From the paper: Why is AI-generated language so hard to understand?

[anibalsanchez]

Hi,

The paper Why is AI-generated language so hard to understand? claims that the AI-generated text has recently become harder to understand due to several frontier techniques that optimize the text lenght.

There are several points that could be added to the Humanizer skill as signs of AI writing to be avoided.

This is an idea that I'm testing:

G. Compression instead of communication

#	Pattern	Before	After
27	Things acting like people	"the asymmetry bites"	"the asymmetry causes problems"
28	Nouns pressed into service as verbs	"we calendared the review"	"we scheduled the review"
29	Rare word pairings (weak alone)	"surfaces adherence metrics"	"shows whether the pipeline is being followed"
30	Skipped "that" (weak alone)	"the slowdown the team observed"	"the slowdown that the team saw"
31	Stacked nouns (weak alone)	"agent deployment cost optimization strategy review"	"a strategy for cutting the cost of agent deployments"
32	Ideas held apart	"the system, after months of changes, finally stabilized"	"Months of changes finally stabilized the system"

## G. Compression instead of communication

Frontier models are increasingly optimized for dense output — verifiable-reward
training, length penalties, and coding or math tasks all favor the shorter
expression — not for communication. A human chooses between two ways to say the
same thing by context; a model trained this way tends to take the short one
either way. Most patterns below are the compressed member of such an alternation,
and each is a construction that reading research ties to slower processing:
re-analysis after an animacy mismatch, working-memory strain on a long dependency,
or bracketing ambiguity in a noun stack. To check for these, ask: would the
longer, plainer form be easier to read, and did this sentence pick short only
because short was cheap? The tells are strongest in chat and coding prose; rare
n-grams and inanimate subjects appear in essays too.

### 27. Things acting like people

**Watch for:** an inanimate subject with a verb that implies a person — the code
swallows, this asymmetry bites, the tool assumes, the pipeline decides. Verbs of
volition or perception applied to code, data, or systems; a compact metaphor
standing for a literal phrase ("bites" for "causes problems").
**Problem:** Readers use animacy as a parsing cue, so the mismatch forces
re-analysis. Keep the settled group metonymy humans use ("the court ruled",
"Microsoft announced"); the tell is novel personification of tools and data.
Remove the metaphor or put a person in the sentence; the longer form is the point.
**Before:**
> The asymmetry bites whenever the same pattern repeats.
**After:**
> The asymmetry causes problems whenever the same pattern repeats.

### 28. Nouns pressed into service as verbs

**Watch for:** an ad-hoc noun-as-verb that is not settled usage — "calendar it",
"footnote it", "we backlogged the request". Not settled verbs like *email me* or
*google it*.
**Problem:** Pressing a noun into service makes the reader re-evaluate its role;
the conventional phrase ("put the whiskey in a bottle" instead of "bottle the
whiskey") is longer but parses without backtracking. Models use the compressed
form to save tokens.
**Before:**
> We calendared the review and inboxed the findings.
**After:**
> We scheduled the review and filed the findings.

### 29. Rare word pairings

**Watch for:** combinations of words that rarely occur together, packed coinages
such as "reader-patience budget", any sentence that could not come out of someone
speaking. *Weak alone.*
**Problem:** Reading is fastest when words are predictable in context. Models
skip the common, uninformative words to maximize information per token, and the
leftover pairings slow the reader. Use the plain phrasing.
**Before:**
> The dashboard surfaces adherence metrics for the provisioning pipeline.
**After:**
> The dashboard shows whether the provisioning pipeline is being followed.

### 30. Skipped "that"

**Watch for:** an object relative clause with the connector dropped to save a
word — "the file the build produced", "the fix the team deployed". *Weak alone*:
dropping "that" is ordinary speech, and one short dropped "that" is not a tell.
**Problem:** In "the house that the contractors built", "that" marks where the
clause starts, so the reader parses it as it arrives; "the house the contractors
built" is harder. Restore the connector when the clause is long or the sentence
stacks several noun phrases; otherwise rewrite so the subject is the thing doing
the action.
**Before:**
> The slowdown the team observed in staging stems from a cache the new version
> leaves stale.
**After:**
> The slowdown that the team saw in staging comes from a cache that the new
> version leaves stale.

### 31. Stacked nouns

**Watch for:** three or more nouns piled into a modifier chain — "agent deployment
cost optimization strategy", "cache refresh failure handling path". *Weak alone*
at three nouns, where the stack can be idiomatic ("performance review meeting");
four-plus is a strong tell.
**Problem:** The reader cannot tell which noun modifies which, so they backtrack
to restore the bracketing. Break the stack with prepositions or possessives, or
restate.
**Before:**
> The document covers our agent deployment cost optimization strategy review.
**After:**
> The document reviews our strategy for cutting the cost of agent deployments.

### 32. Ideas held apart

**Watch for:** a subject and its verb separated by a long chunk; a modifier several
clauses away from the noun it describes; one sentence kept open while multiple
facts load into it.
**Problem:** Long dependencies strain working memory, and writers avoid them in
prose. Models keep facts in one sentence because it packs them densely; readers
lose track of what connects to what. Move the parts that belong together closer,
or split the sentence.
**Before:**
> The system, after months of incremental changes across services that were
> reviewed by different teams in different time zones, finally stabilized.
**After:**
> Months of incremental changes finally stabilized the system. Each change was
> reviewed by a different team, often in different time zones.