July 14, 2026
Delete means DELETE, darling
Accretive Editing
AI won’t delete its exes, and the comments are absolutely losing it
TLDR: The big issue: AI editing tools often keep outdated information and tack corrections onto it instead of cleanly fixing the text. Commenters were split between “just prompt it harder” and “this is a deeper design flaw,” with plenty of jokes about bots being unable to let go of the past.
A small writing problem has turned into full-on comment section theater. In Justin D. Fuller’s post, the complaint is simple: when you ask an AI helper to update old text, it often refuses to fully let go of the past. Instead of cleanly replacing the outdated line, it keeps the old fact hanging around like an awkward ex at a wedding. One example: a project stopped working with Amazon Bedrock and started working with LiteLLM, but the AI rewrote the sentence to mention both, creating a clunky little side-note instead of a clean correction.
The crowd immediately jumped in with a mix of practical advice, eye-rolling, and meme energy. One camp says the fix is easy: tell the bot to do a “hard switch” and write as if the old fact never existed. Another went even more dramatic, saying they sometimes have to instruct it to imagine an “alternate reality” where the wrong information never happened at all. That line alone basically wrote the jokes for everyone.
But the spiciest reactions came from people arguing this is not just a wording issue. Some suspect modern AI systems are now overly scared of deleting things, so they keep stuffing in caveats and warnings instead of making clean edits. And then there was the vocabulary snark: one commenter joked that suddenly it’s “accretive this, accretive that,” as if the entire internet got haunted by one dictionary page. In other words, the docs drama is real, the comments are savage, and everyone agrees on one thing: AI really needs to learn how to break up properly.
Key Points
- •The article defines accretive editing as an AI editing behavior where outdated text is preserved and amended instead of fully corrected.
- •It uses a documentation update from Amazon Bedrock support to LiteLLM support as a real-world example of the problem.
- •The article states that historical changes should be communicated in changelogs, announcements, or callouts rather than embedded throughout revised prose.
- •It argues that prompting the model to write less or changing style guidance does not eliminate the failure mode.
- •The article recommends instructing AI to replace obsolete text with accurate text so the final document reads as if it were correct from the start.