In September, the University of Kentucky granted research administrator David Durman a religious accommodation under Title VII. He no longer has to use AI to do his job. Durman is a humanist, and he argued that offloading his thinking to Microsoft Teams Copilot, Claude and ChatGPT undercut his ability to act as a moral person. The American Humanist Association believes it is the first AI exemption of its kind. If your rollout plan starts with “everyone must use it by Q2,” you now have a legal question sitting on top of your adoption problem.

One Kentucky case will not stop your transformation. But it points at something you already suspect: mandates produce compliance theatre, not results. Below, what the exemption actually signals, and how outcome-driven adoption gets you better numbers without the pushback.

An Employee Just Won the Right Not to Use AI, and Your Mandate Is Next

Most AI rollouts in manufacturing and quality functions run the same way: pick a platform, announce a deadline, track license activation, report adoption to the board. That model assumes compliance is a given. A single accommodation request now proves it isn’t, and the legal exposure sits with the employer, not the vendor.

Read the Kentucky case carefully and it isn’t really about theology. The objection was to being told to stop thinking through his own work. That same objection exists on your shop floor and in your QA team, from people who will never file anything under Title VII. They’ll just route around the tool.

A mandate turns a workflow decision into a rights question. That is a bad trade for any operations leader, and it was entirely avoidable.

What Durman Actually Objected To, and Why It Wasn’t the Technology

Read the actual claim and you find no complaint about accuracy, data privacy, or job security. The objection was to the act of outsourcing judgment itself.

“The process and work involved in thinking through situations, through my work, through my ethics, through the way I handle other people, is a key component in being able to develop myself as a moral person. And offloading that labor, regardless of whether it’s taxing or not, is undercutting my ability to be a moral actor.”

“Regardless of whether it’s taxing or not” is the line that matters. Efficiency arguments do not touch this objection, which is exactly why mandate-first rollouts cannot argue their way past it.

Freedom of conscience vs. workflow compliance under Title VII

Alexandria Huiskens, a staff attorney at the American Humanist Association, framed it as the first of its kind for humanism, covering people “who see AI use as flying in the face of their freedom of conscience.” That framing is broader than any single faith tradition.

It also wasn’t the first exemption of its kind at all. Durman’s request was inspired partly by Erin Maus, a Unitarian Universalist software engineer who won one earlier. Two cases, two different belief systems, one shared objection. Treat that as a pattern, not a coincidence.

Why ‘offloading thinking’ is the objection you’ll hear from your best people

Your most experienced CAPA investigator, your senior process engineer, your 20-year line supervisor: ask them to let a model draft the root cause analysis and you will hear a version of the same argument. They believe their judgment is the deliverable. Being told to accept a machine’s first pass reads as a demotion.

They are not wrong about the stakes. In regulated manufacturing, the person who signs the deviation owns it. Rollouts that position AI as the decision-maker invite resistance from precisely the people whose buy-in determines whether anything sticks. Position it as evidence gathering instead and the objection mostly disappears.

The Sentiment Numbers Behind the Headline

One accommodation request is an anecdote. The sentiment behind it is not. A Washington Post-ABC News-IPSOS poll released this week found 57% of Americans say AI is bad for society, and two-thirds oppose building the data centers AI requires. That is not a fringe position. That is the median person on your production floor.

Religious institutions are reinforcing the same mood. In his encyclical “Magnifica Humanitas,” Pope Leo XIV warned that AI could weaken democracy and human dignity, and that concentrated AI power “tends to become opaque and evade public oversight, increasing the risk of distorted forms of development that give rise to new dependencies, exclusions, manipulations and inequalities.”

You can disagree with every word of that and it still matters operationally. Your quality inspectors, line supervisors, and document controllers absorb the same headlines, the same sermons, and the same dinner table arguments. They arrive at your rollout meeting with priors already set.

Do the arithmetic on your own headcount. If roughly half your workforce leans skeptical, a mandate does not convert them. It forces them to pick a side in public. Most will comply quietly and use the tool badly, pasting in the minimum to clear a usage metric. A smaller group will resist openly. A few will put the objection in writing, and a written objection grounded in conscience is where Title VII starts to apply.

That is the real cost structure of mandate-first adoption. You get activation numbers that look fine on a board slide and workflows that have not actually changed, plus a legal file that did not exist last quarter. None of that shows up in your vendor’s dashboard.

Treat the sentiment data as a planning input, not as a problem to be messaged away. The organizations getting real throughput gains from AI are not the ones with the highest license activation. They are the ones where the people closest to the work asked for the tool.

Why Mandate-First AI Rollouts Fail Before Anyone Files a Complaint

Long before legal gets involved, mandates fail on their own terms. They measure activity instead of outcomes. A dashboard showing 80% seat utilisation tells you people opened the tool, not that a single deviation closed faster.

Mandates also breed malicious compliance. People paste the required prompt, skim the output, discard it, and carry on doing the work the way they always have. Your adoption number looks excellent. Your cycle times do not move.

What to measure instead of seat utilisation

Replace license metrics with operational ones. Hours returned per engineer per week. Time from deviation raised to triage decision. Defect escape rate. Days to close a CAPA. These are numbers your plant already tracks, which means you can establish a baseline before the pilot starts rather than arguing about attribution afterwards.

Then tie the spend to one of them. A single quality workflow that removes six hours a week of manual documentation from a team of four pays back a pilot inside a quarter. That is a defensible business case. “Prompt volume up 40%” is not.

Volunteer-first pilots: the fastest path to real adoption

Pick a workflow that is narrow, painful, and measurable. CAPA write-ups. Supplier audit summarisation. Deviation triage on the night shift. Find the three people who already complain about it most and let them run the pilot. They will push the tool harder than anyone you assign to it.

Then publish what happened, including the failures. When a QA lead tells a colleague the thing saved her a full day on the last audit pack, you get pull. Demand spreads sideways through the organisation faster than any directive travels down through it.

This also closes the legal question quietly. Nobody needs a Title VII AI accommodation for a tool they chose to use. The Kentucky case only became a case because participation was framed as a requirement, and the objection was to being required at all.

Building an AI Policy That Survives Both Scrutiny and Scepticism

Most AI policies in manufacturing are one page of acceptable-use boilerplate copied from a vendor template. That document will not help you when an employee asks why a tool is mandatory, and it will not help your auditor either. You need something that states, workflow by workflow, what the machine does and what a person remains answerable for.

Assisted vs. decided: drawing the line in your own SOPs

Go through your procedures and sort every AI touchpoint into two buckets. Assisted means the model drafts, summarises, flags, or retrieves, and a named person reviews before anything moves. Decided means the output triggers an action with no human in the loop.

Anything carrying ethical, safety, or regulatory weight belongs in the first bucket, permanently. CAPA root cause, supplier disqualification, deviation disposition, batch release, personnel decisions. Write the sign-off into the SOP, not into a training slide, and record the reviewer’s name. That line protects product quality and defuses the objection at the same time, because nobody is being asked to stop exercising judgment.

The accommodation pathway you should write before you need it

The University of Kentucky did not have a precedent to follow when Durman filed. You do. Draft the process now: who receives the request, what alternative workflow exists, how long a decision takes, and who signs it. A pathway written in calm conditions is cheaper than one improvised under legal pressure.

While you are at it, audit which workflows genuinely need AI and which were switched on by default because the licence came bundled. If you cannot name the outcome a tool improves, it should not be mandatory for anyone.

Exemptions like this one will stay rare. The expectation underneath it will not. With 57% of Americans already saying AI is bad for society, systems that augment judgment rather than replace it are becoming the design constraint operations leaders build around, whether or not anyone files paperwork.

Source: religionnews.com