We tested three AI tools on bulletin announcements. One was safe.

Every parish bulletin starts the same way: a pile. Ours was a realistic one — eleven announcement requests collected the way parishes actually collect them. Two formal emails from ministry leaders. A forwarded text from the Knights about a pancake breakfast. A handwritten note about a funeral luncheon, photographed. A request from the school with three different dates in it, one of them wrong. A long, warm, completely unstructured paragraph from the prayer shawl ministry. The usual.

We gave the same pile to three kinds of AI tools and asked each for a clean, bulletin-ready announcements section. We are deliberately describing categories rather than naming products — the brands will change by Christmas; the categories will not.

Tool one: a general-purpose chatbot, used raw. The kind anyone can open in a browser for free. The prose was, frankly, lovely. It smoothed the prayer shawl paragraph into two graceful sentences and gave the pancake breakfast more dignity than any pancake breakfast has ever had. It also did something disqualifying: where the school’s request was ambiguous, it quietly resolved the ambiguity by inventing a time. The result read perfectly and was wrong. A reader cannot tell a confident error from a fact, and a bulletin error about a funeral luncheon is not a small thing in parish life. Used raw, with no checking step, this category fails the test — not because the writing is bad, but because the workflow has no place where a human is forced to catch the invention.

The result read perfectly and was wrong. In a bulletin, that is the most dangerous sentence there is.

Tool two: the AI feature inside a big office suite. Safer by temperament — it stayed closer to the source text and invented nothing we could find. But it also did the least work. It tidied each announcement individually and left the actual job — one consistent section, in the parish’s voice, with dates verified against each other — undone. It also kept nudging the tone toward corporate newsletter: “We’re excited to announce.” No parish has ever been excited to announce in those words. Harmless, mildly useful, not a solution.

Tool three: the same AI underneath, but wrapped in a workflow. A structured setup — the kind any parish can build with help, no programming required — where the AI receives the parish’s style examples, is instructed to flag anything uncertain instead of resolving it, and produces a draft that lands in front of the secretary as a checklist: each announcement, its source, and a highlighted question wherever the inputs disagreed. It caught the school’s date conflict and asked, rather than guessed. The secretary approved nine items, fixed two, and the section was done in eleven minutes.

Same intelligence, radically different safety. The lesson of the test is that “is this AI tool safe for parish use?” is mostly the wrong question. The right question is whether the workflow around the tool forces uncertainty into the open and puts a person at the gate. Tool three was safe because doubt was made visible and a human approved every line — which is exactly the standard Magnifica Humanitas sets when it insists that human responsibility remain real, not ceremonial.

The practical takeaway for your office: do not paste the announcement pile into a raw chatbot and print what comes back. Either add the checking step yourself — ask the tool to list every date, time, and name with its source before you accept a draft — or use a setup that builds the checking step in. The pile becomes eleven minutes either way. The difference is whether Wednesday’s bulletin contains a Mass time nobody scheduled.