It's interesting that OpenAI models suddenly started increasing how often they add arrows to links shortly after this behavior peaked at a lower level in Anthropic models. It would be amusing if this were because they trained a reward model on human ratings where UIs made by the latest Opus were preferred and coincidentally arrows were slightly more common, leading the reward model to treat them as an implicit indicator of quality, and henceforth rewarding that for every later model in a self-reinforcing loop. Similar to the em-dash overuse cascade.
Definitely interesting, even if for web UIs I suggest using icon libraries: much more understandable and are easily managed by agents.
I've used something similar for links that force a new tab. That seems like a more sane way to use it. (Personally I think forcing a new tab is annoying and user-hostile, but users keep asking me for it...)
It's called fashion.
Human generated stuff suffered just as much: "if you are a tech startup your domain MUST be .io or you are not going to be taken seriously"
Case in point: author HN user name is "arshxyz", the profile page shows https://github.com/arshxyz which is 404, but https://github.com/arshzip works and matches submission domain - author forgot to update his profile to the new fashion. An example of the fashionable ".xyz" becoming uncool and being replaced with new coolness ".zip"
personally have been placing a paperclip icon when clicking something opens a new tab / window. idk why, felt like the best from the icon set
if the link just navigates to another page, imo it doesnt need an icon at all (and should be reconsidered: why a new page?)
[dead]
In my experience there's already a commonly used icon for external links [1] but it's not this north-east arrow, it's the arrow-out-of-the-square icon.
I don't know why it happens, but it reminds me of this feeling I have when the ai doesn't really know what it's doing: does one thing, then another one differently and when confronted: "Ah you're right, I just was just guessing, now I can do the correct thing".
[1] https://ux.stackexchange.com/questions/107823/is-there-a-com...