Algorithmic liability and the new tobacco parallel

In March 2026, a Los Angeles jury delivered a $6 million verdict against Meta and Google in K.G.M. v. Meta Platforms, Inc. and Google LLC, marking the first of its kind to hold social media companies liable for harm caused by platform design. The plaintiff claimed that as a minor she had suffered anxiety and depression as a direct result of addictive features deliberately engineered into Facebook and YouTube. For lawyers advising technology, automotive, and financial services clients as well as private individuals, this verdict has significant potential implications. It signals that courts are beginning to hold algorithm-driven systems to account in a similar way in which they once held cigarette manufacturers accountable.

The shadow of tobacco

Early claims against cigarette manufacturers, including Green v. American Tobacco Co. (1963), were undone by causation difficulties and the absence of compelling internal evidence. Then, the disclosure of internal industry documents revealed that manufacturers had long known of nicotine's addictive properties but had chosen to suppress that knowledge. Once corporate knowledge could be proved, the litigation dynamic shifted irreversibly, culminating in a US$206 billion settlement that reshaped an industry and permanently redefined the duty corporations owe to consumers.

The structural parallels between those cases and what is happening today is remarkable. During the KGM trial, Meta employees testified that concerns about risks to children were raised internally over many years, yet the platforms were deliberately designed to maximise engagement among young users, which would in turn drive up advertising revenue. This was admitted to have been done notwithstanding the long-term psychological harm that would result. That pattern, corporate knowledge, internal suppression, and a product engineered to encourage dependency is precisely what drove tobacco litigation from isolated individual verdicts to industry-transforming group claims. 

In the EU, instruments such as the EU AI Act and the revised Product Liability Directive provide the procedural and substantive architecture for aggregated claims against platform operators. Practitioners should not be asking whether comparable claims will emerge from algorithm-related harms. They should be asking when, and in what form. Given the potential impact of adverse precedent on the business model of tech companies as well as the level of punitive damages available in the US, a large proportion of these claims will likely be resolved through confidential settlements rather than through drawn-out public trials.

From public courts to arbitral tribunals

The KGM verdict and the tobacco parallel describe the trajectory of litigation, but the considerations are equally applicable to arbitration. Behind every platform there is a chain of commercial contracts between developers and deployers, technology companies and infrastructure providers. As part of the drive towards electrification, software development in the automotive sector now rivals traditional combustion-engine engineering in importance. Where an algorithm causes harm and liability is disputed within a supply chain, those disputes will frequently be resolved by arbitration. The cross-border dimension reinforces the point: a system manufactured in one jurisdiction, integrated in another, and causing harm in a third presents precisely the kind of multi-party complexity that international arbitration was designed to resolve. 

Three fundamental questions will occupy arbitrators and judges alike.

Adjudication requires an updated skills-set

Before liability can be tested, tribunals face a more elemental challenge: establishing what actually happened. Digitally manipulated documents, fabricated images, and AI-generated evidence are appearing in proceedings with increasing frequency, and detection is no longer straightforward even for trained professionals. Authentication standards across multiple jurisdictions were not designed with synthetic media in mind. In England and Wales, the Divisional Court has already been required to address counsel submitting AI-hallucinations as authority, with wasted costs orders made and referrals to the Bar Standards Board. Arbitrators tend to be lawyers, not data scientists. When asked to establish facts using analytical tools they were never trained to deploy, the award that follows may be based on uncertain foundations.

What happens when the algorithm causes harm?

The second problem is liability attribution. As algorithmic systems are deployed across the board, disputes about what went wrong and who bears responsibility are multiplying. The autonomous vehicle sector illustrates the difficulty most sharply. When an algorithm operating across multiple legal systems causes harm, fault does not obviously rest with any single party: the developer who trained the model, the manufacturer who integrated it, the operator who deployed it, and the regulatory authority that approved it may each bear some responsibility. As the recent Tesla case in the US demonstrates, even in cases involving just one jurisdiction allocating responsibility is far from straightforward. A patchwork of national liability regimes, overlaid with EU law means the legal framework is still being constructed around the technology it is meant to govern. 

Are adjudicators sufficiently equipped to decide?

When a dispute turns on how an algorithm made a decision, can a panel of lawyers actually resolve it with confidence? Understanding whether an algorithm behaved as it should requires knowledge of how that system was built, trained, validated, and operated. The ICC Commission's Task Force on Artificial Intelligence in Dispute Resolution has identified precisely this gap: just as construction disputes bring engineers into the room and financial cases draw on forensic accountants, disputes arising from algorithmic decision-making will require standing interdisciplinary expertise within the tribunal itself.