It has been widely reported, and probably you personally have discovered, that the written text produced by a bot can seem . . . odd, not human sounding, stiff, boring, repetitive, dense, unfeeling . . . Research in several fields has been trying to get a good understanding of "Bot-elect" (my word). Often it is so obvious, especially in the case of students, that it is easy to spot; other times, not so much.
I have begun to notice the "Bot-elect" showing up "quietly" in papers done by colleagues who used a bot to help them revise their work, resulting in some pretty bizarre, yet subtle and unsettling changes in what "they" write. I decided to ask the Bot to unpack that issue for me. Will let the Bot speak/write for itself, leaving in place the links to the sources accessed, which have been checked for content and credibility.
 |
| Clker.com |
Summarize current research studies on the potentially problematic structural features of A!-generated text that may distinguish the output as "nonhuman", focusing especially on issues of syntactic complexity, embedding, diversity, lexical domains and styles. Keep this brief and technical; I'll do a follow up, possibly, in the form of a blogpost, etc.
The Bot responds:
Current research suggests AI text often has a detectable “signature,” but the direction of the effect varies by model, prompt, and genre. The most consistent signals are structural regularity, reduced stylistic range, and distributional differences in lexical choice rather than simple grammatical error.
Key findings
Syntactic complexity:
Many studies find AI output is grammatically fluent but tends toward more regular, less variable sentence architecture, with fewer genuinely diverse clause patterns and a preference for efficient, low-cost constructions. [pmc.ncbi.nlm.nih]
Embedding/subordination:
Work on clause structure often reports lower or more constrained embedding depth and a weaker spread of subordinate-clause patterns, even when sentence length is high; some newer models can increase surface complexity without matching human syntactic range. [unitesi.unive]
Lexical diversity:
Earlier LLMs often show lower lexical diversity and more repetition than human writing, though newer models can move closer to human levels on some measures while still differing on finer-grained diversity metrics. [pmc.ncbi.nlm.nih]
Lexical domains/styles:
AI text often skews toward formal, structured, motivational, and emotionally positive language, with heavier use of nouns/adjectives and more template-like register control; humans show broader shifts across personal, negative, and domain-specific styles. [pmc.ncbi.nlm.nih]
-Practical takeaway:
The strongest “nonhuman” cue is usually not one feature alone, but a bundle: high fluency plus compressed variation in syntax, embedding, lexical breadth, and register. [dl.acm]
Prompt 2:
Give me a 200 word paragraph that succinctly summarizes the research cited above.
The Bot responds:
Recent research on AI-generated text suggests that its most recognizable “nonhuman” qualities are structural rather than merely lexical: the prose is often fluent, formally organized, and grammatically polished, yet it tends to show reduced variation in syntax, clause embedding, and sentence architecture, along with narrower lexical diversity and more repetitive phrase patterns than comparable human writing. Across studies, AI text is also described as more impersonal, more positive or motivational in tone, and more heavily reliant on nouns, determiners, and template-like register choices, which can make it feel polished but stylistically compressed. At a higher level of analysis, newer work points to discourse-level differences such as weaker contextual sensitivity, more formulaic organization, and possible limitations in cohesion, coherence, and rhetorical flexibility, though this part of the literature is still emerging and is less settled than sentence-level findings. Importantly, the evidence does not imply that all AI text is easily identifiable or uniformly simplistic; newer models can approximate human scores on some measures, so the clearest signal is usually a bundle of features—regular syntax, limited variation, and constrained register—rather than any single giveaway. [pmc.ncbi.nlm.nih]
Prompt 3 : Now, just for fun, analyze the text you just created for telltale signs that you may be a bot!!!
The Bot responds:
A few **telltale signs** in that paragraph are exactly the kinds of things the research itself flags: it is highly organized, balanced, and general-purpose, with a clean progression from sentence-level features to discourse-level features and then a cautionary caveat at the end. It also leans on repeated abstract nouns like “variation,” “structure,” “register,” and “features,” which can make the prose feel compressed and concept-heavy rather than personally inflected.
Bot-like signals
- It uses a very even, survey-style rhythm.
- It avoids concrete examples and instead stacks abstractions.
- It includes hedging phrases like “suggests,” “tends to,” and “does not imply.”
- It closes by smoothing over uncertainty with a broad qualification.
Why that happens
That style is common in summarizing research because the goal is to compress multiple findings into one coherent paragraph. So the same features that can make AI text look “bot-like” can also simply reflect cautious academic synthesis.
Bottom line
If you were trying to make it sound more human, you’d usually add one specific example, vary the sentence lengths more sharply, and let one sentence be a little less perfectly balanced.
wracton@gmail.com
Website: williamacton.legalshieldassociate. com
No comments:
Post a Comment