“You look like a thing and I love you” was one of many results when AI researcher Janelle Shane gave a neural network a list of pickup lines from the Internet and asked it to generate its own. From this, you might think that artificial intelligence won’t be writing fundraising copy soon. You’d be wrong.
At first, Shane said, her network spat out nonsense like “De’ver dincing you bangting Won pantt ore.” (Note: this is what Chauncer always looked like to me…) But with more time and raw data it came up with lines that ranged from still nonsensical:
– Hey baby, you’re to be a key? Because I can bear your toot?
– I have a cenver? Because I just stowe must your worms.,
– Hey baby, I’m swirked to gave ever to say it for drive.
To those that start well but fade:
– Are you a candle? Because you’re so hot of the looks with you.
– If I had a rose for every time I thought of you, I have a price tighting.
– Are you a camera? Because I want to see the most beautiful than you.
To plausible, especially around closing time:
– You are so beautiful that you know what I mean.
– I want to get my heart with you.
– You are so beautiful that you make me feel better to see you.
– I want to see you to my heart.
In her book appropriately named You Look Like a Thing and I Love You, Shane talks about the four signs of AI doom:
– The problem is too hard
– The problem is not what we thought it was.
– There are sneaky shortcuts
– The AI tried to learn from flawed data
This is a clear case of flawed data. Going to the Internet to find examples of successful pick-up lines is like trying to strike oil by building an offshore drilling platform in your bathtub. Specifically, there’s no record of whether or not these lines were successful in accomplishing whatever the speaker desired to have happen. Thus, the “training” that’s taking place based on volume, rather than success for failure.
What happens if you do have a record of success or failure? John Seabrook tested this in “The Next Word” in which he used Google’s Smart Compose feature to help him write a New Yorker article. Because Google can see whether its suggestions are taken or not, it can compile a win-loss record for its suggestions, using those data to then suggest more accurately in the future.
And now you can see the first AI-written fundraising letter from late 2018:
Some of this has the linguist challenges of our machine learner who loves you because you look like a thing. But it’s already developed a sense for legitimizing small gifts (“Very little helps”), urgency (“My request today comes with a note of urgency”), and identifying with your audience (“I am writing to you because you either studied at Manchester or you wanted to”).
We know what the future looks like because this has already been mapped out in other disciplines. At first, for example, humans beat AIs at chess. Then, AIs beat humans at chess, but the best players were a team of AI and human to help sand off the neural network’s rough edges. Now, AIs don’t need humans — at chess tournaments, they can tell you are cheating with an AI if you make a move that is too good or too original. Original is now the realm of the AI, not the human.
So too is it here. Our fundraising copy is better, for now. But there are already systems that can help suggest emails for fundraisers after learning their style to help automate some of the more rote tasks. You also see human copy improved by algorithms like the Hemingway app or the scoring system discussed on yesterday’s Agitator blog. But with access to a strong corpus of letters and (and this is what has been lacking in the past) results, you will see AIs being able to do mass writing as well as better customization to an individual than a human writer. We, then, will turn into the guides, setting the goals and parameters in which AIs do selection and communication.