Pope Leo's AI Encyclical Says Technology Is Never Neutral. He's Probably Right.

Pope Leo XIV published Magnifica Humanitas on May 29, 2026, and buried inside the theological framing is a claim that should stop a commerce executive mid-scroll: 'Technology is never neutral.' Not edgy. Not new to philosophy of technology scholars. But now it carries institutional weight that vendor AI ethics pages do not, and that gap is worth examining carefully.

What 'Not Neutral' Actually Means for Your Stack

The Skeptic translation: every model you deploy carries the values of its training corpus, its RLHF annotators, and the inference constraints its vendor chose to enforce. This is not a metaphor. It is a structural property. When your recommendation engine surfaces certain SKUs over others, that ranking reflects choices made by engineers you have never met, optimizing for objectives that probably do not map perfectly to your brand's stated values.

Most commerce teams treat AI outputs as arithmetic. Input goes in, output comes out. But the encyclical's framing forces a more uncomfortable inference: your AI vendor is, in a calibrated sense, a silent co-author of every customer interaction your brand produces. That is a vendor lock-in problem with an ethical surface area most procurement conversations never reach.

The Benchmark: Where Most Brands Are vs. Where They Should Be

Roughly 70 percent of enterprise commerce teams, based on available operator surveys, have an AI usage policy. Fewer than 20 percent have documented what values their deployed models were optimized for. And best-in-class operators — the ones running formal model evals before deployment — are probably closer to 8 percent of that group. That gap is the actual benchmark problem. Not adoption rate. Not token cost. The gap between 'we use AI' and 'we know what our AI is optimizing for.'

The separation between average operators and top-10-percent operators in 2026 is not which LLM they chose. It is whether they built an internal eval framework before deploying customer-facing AI. Evals are unglamorous. They require you to define what good looks like before the model tells you what good is. Most brands skip this step because it is slow and the vendor demo looked fine. The brands that do not skip it are building a defensible moat.

Three Actions Separated by Rigor

Average brands will read the encyclical headline, forward it to legal, and move on. That is a mistake. The document's underlying logic maps onto three operational decisions that are worth taking seriously regardless of your theological priors.

First, audit your deployed models for value alignment. Not vibes. Actual documentation of what objective function your vendor used during fine-tuning, and whether that objective conflicts with your brand's positioning on things like inclusivity, sustainability, or pricing fairness. If your vendor cannot answer that question, treat it as a red flag roughly equivalent to a supplier refusing to share a factory audit.

Second, build a lightweight eval suite for your highest-traffic AI touchpoints. Customer service, product recommendations, search ranking. Define three to five failure modes you would not accept from a human employee, then test whether your deployed model produces them at scale. This does not require a research team. It requires two days and someone willing to write adversarial prompts.

Third, consider open-weight model options for use cases where value alignment is non-negotiable. Open-weight models are not automatically better. But they allow inspection. You can see the training decisions. Vendor lock-in and ethical opacity often travel together, and the encyclical's framing gives you a board-level argument for why inspection matters — one that is harder to dismiss than an engineer saying they prefer open-source.

Three Questions to Pressure-Test Your AI Governance Posture

Can you name the objective function your primary customer-facing AI was optimized for, in one sentence, without looking it up? If your highest-volume AI touchpoint produced a systematically biased output for six months, how many transactions would pass before your team detected it? And when your AI vendor updates their model — which they will, without asking — does your brand have a process to re-eval before the new version hits production, or does the update just silently propagate?

One honest uncertainty: the encyclical's commercial uptake is genuinely hard to predict. Papal documents do not typically move enterprise procurement cycles. What would change this view is evidence that institutional investors or major retail partners begin requiring AI value-alignment documentation as part of vendor due diligence. If that happens, operators who built the eval infrastructure early will have a durable advantage. If it does not, the cost of building that infrastructure was still probably worth it.