As a non-native speaker, expressing yourself is even harder than it is in your mother tongue. One of the (many) moments I was reminded of this was when someone pointed out a mistake in my LinkedIn bio. I had called myself an “Out-of-the-box thinker,” which was an attempt to hint at my ADHD using LinkedIn speak. Most likely, it should have been outside-the-box thinker. (I’ve since updated my bio, in case you were wondering.)
But this post is about AI, not my LinkedIn bio, so why is this relevant, you may ask? Well, recently I asked Google’s Gemini to help write a proposal for my company to be more inclusive to neurodiverse people, developers in particular, to help them reach their true potential. I was surprised to see it use out-of-the-box instead of outside-the-box too.
So I thought, let’s check if it could explain the difference:
Question: Could you explain the difference to me between outside-the-box and out-of-the-box?
Answer: While both idioms sound almost identical, they describe completely opposite concepts: one is about creative thinking, and the other is about practical readiness.
Thinking Outside the Box
- Meaning: Thinking creatively, unconventionally, or from a new perspective.
- Origin: Inspired by the famous 9-dot puzzle, which requires drawing lines beyond the invisible boundary of a square to connect all the dots.
- Usage: “To solve this budget deficit, we need to think outside the box.”
Out of the Box
- Meaning: Ready to use immediately upon arrival or installation, requiring no additional assembly, configuration, or modification.
- Origin: Derived from retail products (like software or electronics) that work straight out of their original packaging.
- Usage: “This software works right out of the box with zero setup required.”
So, when asked explicitly, it can perfectly explain the differences between the two, but it fails to apply this insight in practice when generating text. What is going on here?
Sometimes we seem to forget that LLMs are nothing more (nor less) than text predictors: given a chain of text tokens, what is the most likely continuation? Due to their enormous models and training sets, they can do magnificent things. When it performed the first task, it reused tons of corporate lingo to come up with an answer. Because I am part of a majority of non-native speakers who mix up the expression, the LLM mirrored that exact mistake. In the second case, however, it apparently reused insights derived from an English dictionary. So, among the long list of limitations LLMs have, we have to add a new one: they cannot think (or answer) outside the box.
Note: Written by me, spellchecked by an LLM