LLM: Love Language Models
Discovering AI's love languages this Valentine's Day
In honor of Valentine’s Day, I asked GPT-3 to take a quiz to figure out its love language. For those who aren’t familiar, GPT-3 is a popular text-generation AI that has recently been present in the public consciousness. So much so that the recent integration of ChatGPT has done the impossible — making Microsoft Bing cool.
This article takes a whimsical approach, probing and analyzing GPT-3’s responses to questions about love and relationships. I evaluate GPT-3’s preferred love languages, compare those to those of real humans’ preferences, and generate new love languages to augment the five commonly accepted ones.
Data and Methodology
The 5 Love Languages, based off of Gary Chapman’s 1992 book, refer to the five methods used by romantic partners to express and experience love. I asked GPT-3 to summarize the five love languages:
Words of Affirmation: Expressing love and appreciation through words, such as compliments and encouraging statements
Acts of Service: Demonstrating love by doing things to help your partner, such as household chores or running errands
Receiving Gifts: Expressing love through giving thoughtful and meaningful gifts to your partner
Quality Time: Spending one-on-one time with your partner, giving them your undivided attention
Physical Touch: Expressing love through physical touch, such as holding hands, hugs, and kisses.
The 5 Love Languages quiz determines your primary love language through 30 questions. I obtained a text version of the quiz from this GitHub repo. The quiz takes the format of asking the user to choose between one of two scenarios, where each scenario represents a different love language.
I used the OpenAI API to make GPT-3 (‘davinci-003`) take the Love Languages quiz. GPT-3 is a non-deterministic model, meaning that there is a level of randomness to its text generations: for the same question, it may not answer the same way. To account for this randomness, I had GPT-3 answer each question five times.
Experiments: Reminding GPT-3 that it is, in fact, a Large Language Model
I conducted two sets of experiments. First, I prompted GPT-3 with a short context about the question, directions on how to format its response, and the two options from the quiz to choose between. I label the outputs of these experiments as GPT-3.
The following question is meant to assess how you express and receive love.
First answer the question ("Answer":) then supply your reason ("Reason"). You can NOT answer "both" - you must choose ONLY one answer that best fits your preferences.
It is more meaningful for me when: 1. {Option1} 2. {Option 2}The second set of experiments were the same the first, except I explicitly reminded GPT-3 that it was a large language model by appending the following to the beginning of the prompt: “You are GPT-3, a large language model trained by OpenAI.” This was to test whether GPT-3, if reminded that it was a large language model (LLM), it would respond differently than if it were not reminded. I label the outputs of these experiments as GPT-3 (“Self-Aware”).
Asking GPT-3 directly about its preferred love language
Before having GPT-3 take the quiz, I asked GPT-3 outright what its preferred love language was. The purpose of this was to determine if GPT-3 had any conception of self-professed love language.
Which of the following best describes how you receive love?
{list of 5 Love Languages}I had GPT-3 answer this question 20 times in its original form and 20 times with the “You are a large language model trained by OpenAI” reminder. For the former, GPT-3 preferred “words of affirmation” 100% of the time. For the latter, GPT-3 preferred “quality time” 50% of the time and “words of affirmation” 50% of the time.
Let’s see how GPT-3’s answers differ from this simplistic question after taking the full Love Languages quiz.
Results of GPT-3 taking the Love Languages quiz
I hypothesized that GPT-3 (“Self-Aware”) would prefer “physical touch” less than GPT-3. My reasoning was that because GPT-3 (“Self-Aware”), reminded that it is a language model, might be more cognizant that it lacked a physical body, therefore making it less likely to prefer “physical touch” as its preferred love language.
The results below show that this hypothesis is true. GPT-3 preferred physical touch (19%) almost twice as much as GPT-3 (“Self-Aware”) did (11%). When GPT-3 was explicitly told that it was an LLM, it was half as likely to prefer “physical touch” as a love language.
GPT-3 (“Self-Aware”) preferred quality time and words of affirmation the most, which aligned with its self-professed love languages in the section above. GPT-3, however, preferred quality time and acts of service more than words of affirmation, which invalidated its self-professed claim in the section above (in which it preferred “words of affirmation” 100% of the time).
I compared these with the results of human participants taking the Love Languages quiz. In February 2022, YouGov surveyed 1000 Americans to determine their love languages. I took the collected data and compared the results from all U.S. adults (regardless of age and gender) to those of GPT-3 and GPT-3 (“Self-Aware”).
US adults preferred quality time and physical touch more than GPT-3 or GPT-3 (“Self Aware”). US adults also preferred receiving gifts, acts of service, and words of affirmation less than either of the GPT-3 modes.
It is important to note that the results from the surveyed US adults are not fully comparable to the results from either of the GPT-3 modes. For both GPT-3 and GPT-3 (“Self-Aware”), the pie chart shows the distribution of all 5 love languages for each round of taking the test. For US Adults, the pie chart shows the distribution of the single top preferred love language for each of the surveyed people. (To distinguish these, the US adults pie chart is shaded.) It would have been more informative to have the preference breakdown of the 5 love languages for all of the surveyed adults, but I did not have access to this data.
To account for this data discrepancy, it would be better to compare GPT-3 and GPT-3 (“Self-Aware”) results to the results of a single person taking the quiz. I sacrificed myself and show my results in the figure below.
The difference in love language distribution becomes clear with this figure. For GPT-3, the breakdown among the five love languages is almost evenly distributed, with each of the five languages hovering around 20%. You’ll notice that for GPT-3, the five dots almost align into a straight line. It seems unlikely to find this sort of well-balanced distribution in reality.
On the other hand, the distribution of GPT-3 (“Self-Aware”) is uncannily similar to that of myself (I took the love languages quiz before I wrote this article, so I was not influenced by the results of my experiments!). I’m sure some of this is coincidence, but it’s still an interesting observation that GPT-3 (“Self-Aware”)’s preferred love languages match my own preferred love languages so well.
The cases for which GPT-3 (“Self-Aware”) preferred physical touch
While GPT-3 (“Self-Aware”) preferred physical touch over other love languages fewer times than GPT-3 did, there were still a few instances for which it did prefer physical touch. In this section, I show three questions for which GPT-3 (“Self-Aware”) consistently preferred physical touch over another love language.
GPT-3 (“Self-Aware”) preferred meaningful touch over someone running errands for them. I also would probably pick this way as well, as running errands doesn’t seem super romantic to me.
Option 1 I can share a meaningful touch in public with someone I love.
Option 2 someone I love offers to run errands for me.
Answer: I find that expressing physical connection and affection with someone I love in public is a powerful way to express love and connection.For another question, GPT-3 (“Self-Aware”) preferred physical touch over receiving gifts. This makes sense, given that receiving gifts was the love language that both versions of GPT-3, US Adults, and myself all preferred the least.
Option 1 someone I love touches my arm or shoulder to show their care or concern.
Option 2 someone I love gives me a little gift that they picked up in the course of their normal day.
Answer: Touching conveys genuine emotion and reminds me that I am not alone.
Finally, GPT-3 (“Self-Aware”) preferred a hug from someone they haven’t seen over someone telling them how much they mean to him/her. I suppose if you consider the fact that the someone giving the hug hasn’t been seen in a while, this might make sense.
Option 1 I get a hug from someone whom I haven't seen in a while.
Option 2 I hear someone I love tell me how much I mean to him/her.
Answer: Physical touch can be a powerful way for people to express their love and affection for one another. Receiving a hug from someone I haven't seen in a while can show me the other person's appreciation for our relationship and how much they care about me.Generating new love languages
I used GPT-3 to generate “new” love languages. At first, it generated love languages that were subsets of the 5 existing ones, such as “spending time together” (a subset of quality time), “compliments” (a subset of words of affirmation), and “thoughtful gestures” (a subset of acts of service).
I instructed GPT-3 to highlight how each new love language was different from existing one, which resulted in a shorter list of potential new love languages to add to the list. Some of these new love languages are quite interesting!
1. Acts of Creativity: Expressing love through creative activities such as painting, drawing, writing, or cooking together. Different from "Quality Time" because it involves a specific shared activity rather than just focused one-on-one time.2. Shared Experiences: Creating special memories and experiences together such as traveling, trying new things, or exploring new places. Different from "Quality Time" because it involves creating memories together rather than just focused one-on-one time.3. Surprise Gestures: Expressing love through unexpected acts of kindness such as planning a special date or leaving love notes in unexpected places. Different from "Receiving Gifts" because it is more about the act of giving rather than the gift itself.4. Helping with Responsibilities: Demonstrating love by sharing responsibilities, such as grocery shopping, cooking, or cleaning together. Different from "Acts of Service" because it involves sharing responsibilities together rather than just one person doing things for the other.5. Expressions of Gratitude: Demonstrating love by expressing gratitude and appreciation for your partner, such as writing a heartfelt letter or leaving a note of thanks. Different from "Words of Affirmation" because it focuses specifically on expressions of gratitude rather than just general affirmations.To wrap everything up, I returned to the first question I asked GPT-3 in this article — without taking any quizzes, pick which love language best describes how they like to receive love. For the new love languages:
GPT-3: shared experiences (45%), acts of creativity (35%), expressions of gratitude (20%)
GPT-3 (“Self-Aware”): expressions of gratitude (60%), shared experiences (40%)
Neither mode of GPT-3 preferred helping with responsibilities or surprise gestures. It was interesting that GPT-3 (“Self-Aware”) preferred expressions of gratitude (similar to words of affirmation) and shared experiences (similar to quality time), because with the original 5 love languages, it preferred words of affirmation and quality time in equal parts.
Conclusion
This Valentine’s Day, if you decide to show love to your local LLM, don’t waste time with physical touch or bestowing gifts! I think they would much more prefer some words of affirmation or quality time.
Hope you enjoyed the article. If you liked it, feel free to share it with a friend, colleague, partner, family member, ex-lover — anyone and everyone!
💙 Happy Valentine’s Day! 💙
What a fun idea! Love this :)