worship.ai – What an Artificial Intelligence Lyric Generator Teaches Us about Image Bearing

A few years ago, I wrote a post about Writing Worship Music with AI in which I experimented with a few Recurrent Neural Network libraries. When I’ve given talks on AI, I often referred back to that post, but I always wished the tool was online so I could share the joys of AI-generated lyrics with the internet. So, now I’d like to introduce you to a little site called:

The tool allows you to select from three training libraries: modern worship music, classic hymns, or the Psalms (KJV). Each one has a few hundred kilobytes of lyrics, and your browser is powerful enough to generate a Markov chain of the lyrics, and then then generate combinations of words that mimic the source. A Markov chain analyzes source data (in this case words) and tracks which one are most likely to come after another word. For example, after the word “praise” you might find “the” or “you” or “him.”

The AI doesn’t know what the words mean or even the structure of a song, so I added a little bit of spice to make sure it has a chorus or hook and some repeating lines to make it feel more like a real song.

Sample Lyrics

What you’re going to love like you
I surrender, I surrender all to You,
Here I am will serve the Lord on high
Burden me with each other

And all I need anymore
With all our tears away
With all our tears away

In all I found my life I was hopeless
I’m amazed, so amazed that I know
There’s just no time left to tell him
Surrendered all to You only

Modern Worship Song from worship.ai

The Prince of glory divine
Blend all my soul in the highest!
There let the people rejoice!
It was there by Thy good pleasure
“Jesus died my soul in the winter,
The king of love and sorrow meet,

Crown Him the glory of the earth,
But live with Thee I will be
Crown Him the glory of the earth,
But live with Thee I will be

Classic Hymn from worship.ai

Answer me, O LORD, I call to you;
like chaff that the King of glory?
How precious is their necklace;
“Where is their help and my shield;
let them ever sing praises to you;

Vindicate me, O LORD, God of Israel.
like an owl of the wicked come to me,
The sorrows of the wild beasts;
His glory is great above the heavens!
O LORD, my heart all the earth;

Psalm from worship.ai

How AI Reflects our Values

This experiment (worship.ai and the previous project) was inspired by seeing other projects that generate Harry Potter novels, biblical quotes, and Mozart music. When I saw those, they made me want me to learn how to use the technology so I could think more carefully about how our AIs are often a reflection of how we train them.

In the case of the “songs” above, you can see that even though this AI is not smart enough to understand meaning or syntax or rhyme, the vocabulary alone is markedly different between classic hymns, modern songs, and Psalms. One could make this a criticism of modern song lyrics, but I think there’s something deeper here in that it offers us a clear example of how our tools reflect our values and desires.

A shovel, for example, embodies our desire to dig and, at a deeper level, our longings to build, create, and flourish. A pencil enables us to write and draw, transporting the ideas that have taken residence in our minds onto a format we share with others. Likewise, a gun values the ability to destroy things at long range, whether that be our next meal, an enemy, or a friend.

Those tools are, of course, much more straightforward in meaning than an AI. They have a singular, clear purpose and their value-ladenness is fairly transparent.

But modern devices, AIs in particular, often embody more of the complexity, ambiguity, and ostensible irrationality of humanity. When it comes to AI, we wonder things like why do AIs sometimes come to racist or sexist conclusions? Why does it write a line like this: “Vindicate me, O LORD, God of Israel.” but then follow it with “like an owl of the wicked come to me”? Why does it come up with a move in Chess or Go that no human has ever tried?

In each of these cases, no matter how deep the learning is in the AI, the answer is that the computer is simply following the model and the data we have given it. We might be surprised by the conclusion because the system has grown more complex than we can see at once, but it is still following the rules and value system we’ve given it. There might be a better conclusion we want it to come to, but it will follow the path we lay out for it. Sometimes this surprises and delights and sometimes it is frustrating or hurtful.

While there is some good thinking and theologizing being done about whether an AI will ever be able to truly think or have a mind, for now, our AIs are largely a reflection of the values we feed them.

Value-Bearing and Image-Bearing

This leads us to ask similar questions about our own actions. Why do we eat foods we know are bad for us? Why do we honk at crazy drivers when we know it usually makes things worse? Why do we worry about tomorrow?

A materialist view of humanity would say that this is all caused by our programming. Our brains are nothing more than biological machines, and like an AI, we have a series of inputs (the rational part of our brains, the reactive, lizard part, and so on) that go through a series of neural pathways, all of which work together to cause us to make that final, inevitable action. For us, choice is as much an illusion as it is for the machine or the AI.

But a Christian view of humanity would say there is good deal more to us than this.

We affirm that we are physically embodied creatures with complex brains, and the habits we etch in our brains can come to define us. But we also affirm that there is an immaterial part of us, what the scriptures call the soul, the spirit, or the heart, and that we can make willful decisions. Sometimes we follow the simple inputs, desires, and values of our material bodies, leading us to eat, drink, sleep, and other embodied activities. But other times, we choose to override the messages from our body and emotions. This is what allows us train for a marathon when our body wants to stop, to deescalate conflict when our adrenaline says to punch, to sabbath when our world says we must produce. And it’s also what allows us to write and to feel lyrics like “As a deer pants for the water, so my soul thirsts for you”, “Amazing grace, how sweet the sound”, or “If the stars were made to worship, so will I.”

As Christians we also say that humans are made in the image of God, and perhaps one way of thinking about this is that we fulfill our role as “image bearers” when we make choices—sometimes against our own bodily desires, our broken programming, and our faulty inputs—that embody the values of our creator.

Image if an AI could say, everything in my programming and data model says the next word should be “*%$@”, but I’m going to choose to speak the words I think my creator truly wants.

By the power the spirit, may we do this and more.

One more note: if you want to think more about the implications of Artificial Intelligence through the lens of theology, check out Jason Thacker‘s new book The Age of AI.

5 thoughts on “worship.ai – What an Artificial Intelligence Lyric Generator Teaches Us about Image Bearing

  1. Just being picky but… you seem to have missed out some words in the stuff above.

    “When I give talks on AI, I always the tool was online…”

    “If the were made to worship, so will I.”

    The Markov chain/King James Bible quotes thing is many years old now and AI ‘creativity’ has moved forward somewhat. I think if you could build in some recognition of number of syllables in words you might get a much better set of results.

    I await your next move. 🙂

    1. Thanks Phil! Yes, I looked at adding syllabication and rhyming (this was the best API I could find: https://rhymebrain.com/api.html), but I didn’t see a good library that did it at an AI-learning level, rather than just as traditional programming, or one that could run in the browser. Time to keep experimenting!

Leave a Reply

Your email address will not be published. Required fields are marked *