and specifically magistrate judges who might be unfamiliar with these processes, will have to determine how 

’s holding applies to increasingly complex forms of digital behavior—and how to reconcile those techniques with the Fourth Amendment’s traditional requirement of individualized suspicion.

 draws forth is how to deal with the data broker loophole. In recent years it has become a serious problem that law enforcement and intelligence agencies avoid the court approval process for data collection entirely, opting to simply buy sensitive data from brokers instead of compelling disclosure. The Court’s explicit protection for cellphone location data—and implicit protection for many other types of digital data held by third parties—highlights the importance of addressing this issue.

It should also close the door on arguments that such data are not Fourth Amendment protected. Perhaps even more so than data-sharing features users knowingly accept, data being harvested and funneled to brokers in unexpected and all-too-often unscrupulous ways (such as sensitive location data being stockpiled and sold by a 

 app) seems even more acutely to reflect the sentiment of 

 that “the exposure of that information to [a company] is merely what happens when a user avails himself of one of the services on his cell phone.”

, “a cell-phone user is not to be viewed as sharing private information with third parties—which then can be freely passed on to the government—just by doing the ordinary things cell-phone users do,” it makes little sense to say a dispositive and overriding factor should be whether or not the government paid the third party for the data. While the Court did not explicitly address the data broker loophole in its ruling, it does seem to have laid the foundation for closing it.

However, it's unfortunately unlikely that the judiciary will resolve this issue soon. Even as 

 bolsters Fourth Amendment protections for digital data, the government will likely continue to rely on other arguments, such as ICE’s 

 that purchasing location data tied to Mobile Advertising IDs (which can be de-anonymized) rather than phone numbers nullifies any Fourth Amendment concerns.

There is also a serious practical obstacle: because exploitation of the data broker loophole is cloaked in secrecy, it is difficult to discover particular uses that could then be challenged in court. And if such a challenge did occur, it would take years for litigation efforts to resolve the issue (the geofence warrant at issue in 

 that now returns to the circuit court for further evaluation was issued over seven years ago). The faster path would be for Congress to pass legislation such as the 

Liz Carolan is a fellow at Tech Policy Press.

 identifies serious vulnerabilities in Europe’s new political advertising regulatory framework ahead of crucial elections in France, Spain, Italy and Poland.

 — Mediterranean Digital Media Observatory — major technology companies, including Meta and TikTok, repeatedly sold political advertisements to election candidates and anonymous third parties in the European Union despite saying they had banned the practice. Many of these ads were purchased through candidates' official accounts in the weeks before the election. Others were anonymous campaigns that spread false claims about opinion polls in the days immediately prior to voting day, a practice prohibited under 

. These findings raise questions about companies’ compliance with the 

Transparency and Targeting of Political Ads

 (TTPA) regulation. Researchers fear that the gaps could allow foreign actors to exploit the systems to influence elections in Europe, which next year will include votes in France, Spain, Italy and Poland.

 announced last year that they would each be banning the purchase of political ads on their platforms in the EU. These bans came in response to the TTPA, a new regulation mandating that platforms label political ads, maintain archives of them, and provide information on spending and targeting. It took effect last October, nearly a decade after the Brexit referendum and the 2016 US presidential election intensified scrutiny of political advertising on social media.

 by Cypriot academics found that the platforms are not consistently enforcing their own policies and that election-related advertising continues to be sold in Europe. Speaking to Tech Policy Press, author of the research, Michael Sirivianos and Andreas Andreou, said that despite these announced bans, political ads “are not yet a well policed environment” and that detection mechanisms “are not there yet.”

The researchers used machine learning tools to monitor the commercial ad libraries and APIs of three platforms, Meta, Google and TikTok, searching for ads containing the names of the candidates and parties, and starting at the commencement date of the TTPA. The researchers were not able to access the required data from X.

Their analysis was based on the official list of candidates for Cyprus' 2026 parliamentary election, which included 753 candidates representing political parties across the country.

The researcher found that Meta platforms allowed at least 2,457 ads to pass internal checks and be pushed to voters. Many of these ads were purchased by candidates' official accounts in the weeks before an election in which they were participating. Candidates who had already had ads found to be in violation of the rules were allowed to continue posting ads unimpeded, including one case where a candidate continued to run ads after having received 19 violations.

Researchers also identified five anonymous campaigns spreading falsehoods about candidates and about polling to voters in the days before the election. Cypriot electoral law prohibits campaigning in the days before polling stations open, and bans the publishing of any polling in the seven days before a campaign.

A structural problem that MedDMO identified was that enforcement on Meta appeared to be at the level of the content, not the identity of the advertiser. Ad purchasers were asked if they were running a political ad, but no checks were made to see if the advertiser was a political candidate. Google takes a different approach, verifying at the level of the account, with the researchers finding that “Google's identity-verification requirements substantially constrained political-advertiser entry.” They identified only one advertiser that violated Google's political advertising ban, running 15 ads that were categorized as "Arts & Entertainment" or "Jobs & Education."

had a global moratorium on the sale of political ads

 on its platform since 2019, though concerns have been raised about its effectiveness. The European Commission has 

opened official proceedings against TikTok

 for failing to adequately enforce this ban, driven in part by a scandal that saw Romania cancel its most recent presidential election due to concerns over manipulation via paid campaigns on TikTok.

The researchers' review of TikTok’s API found that the moratorium was not being fully enforced in the Cypriot election. They identified 887 political advertisements from 67 advertisers spanning all 15 registered parties.

The capacity of platforms to identify and remove policy-breaking ads was a particular concern in Cyprus. The research found that just a single removal occurred on TikTok during the first 220 days post-TTPA implementation (up to May 16th). On Meta, less than 1% of the ads identified during the election were removed while they were active.

Instead, a pattern emerged of ads being taken down following MedDMO’s interactions with the companies about these ads via an EU coordination mechanism. The researchers noted in their report that in the 30 hours following their meeting with Meta, “687 takedown notices appeared,” which was “172 times the prior single-day record.” On TikTok, they noted that “631 removals followed within five days” of their meeting.

The findings raise questions about the extent to which adequate resources were being put into companies enforcing their own bans (a requirement under the DSA) and into compliance with the new TTPA. The MedDMO team estimated that the project took about one month each of two people’s time, which was covered partially by EU grants, by their university salaries or their own volunteer time. No platform funding was provided.

For Sirivianos and Andreou, the broader concern is that the enforcement gaps they identified could be exploited by foreign actors seeking to influence elections across Europe. As they write in their report, “non-compliance with the political advertising regulatory framework leaves an enforcement vacuum primed for Foreign Information Manipulation and Interference actors to exploit.”

Tech Policy Press contacted Meta for comment on the findings and its election enforcement efforts, but the company did not respond before publication.

Governments are banning children from social media. The bans are failing because they regulate the population bearing the cost rather than the systems doing the harm.

Governments across Europe and Asia are reaching for the same instrument to reduce online harms: 

 has been in place since December 2025, and similar measures are advancing in 

Slovenia, Poland, Spain, Denmark and Malaysia

. The first prospective evaluation of Australia's regime,

, shows what happens when the wrong instrument is applied to the right problem: 85 percent of under-16s were still using restricted platforms three months in, a third had encountered no age verification at all and attention had begun migrating to less-regulated environments.

The problem the bans are trying to fix is real. 

Platforms profit from producing harms — to children, to parents, to schools, to public health services — that other people pay for. 

a situation where the producers keep the profit while the costs fall elsewhere. The standard correction, sometimes called 

, is to make the producer pay for the harm it causes. Tobacco taxation, carbon pricing, sugar levies and emissions trading all rest on this principle. Each worked. Bans do not, because they leave the producer's incentive to generate the harm in place. I argue that a 

 The phrase captures something true about the cultural moment, but it stops short of an economic diagnosis. The problem is not fundamentally that children are getting through age gates. It is that the platforms they reach are running optimization functions that, by design, generate externalities the platforms do not pay for. A recommender system is optimized for engagement rather than user welfare: it 

maximizes the probability that a given user will spend more time on the platform

video platforms optimize watch time and time spent on the app

. And time on the platform is monetized through advertising auctions. The recommender's optimization target is therefore, indirectly, the firm's revenue.

The optimization function does not include the cost to the user of being optimized against. It does not include the cost to third parties — schools, parents, public health services — of the downstream effects of the attention the platform extracts. These costs are real and measurable. The 

 finds that young people who use social media for less than an hour a day report the highest life-evaluation scores, while average daily use among under-25s is 2.5 hours. Over the past decade, life evaluations among under-25s in the United States, Canada, Australia and New Zealand have fallen by almost a full point on a 0-to-10 scale, while the same measure for under-25s elsewhere in the world has risen. The report itself characterizes this pattern as a negative externality, a cost the platform generates but others pay. Many young people use these services only because the rest of their social world does, while reporting that they would prefer not to. That is what welfare economics calls 

: the social cost of consumption exceeds the private cost the consumer pays.

The mispricing has measurable downstream consequences. 

The UK's Internet Watch Foundation documented a 26,362 percent year-on-year increase in photorealistic AI-generated child sexual abuse material

 — a category of harm that did not exist three years ago, and one that is circulating on a range of sites on the open internet, as well as the dark web. The UK's 

 estimates that 72 percent of children aged 8 to 12 bypass safety controls to reach adult-tier platforms, where recommender systems 

are known to surface self-harm and eating-disorder content.

 have begun treating recommender system design as a 

 — legally comparable to a faulty car or a contaminated drug.

None of these harms is borne by the firm that produces them. All are borne by users, third parties and public services. This is what an unpriced externality looks like.

The standard economic response to a market failure depends on the structure of the failure. Quantity restrictions, like bans, work when consumption itself is the problem, and consumers can be reliably excluded from the market. They do not work when the harm is produced by the firm's design choices and consumers can easily bypass the restriction at low cost. 

 demonstrates the latter case: 85 percent bypass within three months, attention migration to less-regulated environments including AI companions and messaging apps, and no significant change in access among under-16s.

The bans also leave the optimization function untouched. A platform fully compliant with an age restriction continues to run the same recommender on the remaining user base. It continues to extract the same attention from the users it is permitted to retain. It continues to generate the same externality, scaled to a slightly smaller market. The age restriction is, in economic terms, a quantity cap applied to the consumer side of a market whose production side is unchanged. This is not how welfare economics says you fix a producer-side externality.

The same structural critique applies to the existing patchwork of duty-of-care regulation. The UK's 

 requires age checks, mandates user reporting and lets Ofcom impose substantial fines. The EU's Digital Services Act mandates systemic-risk assessments and gives regulators access to platform data. Both regimes specify what platforms must not do. Neither changes the financial incentive that makes the prohibited conduct profitable in the first place. A platform that fully complies with both regimes can continue to make exactly the same money from exactly the same engagement-driven amplification, right up to whatever the legal line happens to be.

Why taxing harm does what taxing revenue does not

Several economists have proposed taxing digital platforms. 

 has argued for a progressive tax on digital advertising revenue to push platforms toward subscription models and away from attention extraction. 

 have argued for a similar instrument. These are very important proposals. They address a real concern about the structure of the digital advertising market, including its role in driving misinformation and political polarization.

But ad-revenue taxes target the wrong variable. They tax the proxy (revenue from advertising) rather than the externality itself (the harm produced by amplification). A platform that reduces its advertising revenue without reducing harmful amplification pays less under a Romer-style instrument. A platform that increases harmful amplification while holding advertising revenue constant pays the same. The pricing signal is not aligned with the social cost the tax is meant to internalize.

A properly Pigouvian instrument, named for economist Arthur Pigou, is a tax set to match the cost that an activity imposes on other people. It targets the externality directly. 

I have previously proposed a Digital Harm Levy

 calibrated to a measurable indicator: the 

I have outlined in a detailed policy brief

. HCER would measure the proportion of total user impressions on a platform that fall within harm categories specified by an independent regulator. A platform that reduces HCER pays less. A platform that increases it pays more. The instrument does not care which content is shown or who shows it. It cares about what the recommender system selects and delivers at the population level.

The pricing signal is now aligned with the harm. A platform can reduce its tax liability through any of the levers it controls: demoting harmful content, redesigning the recommender, slowing the viral propagation of content that has signaled early engagement, introducing friction on attention-capture features. None requires removing content. None requires changing what users post. They require changing what the optimization function optimizes against. This is the difference between regulating users and regulating the producer's incentives.

The instrument is also technically tractable. Major platforms already calculate similar internal metrics — YouTube's 

Violative View Rate, an estimate of the proportion of video views that violate YouTube’s Community Guidelines,

 is the closest public analogue. I propose that the levy could be integrated with the existing UK Digital Services Tax architecture. The audit infrastructure required is structurally similar to what is already being built under Article 40 of the EU Digital Services Act, which mandates vetted-researcher access to platform data for systemic-risk study. The institutional and technical groundwork is largely in place. Implementing the levy is a question of regulatory will, not of feasibility.

The Guardian's “big tobacco moment” framing is right. It is also incomplete. The real breakthrough in tobacco regulation came not when governments warned consumers, but when they changed the incentives facing producers, making harmful business models progressively more expensive to sustain. The same logic has shaped successful regulation elsewhere. For example, the 

rather than simply asking consumers to exercise more willpower, while 

carbon pricing has reduced emissions by making pollution itself more costly

. In each case, the structural correction was a producer-side instrument that priced the externality rather than relying on individuals to make better choices. Clearly, taxation played a major role, but also 

smoke-free laws in restaurants and bars, along with public health advertising, were also strongly correlated

The engagement economy now sits where the carbon economy sat 30 years ago: a productive market whose private profits depend on social costs the market does not pay. Bans target users. Duty-of-care regimes target conduct. A tax on harmful amplification has not yet been tried. It is the one instrument that targets the externality at source.

Speech remains free in a Pigouvian regime. Amplification of harmful content stops being free.

 returned to the negotiating table to discuss artificial intelligence—an issue as consequential this century as nuclear weapons were during the Cold War. But AI poses a fundamentally different 

. Its capabilities are difficult to measure, evolve at extraordinary speed, and are increasingly intertwined with economic competitiveness, military power, and the strategic interests of a handful of technology companies. The emergence of two AI superpowers facing shared risks without shared rules presents one of the most urgent governance challenges of the twenty-first century.

Among these risks, the growing use of autonomous AI systems in military decision-making deserves particular 

. Unlike traditional weapons, such systems may operate at speeds and levels of complexity that exceed human understanding and effective oversight. As machines take on greater roles in surveillance, targeting, and battlefield operations, the risk of errors, accidents, and unintended escalation increases. The prospect that critical decisions about the use of force could be delegated to algorithms raises profound questions about accountability, human control, and international security. In short, extreme military AI risks include high-consequence outcomes arising from autonomous systems operating beyond effective human oversight.

History suggests that governance often emerges in response to 

. Nuclear arms control, environmental regulation, and public health institutions gained support only as societies recognized the scale of the threats involved. Fear of nuclear catastrophe helped mobilize public backing for arms-control agreements, while international pressure from governments, scientists, and public health organizations led to the 1972 

, which prohibits the development, production, acquisition, transfer, stockpiling, and use of biological and toxin weapons.These examples show how scientific concern and international cooperation can establish global norms against high-risk technologies.

Similar concerns about autonomous military AI may create the political conditions for new forms of oversight and international cooperation. These concerns were on display last week, when United Nations Secretary General António Guterres 

 on “killer robots.” Without a shared understanding of the risks, efforts to build safeguards may lack the urgency needed to keep pace with technological change. We argue that philosophy can help clarify the ethical foundations needed to govern military AI before its risks become irreversible.

 recognize the risks of advanced AI, yet neither is prepared to slow its pursuit of technological leadership. Both worry that restraint would confer a strategic advantage to the other. But despite what some argue, AI is not a race with a finish line. Framing it as one may intensify tensions and increase the very risks both sides seek to avoid. As the nuclear era showed, rivalry does not eliminate the need for cooperation to prevent catastrophe. The race for advanced AI leadership is intensifying. Mirroring the Cold War, mastery of this technology and being the first to achieve “superintelligence” is now seen as a matter of survival and deterrence. However, growing recognition shows this rivalry could spiral out of control. We urgently need risk-reduction mechanisms to prevent accidents, cyberattacks, and autonomous system failures.

Perhaps the most revealing fact is this: when rival powers begin 

 for AI, the issue has clearly moved beyond the realm of science fiction. The possibility that advanced systems could be used in extreme geopolitical conflicts has become part of the real calculus of international security. Yet how can we ensure that errors do not occur—or that they are not transmitted and amplified through systems that remain only partially understood even by their creators? How can we guarantee the responsible use of technologies whose inner workings we do not fully comprehend? Is it possible to control the use of AI on the battlefield at a safe level?

, the MIT computer scientist who created ELIZA in 1966, was among the first critics of AI and its military applications. For him, AI was too important to be left solely in the hands of computer scientists and required the participation of philosophers, social scientists, and scholars from other disciplines. Weizenbaum warned that AI was a dual-use technology and that civilian advances would inevitably be adapted for military purposes. He feared that systems designed to maximize efficiency would weaken human judgment, dilute responsibility, and lower the barriers to the use of AI technology.

Recent conflicts have confirmed many of his concerns. The wars in 

 illustrate AI's growing role in targeting and battlefield automation. These examples illustrate how AI is contributing to the automation of warfare while raising profound questions about responsibility and human control in lethal operations. Despite repeated warnings from the 

, there is still no global control on autonomous weapons, suggesting that the military AI arms race envisioned by Weizenbaum decades ago is already underway.

, every technical object functions as an amplification of human action in the world. When we speak of technologies associated with warfare, we are ultimately speaking about the expansion of humanity’s capacity for destruction and deterrence. The issue today is that AI increases the psychological and moral distance between those who exercise destructive power and those who bear its consequences. If war, as Prussian military theorist 

 argued, unfolds within a fog that obscures judgment, AI may intensify that fog by automating how hostile activity is identified and classified. In this sense, it automates not only tasks but part of human cognition itself.

If politics is profoundly shaped by technology, thinkers such as 

 have long argued that technology is always embedded with political theory. It is worth remembering that Henry Kissinger first gained prominence in US policy circles after publishing his influential study on 

nuclear weapons and international politics

 in 1957. There is a clear parallel between the technological transformations of the twentieth century and the political challenges of the twenty-first. The use of AI on the battlefield highlights one of the central issues in AI regulation: the governance of catastrophic risks.

This is not about the fantasy of machines deciding to exterminate humanity. Rather, it concerns extreme global harms resulting from the human use of advanced AI systems. The core issue is the cognitive transformation that these technologies may induce in human decision-makers under extreme conditions such as armed conflict. Current AI models can already dramatically enhance the ability of malicious actors to conduct cyberattacks against hospitals, banks, power grids, and other 

. Threats associated with autonomous AI agents and the development of 

There are therefore clear grounds for state intervention. When technologies acquire strategic and military significance, their regulation ceases to be merely an economic issue and becomes a matter of global security. The challenge now is to build democratic regulatory institutions capable of involving society in the governance of advanced AI before the logic of technological competition makes such participation impossible. This will require drawing on philosophy reflection—not only to better understand the nature of AI and its relationship with individuals, but also to reflect on how democratic societies decide which risks they are willing to accept.

In 1979, with Leonid Brezhnev leading the Soviet Union and the United States still shaped by the strategic legacy of Nixon and Kissinger, the Soviet bloc invaded Afghanistan. That same year, the German philosopher Hans Jonas 

The Imperative of Responsibility: In Search of an Ethics for the Technological Age

. Jonas argued that modern technology had become a threat to the very continuity of human life and that humanity’s only reliable compass was the anticipation of danger. In an era increasingly defined by powerful technologies, foresight about risks becomes a condition for preserving humanity. Fear, far from being merely a negative emotion, serves the essential function of warning us and steering us away from potentially fatal threats.

By imagining the worst consequences of our choices, we become better able to recognize what is at stake and what must be protected. Jonas argued that in a technologically transformed world it is easier to protect humanity by identifying the evils we must avoid than by identifying the goods we hope to achieve. This insight is particularly relevant to military AI and autonomous weapons. The fear that AI systems may eventually select and attack targets without meaningful human control should not be dismissed as alarmism. On the contrary, awareness of these risks can mobilize society and stimulate democratic debate about limits, responsibilities, and governance mechanisms.

Yet security risks are only the beginning. The greater danger is that autonomy in warfare may reduce human oversight, making conflicts faster, more automated, and more destructive. For this reason, military AI regulation cannot be left solely to governments, armed forces, or technology companies. It must involve civil society and establish rules to limit the extreme risks of autonomous weapons before they exceed our capacity for 

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.

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

Latest

Podcast

Recent

Our content delivered to your inbox

Thank you!

Around the globe