To Curb Online Harm to Children, Tax Amplification

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: bans. The UK is committed to under-16 restrictions for spring 2027. France is attempting to legislate for under-15s. Australia's ban 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, recently published in The BMJ, 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. Economists call this a market failure, a situation where the producers keep the profit while the costs fall elsewhere. The standard correction, sometimes called polluter-pays, 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 tax on harmful amplification would.

Who pays for the harm

The Guardian has called this technology's "big tobacco moment." 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. For example, 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 World Happiness Report 2026 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 mispricing: 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 Ofcom 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. US civil juries in the New Mexico Attorney General's case against Meta and the K.G.M. trial in Los Angeles have begun treating recommender system design as a defective product — 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.

Why bans are the wrong instrument

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. The Australian data 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 Online Safety Act 2023 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. Paul Romer has argued for a progressive tax on digital advertising revenue to push platforms toward subscription models and away from attention extraction. Daron Acemoglu and Simon Johnson 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 Harmful Content Exposure Rate (HCER)I which 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.

What the engagement economy needs

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 UK's Soft Drinks Industry Levy prompted manufacturers to reformulate products 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.