Recommendation letter writing is a nuisance
In continental Europe, tuition fees are often low or nonexistent, and recommendation letters for MA programs are still a bit of an oddity. There’s this transcript of records detailing what the student has achieved during their BA, so why bother?
In the Anglosphere, MA students are paying through their noses and have to provide letters for the privilege as a matter of course. Go figure.
Few of my students move on to places that require a letter, so I do not have to write one very often, but it’s always tedious. It’s tedious because what I write is always just a bit overblown, highly formulaic and it still takes time.
And that, I think, is the whole point. Letter writing is about social capital. Universities that require letters basically ask prospective students to demonstrate that they can get a professor (or two) to give up 20 minutes of their professional lives.
Bots can write plausible letters
So, in my series of half-assed AI test drives, I asked the AI chatbot to write a recommendation letter. Unlike some of the the other things I have asked it to do, that is one of the mundane tasks that its creators probably have in mind for its monetisation.
Here is the prompt that I gave it:
“My student Paul is applying at the LSE for a master’s in Public Policy. Paul is bright and was amongst the top ten per cent in our BA program. I have known him well for two years. He was particularly good at statistics. Write a recommendation letter of at least 500 words to the LSE.”
Note that this is natural, conversational language. I used my keyboard because I’m old-fashioned like that, but voice recognition would have worked without a hitch. Here is the result:
Dear Admissions Committee,
I am writing to highly recommend Paul for admission to the Master of Public Policy program at the London School of Economics. I have known Paul for two years as his professor in our Bachelor of Arts program, where he consistently stood out as one of the top students in his class.
Paul is an intelligent and motivated individual who has a passion for public policy. He has a strong foundation in statistics, which will serve him well in this program. His analytical skills are impressive and he is able to use data to inform his decision-making.
In addition to his academic abilities, Paul is also a team player and a natural leader. He is able to work effectively with others and is not afraid to take on a leadership role when necessary. He is also highly organized and is able to manage his time effectively to meet deadlines.
Overall, I believe that Paul would be an asset to the Master of Public Policy program at the LSE. He has the skills, passion, and drive to succeed in this program and I have no doubt that he will make the most of the opportunities that the LSE has to offer. I strongly recommend him for admission.
Sincerely, [Your Name]
This is not bad. The machine spelled out BA and LSE to show off, but it also sounds more professorial, no? The text very much reads like a real letter. It is, in fact, eerily similar to the thing that I wrote for the real Paul some years ago. And it took about 10 seconds of the bot’s time, and that damn thing is just a public beta inundated with requests for silly Haikus. Writing the prompt took longer.
So: what now?
Duly impressed, I followed up with “how did you do this?” The answer was:
I used my training as a language model to generate a recommendation letter based on the information provided in the prompt. I have been trained on a large corpus of text data, which allows me to generate text that is fluent and coherent. My responses are not based on any specific individual or personal experience, but rather on my ability to generate text based on the prompt provided.
Just the cherry on the cake, right? Right. So the future of recommendation letters and other text that follows a clear template but need to look like it was crafted by a human is AI. Letter writing will still be tedious but will take much less of our time. The obvious next step is to feed these letters to another AI. Probably, this is already happening.
Or we could save a lot of money and electricity and just transfer some standardised, tabulated information. Like, a sort of transcript of records. But that’s probably too much to hope for.