We didn't build an AI app. We built a prayer app that uses a little AI.
Most of what Tower of Babble does is ordinary software. playing audio, syncing lists, sending a notification when your leader posts. AI writes a first draft of a prayer, only when you ask.
What the app actually is
Strip out the buzzword and look at the parts. A library that stores your prayers. A player that speaks them in a warm voice. A way for a leader to share a recording with the people who listen for their guidance. Playlists. A QR code. Reminders. None of that is AI. It's the same computer science that runs a podcast app or a notes app.
We hear the concern
People worry about AI for good reasons: the energy it burns, the water it churns through, what it was trained on, whether it flattens human craft, whether a machine should be anywhere near something as personal as prayer. You're not wrong to ask. We asked the same things before we shipped a single generation.
So we kept it small. The AI never speaks for you and never speaks to God on your behalf. It hands you a draft. You decide whether it's truly yours.
How we use it — lightly, and on purpose
Only when you ask
No background generation, no auto-suggestions, no AI in your feed. You tap "create," or nothing happens.
You approve the request before a word is written
You see exactly who and what the prayer will be about, and confirm it. The question we want you asking is "would I really say that?"
You edit every word after
The draft lands in a text box, not a stone tablet. Change it, cut it, rewrite it, or throw it away.
A small model, not a frontier one
Prayers are generated with GPT-4o-mini — a lightweight model. We don't need the biggest engine to help you start a sentence.
Your prayers are never training data
Generations run through OpenAI's API, which is not used to train their models. Your words and the people in them stay yours.
We tell you when AI is involved
Every AI-generated prayer is labeled as one. Classic prayers and your leader's prayers are exactly what they say they are.
A little carbon, measured
We didn't want to hand-wave this, so we pulled every prayer ever generated in our development and production databases and averaged the tokens. Here's what one AI prayer really costs, next to something that is easily recognizable.
AI-generated prayers to match the carbon footprint of a single 10-mile drive to work.
That's a lot of good for a little carbon.
The commute figure uses the average car on U.S. roads — about 22 miles per gallon (U.S. EPA). A hybrid would roughly halve it; a gas SUV or pickup would raise it by a quarter.
How we got the number
Across the development and production databases, a successful generation averaged 481 tokens (295 in, 186 out) on GPT-4o-mini, a small model. We priced that at roughly 0.3 Wh and 0.2 g CO₂e per prayer — deliberately generous, several times higher than published small-model estimates, to keep the claim safe. The U.S. EPA's average car on American roads — roughly 22 miles per gallon — puts out about 400 g CO₂ per mile (one gallon of gasoline burns to 8,887 g CO₂), so a 10-mile commute is roughly 4,000 g. 4,000 ÷ 0.2 ≈ 20,000. Audio playback and text-to-speech are separate, and TTS is cached so a given prayer is only ever voiced once.
The line we hold
If any of this ever stops being true, this page changes along with the app.
A human is always in the loop
You review the request and edit the result. AI never publishes a prayer on its own.
AI stays a drafting tool
It won't creep into your voice, your data, or your leader's relationship with their flock.
No dark patterns dressed as intelligence
No engagement-maximizing suggestions, no synthetic urgency, no "the AI thinks you should pray more."
We'll keep measuring
The footprint figures on this page come from real usage data, and we'll continue to revisit this; both for our knowledge and your peace of mind.
Prayer first. Software second. AI last.
That's the order, and it's not changing.