Handcrafted Letters
When AIs do the writing and reading
After Google Translate launched back in 2006, I remember occasionally playing a “game” where we’d take some English text, translate it into another language, then take the output and translate that into yet another language, bouncing through French, German, Mandarin, Afrikaans and the like, before finally returning to home to English. Usually this produced hilarious results via a sort of digital game of telephone.
Translation quality has improved significantly since then, and the results long since stopped being very funny, but this memory resurfaced recently in the aftermath of the blitzkrieg launch of AI products bombarding us on every front, each promising to make us more productive than ever before. Tools like Superhuman AI or Torq let users type a few key points, and then use AI to expand it into a friendly, human-sounding email, while also enabling recipients of any long-winded emails to (one-click, bam!) summarize things back down into bullet points.
Here is a scenario: I am emailing you to update you on some <insert workplace topic here>, but in the interest of politeness, I can’t just send you an email with a bulleted summary and nothing else. As much as it would make my life easier, it risks offending you — who do I think I am, your boss? — and so I run it through an AI which converts it into a socially acceptable email. On your end, you are also busy and can’t be bothered with reading my wordy email, so you run the email through an AI that summarizes it back down to key items.
There is something funny going on here! Bullet points are the input into this chain, and bullet points are the ultimate output as well, but in between we’ve introduced two AI transformation steps for no reason other than the appearance of politeness. This seems like a silly cost to pay; invoking an AI costs something, and with every step some risk of errors creeps in, just like the repeated Google Translate loop. AI hallucinations remain unsolved, and we lack any sort of error correction codes for ensuring messages avoid corruption as they travel through the AI pipes connecting us.
This isn’t intended to be some luddite anti-AI “death of human interaction” doompost. Much of the creation of basic business artifacts (email, docs, presentations) stands to benefit from AI-assisted productivity tools. How important is it that one learns the fine art of pixel wrestling in Google Slides or PowerPoint, for example? Why not just use an AI presentation generator? Even if I wanted to take up arms alongside the anti-AI crowd, the genie is out of the bottle, and there’s no going back. Google already has a “Help me write” button and extension in testing, Gmail already has auto-write features, and we’re collectively barreling towards this future regardless of what you feel about it.
Still, I can't shake the feeling that this is one of the many instances where we find ourselves on the brink of irreversible change. Not since the arrival of email have we encountered something with the potential to change how we communicate as much as AI. It may start with a few services offering to write and summarize your emails, but it’s not hard to imagine a world where the vast majority of our emails and documents are handled primarily by AI. Sitting down to “hand-type” an email may seem as quaint as writing a letter does today.
There is ongoing discussion on watermarking AI-authored text, as a way of combating AI-written homework, essays, and other documents where we (for whatever reason) prefer a human writer. Maybe email will be another. Perhaps when you open an email in the future, it might contain some verifiable proof that a human wrote it. And in this future, perhaps it will again be the thought that counts; you will smile wistfully, grateful that someone bothered to spend their precious time laboriously crafting some sentences, and think back to a simpler time when we all typed the letters out ourselves.
If all else fails, there’s always this:
Thanks to Nundu Janakiram for inspiring this post.