AI Localization Governance Is Now the Hard Part

Production Is Getting Easier. Control Is Getting Harder.

At SlatorCon San Francisco in September 2026, eBay Senior Product Manager Marco Rotelli and XTM International CEO Lorcan Malone put a sharp name on something many localization teams already feel but struggle to articulate. The shift, as Malone framed it, is from a production problem to a governance problem.

Translation output is more available than ever. The bottleneck is no longer generating a translation. It's knowing whether that translation is accurate, traceable, brand-compliant, and legally defensible — and knowing who is accountable when it isn't.

That framing matters a lot if your organization works with languages where AI fallback is unreliable and where federal law requires meaningful language access.

What Governance Actually Means in Practice

At eBay, governance is not a policy document. It's a set of active controls built into the workflow. Before any AI capability goes live internally, it must pass through a Responsible AI Intake process. Translations are risk-profiled by content type. Marketplace listings move through machine translation quickly. Internal and regulated content moves through a pipeline that combines AI output with translation memories, glossaries, automated quality checks, calibrated confidence scores, and human review at defined thresholds.

The goal is not to remove human oversight. It's to direct human attention to the translations that most need it.

Malone added a warning about what happens when governance is absent or inconvenient: employees route around the localization team entirely, using whatever AI tool is sitting in their browser. He called this "shadow localization." The fix is not to block access to AI tools. It's to build an approved workflow that is fast enough and easy enough that people actually use it — while brand guidelines, terminology controls, and audit trails run in the background.

The Quality Bar Has Not Dropped

One assumption worth challenging: that AI-generated translation is "good enough" and governance is just a compliance formality.

Malone pushed back on this directly. Teams now receive AI-generated source content that is already degraded before translation begins. That translated output then needs to serve both human readers and AI agents that use it to answer queries or support decisions. The quality bar, he argued, has probably gone up, not down.

Rotelli added that relying entirely on AI to generate and translate content without human input risks making language homogeneous and repetitive. People are not just reviewers. They are the check against drift.

Why Rare Languages Make Governance Non-Negotiable

Here is where the governance conversation looks different for organizations working with languages like Chuukese and Pohnpeian.

For major language pairs — Spanish, Mandarin, French — AI systems have been trained on billions of words. Quality is still variable, but confidence scores mean something. You can calibrate when to trust the output and when to flag it for review.

For Chuukese and Pohnpeian, the training data is sparse. AI output for these languages is not just imperfect. It is unpredictably wrong in ways that a confidence score will not catch, because the model does not know what it does not know. A high confidence score on a Chuukese translation can still produce a clinically or legally inaccurate sentence.

This matters enormously for US healthcare systems and school districts serving Pacific Islander and Micronesian communities. Title VI of the Civil Rights Act requires meaningful language access for limited English proficient patients and families. "Meaningful" is a legal standard. A translation that sounds fluent but misrepresents a diagnosis, a consent form, or a special education procedural safeguard does not meet it — regardless of what any automated quality score says.

The governance framework eBay and XTM described — risk-profiling content, applying human review to regulated material, maintaining audit trails, defining accountability — is not optional overhead for rare-language pairs. It is the minimum viable process.

The EU AI Act Is Coming, and US Healthcare Is Already Regulated

Malone flagged the EU AI Act as a regulatory layer localization teams need to monitor. For US-based buyers, that may feel distant. But the regulatory environment for language access in US healthcare and education is already demanding accountability that most AI-only workflows cannot provide.

OCR complaints under Title VI increasingly scrutinize the quality and appropriateness of translated materials. Documented workflows, qualified translator credentials, and evidence of review are what protect an organization when a complaint is filed. Shadow localization — someone using a browser-based AI tool to send a message to a Chuukese-speaking patient — leaves no trail and offers no defense.

Malone's point that governance belongs with the central localization team applies directly here. The team sets the rules, holds the approved vendor relationships, and provides the audit record. The rest of the organization gets fast access to translation that they can actually trust.

The Takeaway

If you are buying translation services for a population that includes Micronesian or Pacific Islander community members, ask your vendor two specific questions. First: what is your quality control process for Chuukese and Pohnpeian, and does it include a qualified human reviewer at every stage? Second: can you provide documentation of that process if a compliance review requires it?

If the answer to either question is vague, you have a governance gap — and AI making production easier does not close it.

If you want to talk through what a compliant rare-language workflow looks like for your organization, we are happy to walk through it with you.