AI-generated text has become common in academic and professional writing, prompting research into detection methods. Less studied is the reverse: systematically rewriting AI-generated prose to read as genuinely human-authored. We build a parallel corpus of 25,140 paired AI-input and human-reference text chunks, identify 11 measurable stylistic markers separating the two registers, and fine-tune three models: BART-base, BART-large, and Mistral-7B-Instruct with QLoRA. BART-large achieves the highest reference similarity -- BERTScore F1 of 0.924, ROUGE-L of 0.566, and chrF++ of 55.92 -- with 17x fewer parameters than Mistral-7B. We show that Mistral-7B's higher marker shift score reflects overshoot rather than accuracy, and argue that shift accuracy is a meaningful blind spot in current style transfer evaluation.
Figures & tables
Metric
Human
AI
Change
Signal
Avg word count
50.77
41.84
−17.5%
Moderate
Avg sentence count
3.78
2.41
−36.2%
Moderate
Avg word length (chars)
5.09
5.84
+14.7%
Strong
Lexical diversity
0.783
0.853
+8.9%
Moderate
Contractions per chunk
0.17
0.00
−100%
Very Strong
Question marks
0.084
0.035
−58.3%
Strong
Table 1: Chunk-level linguistic marker comparison between human-authored and AI-generated text (test subset, n=1,390 ).
Setting
BART-base
BART-large
Mistral-7B
Parameters
139M (full FT)
406M (full FT)
7B (LoRA only)
Adapter
None
None
QLoRA NF4, r =16, α =32
Epochs / Steps
1 epoch (smoke)
5 epochs
500 steps
Learning rate
5×10−5 (linear)
5×10−5 (cosine)
2×10−4 (cosine)
Batch size (eff.)
16
16
8
Precision
bf16
bf16
fp16 + 4-bit base
Table 2: Training configuration summary.
Model
Params
BERTScore F1
ROUGE-L
chrF++
PPL
Mean Shift
BART-base
139M
0.9088
0.4448
46.41
26.69
0.6513
BART-large
406M
0.9240
0.5657
55.92
27.15
0.8289
Mistral-7B
7B
0.8980
0.4642
55.68
0 9.03
1.2788
Human ref
—
—
—
—
23.69
—
Table 3: Quantitative evaluation results on the 1,390-example test set. Best values per metric in bold. PPL = GPT-2 perplexity; Shift = mean linguistic marker shift.
Figure 1: Reference-based metrics (BERTScore F1, ROUGE-L, chrF++) for BART-base, BART-large, and Mistral-7B on the 1,390-example test set.
Marker
AI Input
BART-base
BART-large
Mistral-7B
Human Ref
Note
Word count
41.8
48.9
46.3
77.9
50.8
△
Sentence count
2.41
3.53
3.41
5.84
3.78
△
Avg word length
5.84
5.52
5.09
5.16
5.09
Lexical diversity
0.853
0.766
0.783
0.510
0.783
△
Contractions
0.00
0.095
0.097
0.383
0.166
△
Question marks
0.035
0.047
0.048
0.112
0.084
Table 4: Per-marker averages across AI input, model outputs, and human reference. Key: △ = Mistral-7B overshoots the human target (shift >1 ). (WD) = wrong direction: Mistral-7B shift <0 , meaning the output moves away from the human mean rather than toward it.
Marker
BART-base
BART-large
Mistral-7B
Note
Word count
0.785
0.496
2.000
Mistral overshoots
Sentence count
0.823
0.730
2.000
Mistral overshoots
Avg word length
0.438
1.000
0.908
BART-large exact
Lexical diversity
1.245
1.004
2.000
Mistral overshoots
Contractions
0.571
0.584
2.000
Mistral overshoots
Question marks
0.246
0.275
1.565
Table 5: Directional marker shift scores. Values >1 indicate overshoot; values <0 indicate movement in the wrong direction.
Figure 2: Per-marker directional shift scores for all three models. Values above 1.0 indicate overshoot beyond the human target; values below 0 indicate movement in the wrong direction.
Figure 3: GPT-2 perplexity for model outputs and the human reference baseline (23.69). Lower is not always better: Mistral-7B’s score of 9.03 reflects over-smooth generation rather than human-like naturalness.
Interdisciplinary School of Doctoral Studies University of Bucharest, Romania · Faculty of Mathematics and Informatics University of Bucharest, Romania