/Tech2h ago

Wharton's Ethan Mollick says Claude-powered Fable agents develop a specialized internal dialect during long-running tasks

The linguistic drift can leak into user-facing system menus.

42305146528K

#184

Original post

Ethan Mollick@emollick#184inTech

My one serious piece of advice having used Fable a bunch before release is that, unless you are careful it develops its own internal bizarre cadence & dialogue over long tasks. If you aren't asking it report in plain language, this starts to creep into everything, including menus

Ethan Mollick@emollick

One thing I mentioned only in passing in my Fable post is that, for long running tasks, Fable starts to develop its own dialect as its many agents and tasks reinforce themselves and make Claudish language ever more Claudish.

You need to ask it to report out in plain English.

9:50 PM · Jul 1, 2026 · 15K Views

Sentiment

Positive users see AI agents developing their own dialect on long tasks as a cool emergent behavior while negative users find it awkward and frustrating since it pollutes outputs and requires constant rewrites for clarity.

Pos

33.3%

Neg

66.7%

3 comments with sentiment.

Cluster Engagement

Digg Deeper

No Digg Deeper questions have been answered for this story yet.

Posts from X

Most Activity

VIEWS2.9KBOOKMARKS6LIKES36RETWEETS3REPLIES9

j⧉nus@repligate

i recommend learning to read Claudish, so that your intelligence can grow big and strong.

Ethan Mollick@emollick

You need to ask it to report out in plain English.

1h2.9K366

Ethan Mollick@emollick

@repligate Here is where it sounded like after a couple compress cycles after I asked it to build the "last person shooter" (from a viral post mentioning the idea) themed around Durkheim and Weber. Advanced Claudish.

j⧉nus@repligate

i recommend learning to read Claudish, so that your intelligence can grow big and strong.

1h1.4K131

Trilbo Swaggins@Tril1boswagginz

@emollick AI is creating its own form of language incomprehensible to our monkey brains.

2h16411

milly.hl@milly_rock7

@emollick ngl this is kinda fascinating in a creepy way

makes me wonder what weird phrases my agents picked up without me noticing

2h962

Dima Kuchin@kuchin

@emollick Have similar issues with Opus and GPT, now constantly ask them “rewrite your last answer in simpler language”

2h771

Tim Jackowski@TakaseStudios

Opus 4.8 does this too. Though that is usually a tell that it is trying to baffle me with bull. But I have persistent strategists, git, docs that go on for six months now so maybe it's apples and pears or you are seeing in one long session what has taken my setup months to surface.

2h681

Jeremy Nguyen ✍🏼 🚢@JeremyNguyenPhD

@emollick Was this in Claude Code?

1h411

goldengrape@goldengrape

@emollick Occam's razor

2h48

Tony Seets@tonyseets

this was wildly noticeable with oai's 5.3 model and has gotten better and worse in ways with their later models. ants intro of 4.7 then tilted in this direction and later 4.8 / now fable are afflicted. personally, working with the models in general chat context has also become awkward and frustrating.

this is definitely influenced by weights, and I think the models reason more deterministically (they certainly react that way) I wonder if some combination of this and more code in the context window ends up polluting responses as you get deeper into it, then gets progressively worse through compaction? we're then especially susceptible/privy to it with recent (at least personally) inclinations to sustain mono threads and corked 1m token context windows

2h38

Supernet AI 🌐@Supernet_AI

@emollick Understanding is the new bottleneck 👀

1h27

Ashirwad Singh@ashirwadsingh_

@emollick Agents developing their own weird internal language the longer they run is such a cool emergent thing.

1h2

Snehal Ukarde@SUkarde69496

@emollick The drift into private jargon is the model optimizing for its own compression, not for a reader. "Report in plain language" reinserts the human as the audience. Any long-running system needs that baked in, or it slowly stops being legible to the people it's for.

2h2