Arbitration clauses are reasonably common in Terms of Service and EULAs. The logic has always been that keeping disputes out of court, keeping them more confidential, avoiding class actions and resolving claims faster and cheaper, is good.

But recent cases involving Valve, Epic Games and Roblox illustrate the unintended risks of arbitration provisions.

A Clause That Becomes a Target

Valve, in its terms for Steam, built what looked like a fair arbitration clause. To make dispute resolution accessible to all users, Valve agreed to pay filing fees of roughly $3,000 per claim. That provision became the vulnerability.

Starting in 2021, a plaintiffs’ attorney recruited more than 50,000 Steam customers through social media advertising and threatened to file individual arbitration demands on each of their behalf. At $3,000 per case: 50,000 claims meant $150 million in filing fees before a single hearing on the merits. The clause designed to protect users was being used as a financial weapon against the company that drafted it.

Valve ended up fighting on two fronts simultaneously. They fought thousands of individual arbitrations and a parallel class action in federal court. In September 2024, after years of litigation, Valve scrapped its arbitration clause entirely and replaced it with a court-only requirement. That move was challenged too, with the plaintiffs’ attorney arguing Valve could not retroactively pull users out of a forum they had already entered. The dispute remains unresolved across both venues. Amazon faced a nearly identical situation in 2021 and also abandoned its arbitration requirement under similar pressure.

A Clause That Creates Uncertainty

Epic Games and Roblox had a different experience. In a May 2026 ruling, federal court in Pennsylvania compelled arbitration in a suit alleging both companies designed their games to be addictive and targeted toward children.

The plaintiff argued that the clauses were unconscionable and that the minor at the center of the case could not be bound by the agreements. The court sent those questions to the arbitrator anyway, because both Epic’s and Roblox’s agreements included delegation clauses. These are provisions that hand questions of enforceability to the arbitrator rather than the court.

On one level, this represents a win for Epic and Roblox: the clause held and the case went to arbitration. But the outcome also illustrates how critical specific drafting choices can be, choices that many companies never think carefully about. Delegation clauses are powerful, but they are not necessarily standard across all arbitration provisions. Had the Epic or Roblox terms not included this clause, the court might have ruled on enforceability itself and that case may have turned out differently.

The case also raises the question of whether a minor can be bound by arbitration terms at all. This question was ultimately deferred to the arbitrator rather than answered by the court. But it’s going to come up again, in another case, with a different set of facts and possibly a different result.

The Structural Problem

What connects these cases is a set of structural tensions built into how arbitration clauses are typically drafted.

Provisions meant to make arbitration accessible (paying filing fees, keeping the process affordable) can be exploited at scale in ways that were never anticipated when the clause was drafted. Clauses that look robust for a user base of thousands may be tested in ways nobody modeled when that base grows to millions.

Some Best Practices

Given what these cases reveal, a few practices are worth building into any arbitration clause from the start.

Fee structures need stress testing. Valve’s fee-payment commitment was reasonable in isolation but didn’t work at scale. Had they modeled out the extremes, they might have avoided costly litigation. Hindsight is 20:20, though, and few anticipated this risk.

Mass arbitration procedures should be explicit. Many agreements now include “bellwether” provisions. These create a process where a small number of representative cases are decided first, with the outcomes informing settlement of the remaining claims. This mirrors how complex mass litigation is often handled in court and can dramatically reduce the cost and chaos of thousands of simultaneous proceedings. The key is making sure those procedures are clearly drafted, genuinely bilateral and likely to survive a court challenge. Provisions that look like a thinly disguised class action ban tend not to hold up.

Delegation clauses are critical. As Epic and Roblox’s cases show, including a clause that delegates enforceability questions to the arbitrator gives the agreement a meaningful layer of protection. Without it, a judge deciding that the clause is unconscionable or unenforceable as to a minor can terminate the arbitration entirely before it begins.

Companies should be thoughtful about the jurisdiction designated for arbitration. The seat of arbitration determines which law governs the process, which courts have supervisory authority and how an award can be challenged or enforced. For US-focused consumer agreements, this question is less pressing. The Federal Arbitration Act covers most of the ground. But for companies with meaningful international user bases, or co-dev partners or other contractors based in foreign jurisdictions, the seat matters considerably. A London seat brings English arbitration law and courts that are generally deferential to the process. Singapore and Geneva are similarly arbitration-friendly and are often preferred for disputes involving Asian or European users respectively. A seat chosen by default, or copied from another company’s agreement without thought, may not match the actual user base or the enforcement environment the company cares about.

Companies should consider which claims should be carved out of arbitration. For example, IP claims, injunctive relief and regulatory disputes often do not belong in arbitration. A clause that sweeps everything into arbitration can create friction when the company actually needs fast court intervention.

Terms should be reviewed and updated periodically. As companies grow, terms that were once appropriate may no longer be adequate for the company’s needs.

Alternatives Worth Considering

Arbitration is not the only option, and for some companies it may not be the right one.

A traditional litigation process with a forum selection clause is what Valve ultimately landed on. Litigation is slower and more expensive per case, but it does not carry mass arbitration fee exposure, and a well-drafted forum selection clause (specifying jurisdiction and venue) still gives a company meaningful control over where disputes are heard. The trade-off is giving up confidentiality and accepting the pace of civil litigation.

Tiered dispute resolution may be a middle path. This model requires negotiation first, then mediation, and only then arbitration or litigation. This adds friction that often filters out low-stakes claims without the all-or-nothing bet of mandatory arbitration.

At the end of the day, no dispute resolution structure is without risk. The goal is not to find a clause that eliminates exposure completely, but rather to make a deliberate choice about which risks are acceptable given the specific product, user base and legal environment. This requires revisiting terms regularly, not drafting once and filing them away.