Building Auditable AI Text Detectors with Sparse Autoencoders

Liam Dugan; Callum McDougall; Matthew Denton; Philip Pham; Christine Kaeser-Chen; Chris Callison-Burch

cs.CLOct 21, 2026

Building Auditable AI Text Detectors with Sparse Autoencoders

Authors: Liam Dugan, Callum McDougall, Matthew Denton, Philip Pham, Christine Kaeser-Chen, Chris Callison-Burch

Organizations: University of Pennsylvania · Google DeepMind

Abstract

AI-generated text detectors are increasingly deployed in consequential settings---peer review, education, admissions, journalism, and content moderation---where their outputs directly affect individuals. Yet the strongest detectors are fine-tuned neural classifiers whose decisions are opaque, while transparent alternatives (stylometric and linguistic-feature classifiers) underperform them by wide margins. We present a pipeline that converts a fine-tuned neural detector into a classifier that operates on human-readable concepts. We accomplish this by training sparse autoencoders (SAEs) directly on a detector's residual stream, assigning auto-interpretation labels to each SAE feature, and retaining only features whose generated descriptions pass a global fidelity threshold. We then train a lightweight classifier on just the interpretable features. We find that our SAEs recover latent concepts that correspond to highly specific and predictive linguistic "tells" of AI-generated text. Classifiers built on these extracted features are not only highly interpretable but exhibit stronger performance on out-of-distribution texts than either neural or stylometric alternatives.

Figures & tables

Figure 1: In our work, we use Sparse Autoencoders to convert existing black box AI detectors into ones that operate on human-interpretable concepts—allowing for more auditable deployment.

Figure 2: Our pipeline. Starting from a fine-tuned AI-text detector, we train SAEs on its residual stream at four layers, label each latent with an auto-interp explanation, mask out all features that do not meet our global threshold and train a lightweight classifier over the remaining concepts. The result is a detector that retains the strength of the original neural model while being composed of auditable and intervenable human-understandable concepts.

Figure 3: Distribution of generative models for COLING, PRDB, and RAID datasets broken up by model family. We see that both our training and evaluation data span a large variety of model families.

Figure 4: F1 Score (COLING) and TPR@FPR=5% (RAID) (y-axis) vs. Interpretability Threshold (x-axis) for classifiers trained on the COLING Shared Task 1 dataset evaluated on ID (grey) and OOD (white) data. We see that FT-Model+SAE matches the performance of the underlying neural detector even at high interpretability thresholds (0.7+) and that Base-SAE and FT-SAE models exceed the detector’s performance on OOD data.

	COLING (dev)			COLING (devtest)			COLING (test)			RAID (test)
	AUC	F1	TPR@ $5$	AUC	F1	TPR@ $5$	AUC	F1	TPR@ $5$	AUC	F1	TPR@ $5$
Binoculars †	75.8	73.1	56.0	88.6	81.6	67.2	86.7	81.4	59.9	85.0	87.6	79.1
Fast-DetectGPT †	75.7	70.5	56.2	83.2	75.5	51.6	85.1	77.5	58.5	83.3	85.2	75.1
ReMoDetect †	76.5	65.1	45.6	91.8	85.2	65.2	91.6	86.7	75.0	73.2	67.8	44.1
SpecDetect †	80.2	74.4	57.2	81.3	74.2	48.9	83.6	74.6	56.4	91.6	88.1	77.3
Baselines

Table 1: Model performance across held out splits of RAID and COLING datasets. Dagger

(\dagger)

indicates off-the-shelf models not trained on the COLING train set. We see that FT-Model+SAE matches the performance of the underlying neural detector, that Base-SAE and FT-SAE models exhibit strong generalization performance, and that interpretability masking does not substantially reduce performance on OOD examples.

	GPT-4o			Llama 3.1 70B
	AUC	F1	TPR@1	AUC	F1	TPR@1
Binoculars †	95.7	89.8	34.5	99.9	99.0	99.1
SpecDetect †	92.2	84.4	29.4	98.4	93.8	80.9
ReMoDetect †	99.4	98.4	89.6	99.8	99.0	98.5
Fast-DetectGPT †	81.0	74.1	10.6	99.3	98.6	85.4
Baselines

Table 2: Performance on in-distribution test split of PRDB. Dagger

(\dagger)

indicates off-the-shelf models. We see that all models achieve near perfect performance.

Figure 5: Percentage of AI-written documents whose predictions were flipped to “Human” after paraphrasing with Gemini using each of our five prompting strategies: Baseline, Few Shot, and Guided are generic paraphrase prompts, while SAE and SAEx3 use our interpretable features one-shot and iteratively (

n=3

). We see that having access to SAE features in the prompt allows the model to more easily humanize the text. See Table 23 for extended metrics.

	w/o Mask				w/ Mask ( $\tau\!>\!0.7$ )
	Num	Avg. Feat Shift	Same	Importance	Num	Avg. Feat Shift	Same	Importance
	Rewrites	$\{\Delta_{\text{tgt}},\ \Delta_{\text{non}}\}$	Sign	vs. Para. $\boldsymbol{\rho}$	Rewrites	$\{\Delta_{\text{tgt}},\ \Delta_{\text{non}}\}$	Sign	vs. Para. $\boldsymbol{\rho}$
COLING
Base-SAE	16,052	{0.31, 0.17}	65.0%	0.36	19,318	{0.38, 0.12}	59.9%	0.41
FT-SAE	15,968	{0.36, 0.23}	72.9%	0.52	17,826	{0.37, 0.14}	80.8%	0.54
FT-Model+SAE	28,965	{0.25, 0.22}	68.9%	0.54	27,866	{0.63, 0.27}	73.2%	0.62

Table 3: Results of the feature-level paraphrase ablation. We report correlation between importance score and paraphrase ablated logit difference.

\Delta_{\text{tgt}}

and

\Delta_{\text{non}}

give the normalized activation shift of the targeted and non-targeted features under the paraphrase as well as the sign agreement. We see features are attributed faithfully across both datasets and see meaningful improvement with

\tau>0.7

indicating its utility as an interpretability heuristic.

Model	Feature	Description (auto-interp: gemini-3.1-flash-lite)	$\tau$	Logit $\Delta$	#Docs (Human / AI)
GPT-4o	13_15950	academic paper review questions asking for more information details or clarification	0.866	$+1.002$	$1{,}400$ / $79{,}922$
	17_8459	academic review feedback or peer review decisions	0.870	$+0.814$	$229$ / $96{,}612$
	22_4800	double colons following bolded markdown section headers	0.860	$+0.672$	$11$ / $54{,}293$
	17_12216	activates on the word “potential”	0.884	$+0.515$	$15{,}012$ / $92{,}929$
Llama	17_4373	the word “thorough”	0.892	$+0.993$	$9{,}606$ / $79{,}708$
	13_12061	numbered list items in parentheses within question sections of research papers ( e.g., (1) , (2) )	0.858	$+0.898$	$21$ / $27{,}425$

Table 4: Top features for FT-Model+SAE (

\tau>0.85

) on the PRDB dataset Yu et al. (2026) , ranked by the mean change in detector logit when the feature is zero-ablated (Logit

\Delta

). Each row reports a feature’s auto-interp explanation, simulation score

\tau

, Logit

\Delta

, and number of occurrences in human and AI text.

Appendix figures & tables36 assets

Supplementary material from the paper’s appendix.

Appendix

SAE Model Hyperparameters
optimizer *	optim.Adam
$\beta_{1},\beta_{2},\epsilon$ *	$0.0,0.999,10^{-8}$
sae_width *	$16,384$
sparsity_coeff *	$6\times 10^{-4}$
sparsity_warmup_steps *	$10,000$
sp_scheduler_type *	linear

Appendix

Table 5: Hyperparameters used for fine-tuning the GemmaScope Lieberum et al. (2024) SAEs. Asterisks indicate settings inherited from GemmaScope.

	Layer 7		Layer 13		Layer 17		Layer 22
	L0	FVU	L0	FVU	L0	FVU	L0	FVU
COLING (base)	48.73	11.27%	67.77	15.49%	66.60	13.53%	73.97	13.8%
COLING (open)	40.19	3.71%	24.03	3.37%	16.04	3.78%	7.18	5.22%
PRDB (base)	31.57	7.18%	46.35	10.50%	45.01	8.53%	44.87	8.56%
PRDB (GPT-4o)	12.59	2.50%	7.20	2.95%	5.87	2.68%	4.64	3.47%
PRDB (Llama)	16.02	1.50%	9.28	1.38%	6.67	1.50%	5.24	2.15%

Appendix

Table 6: Evaluation results for our fine-tuned SAEs on the eval splits of the COLING and PRDB dataset. We report number of active features per example (L0) and reconstruction fidelity (FVU) and note that on average, base model SAEs have worse reconstruction performance than the fine-tuned model SAEs.

	SAE	gram2vec
Optimizer	SGD + Momentum
Momentum	$0.9$	$0.9$
Learning rate	$3{\times}10^{-2}$	$3{\times}10^{-2}$
L1 penalty ( $\lambda_{1}$ )	$10^{-1}$	$10^{-2}$
Epochs	$3$	$3$
Batch size	$1024$	$1024$

Appendix

Table 7: Hyperparameters used for training the logistic regression classifiers for the SAE detectors and the gram2vec linguistic feature baselines.

Figure 6: COLING-dev F1 against the mean number of active features per document (log scale) for the SGD +

L_{1}

grid, three epochs. Columns are the feature settings, rows the mask thresholds; each line follows one learning rate through the five

\lambda

values (leftmost point

\lambda=1

, rightmost

\lambda=10^{-2}

XGBoost Hyperparameters
objective	reg:squarederror
n_estimators	150
max_depth *	6
subsample *	1
reg_lambda *	1
learning_rate	0.1

Appendix

Table 8: Hyperparameters used for fine-tuning the XGBoost classifiers for both the linguistic feature baselines and the SAE classifiers.

	dev	devtest	test
RoBERTa-baseline	96.2	83.1	74.9
gemma-2-2b	98.0	86.8	75.2

Appendix

Table 9: F1 score of the RoBERTa baseline provided by the COLING shared task organizers against our fine-tuned gemma-2-2b neural detector (§ 5.2 ).

Bin	Auto-interp $\tau$	# Feats	# Pairs	$\rho$
0	$[\,0.700,\ 0.707\,]$	44	14,733	0.778
1	$[\,0.707,\ 0.719\,]$	20	6,162	0.791
2	$[\,0.719,\,0.750\,]$	31	10,276	0.825
3	$[\,0.750,\,0.866\,]$	29	10,169	0.871
All	$[\,0.700,\,0.866\,]$	124	41,340	0.809

Appendix

Table 10: Agreement between zero-ablation and paraphrase logit deltas as a function of auto-interpretation score ( FT-Model+SAE ,

\tau\!>\!0.7

, PRDB, GPT-4o target). Features are binned by quartiles of their auto-interp score; within each bin, Spearman

\rho

is computed between the per-feature mean zero-ablation

\Delta

logit and the per-feature mean paraphrase

\Delta

logit. We see that correlation between paraphrase ablation and zero-ablation ablation score goes up with increasing autointerp score.

Neural AI Detector Hyperparamters
optimizer *	adamw_torch
$\beta_{1},\beta_{2},\epsilon$ *	$0.9,0.999,10^{-8}$
precision *	float32
lr_scheduler_type *	linear
gradient_accum_steps *	$1$
max_grad_norm *	$1.0$

Appendix

Table 11: Hyperparameters used for fine-tuning Gemma-2-2B model on AI detection. Asterisks indicate defaults inherited from Huggingface Trainer.

Figure 7: Histogram of Log Odds Ratio (x-axis) of all features from the Open Domain AI generated text detectors. We see that FT-Model+SAE produces far more highly predictive features on the in-distribution data but once out-of-distribution the features are no more predictive than those from FT+SAE or Base-SAE .

Data	SAE	$\tau\!>\!0.0$	$\tau\!>\!0.7$	$\tau\!>\!0.85$	$\tau\!>\!0.0$	$\tau\!>\!0.7$	$\tau\!>\!0.85$	$\tau\!>\!0.0$	$\tau\!>\!0.7$	$\tau\!>\!0.85$
		Interpretable			Interp. & $w>0$			Avg. # active / doc
COLING	Base-SAE	65,512	33,624	9,355	201	189	149	101	107	108
COLING	FT-SAE	66,245	43,480	13,051	118	136	155	100	101	101
COLING	FT-Model+SAE	29,489	16,893	2,983	63	65	65	53.7	41.9	25.6
PRDB	Base-SAE	65,153	22,332	5,544	82	156	71	79.6	116	68.7
PRDB	FT-SAE	64,997	39,231	7,967	71	85	115	99.3	83.7	85.1

Appendix

Table 12: SAE feature counts summed over the four probed layers. Interpretable : features whose auto-interp score exceeds

\tau

. Avg. # active / doc : mean number of nonzero-weight interpretable features that fire on a held-out test document, i.e. that contribute a nonzero logit term to the prediction. PRDB rows use the GPT-4o target.

Figure 8: Density normalized histograms of feature auto-interp correlation scores (x-axis) across our conditions. We can see that in both cases, the FT-SAE model retains a higher proportion of highly interpretable features.

Feature-Level Paraphrase Results $\tau>0.7$ — Paper Reviews
Feature	Concept Description	# Docs	Ablate $\Delta$	Para. $\Delta$	$\boldsymbol{\Delta_{\text{tgt}}}$	$\boldsymbol{\Delta_{\text{non}}}$
Base-SAE (GPT-4o)
7_16119	words related to claims assertions and disclaimers	495	$+0.0139$	$+0.0382$	0.70	0.02
13_6062	period punctuation mark at the end of sentences	434	$-0.0690$	$-0.151$	0.67	0.09
13_377	markdown formatting for section breaks headers or newlines in academic reviews	424	$-0.0495$	$-0.109$	0.22	0.06
22_632	subject pronoun it followed by a verb phrase	386	$-0.0189$	$-0.230$	0.48	0.06

Appendix

Table 13: The 6 features most often sampled for paraphrase by each of the Base-SAE , FT-SAE and FT-Model+SAE (

\tau>0.7

) counterfactual experiment on PRDB, ranked by the number of documents in which they were targeted (descriptions from gemini-3.1-flash-lite auto-interp). Each row reports the feature’s mean change in detector logit under zero-ablation (Ablate

\Delta

) and under Gemini’s rewrite (Para.

\Delta

), and the mean normalized change in activation that the rewrite induced in the targeted feature (

\Delta_{\text{tgt}}

; 1.0 is complete removal) and in all other active features (

\Delta_{\text{non}}

Feature-Level Paraphrase Results $\tau>0.7$ — General Domain Text
Feature	Concept Description	# Docs	Ablate $\Delta$	Para. $\Delta$	$\boldsymbol{\Delta_{\text{tgt}}}$	$\boldsymbol{\Delta_{\text{non}}}$
Base-SAE
22_1866	introductory phrases or transitional clauses introducing topics or evidence	523	$+0.148$	$+0.0391$	0.09	0.26
17_9942	punctuation marks like periods and apostrophes	519	$+0.374$	$-0.122$	0.01	0.22
17_15019	opening quotation marks followed by quoted dialogue or official statements	422	$+0.0545$	$-0.121$	0.16	0.05
13_13241	period punctuation mark at the end of sentences	407	$-0.326$	$-0.169$	0.35	0.11

Appendix

Table 14: The 6 features most often sampled for paraphrase by each of the Base-SAE , FT-SAE and FT-Model+SAE (

\tau>0.7

) counterfactual experiment on COLING, ranked by the number of documents in which they were targeted (descriptions from gemini-3.1-flash-lite auto-interp). Each row reports the feature’s mean change in detector logit under zero-ablation (Ablate

\Delta

) and under Gemini’s rewrite (Para.

\Delta

), and the mean normalized change in activation that the rewrite induced in the targeted feature (

\Delta_{\text{tgt}}

; 1.0 is complete removal) and in all other active features (

\Delta_{\text{non}}

Figure 9: Per-feature agreement between the two interventions (§ 7 ), for the logistic-regression heads. Each point is one SAE feature. The

x

-axis is the feature’s mean change in the detector’s AI logit when the feature is zero-ablated, and the

y

-axis is the mean change when the text is instead paraphrased to remove the feature, both averaged over the documents in which the feature was sampled. Positive values on both axes mean the feature pushed the prediction toward AI and removing it from the text lowered the AI logit. Columns are the feature settings; the top two rows are COLING without and with the interpretability mask (

\tau>0.7

), the bottom two rows the same for PRDB (GPT-4o).

Input text	Active features (Logit $\Delta$ )	Score
(GPT-4o) The paper titled “Sub gradient Descent 5 Lear 5,e ns Orthogonal Dictionaries 5 ” presents 3 a significant contribution to 3 the field of dictionary learning, particularly 4 in addressing the challenge of learning orthogonal dictionaries through a novel approach involving subgradient descent. The authors successfully demonstrate that a subgradient descent algorithm with random initialization can learn orthogonal dictionaries using a nonsmooth, nonconvex $\ell$ 1 minimization formulation. This approach contrasts d,e with prior methods that necessitate complex initialization strategies or computationally expensive operations. The research marks a 7 noteworthy advancement 7 by providing the first c theoretical guarantee for dictionary recovery via subgradient descent on this specific problem formulation. The paper’s originality stems from the novel application of subgradient descent in a nonconvex, nonsmooth optimization landscape, tackling issues directly rather than circumventing them through smoothing techniques. This contributes valu 2,a ably to the 2 dictionary learning domain and offers insights that could extend 2 to other nonconvex, nonsmooth optimization problems 6 . The use of Riemannian subgradient descent on a nonsmooth formulation is particularly 2,4 innovative 1 , and their theoretical analysis of the optimization landscape dynamics is rigorous. In terms b of scientific rigor, the paper is comprehensive in its theoretical underpinnings.	Toward AI 1 $+0.27$ innovative variations like innovative innovations or invasive 2 $+0.22$ academic phrases predicting future influence or impact on researchers 3 $+0.22$ academic paper verbs introducing research contributions 4 $+0.19$ word particular and variations like particularly in academic writing context 5 $+0.16$ multi-word academic terminology specifically within titles or descriptive naming conventions 6 $+0.16$ word environments appearing in technical literature regarding system deployment or model applicability 7 $+0.13$ words related to progress or improvements such as advances advancement advancements or shifts + 15 more $+0.88$ Toward human a $-0.10$ typos or misspellings within technical or academic documents b $-0.03$ idiomatic transition phrases like to put it bluntly to be clear and to illustrate this with a c $-0.02$ phrase to the best of our knowledge or similar personal epistemic qualifiers d $-0.02$ comparing similarity or difference between methods or concepts e $-0.01$ first person subjective expressions or comparisons of ideas to previous work	$\Sigma$ AI $+2.22$ $\Sigma$ human $-0.18$ bias $-0.51$ logit $+1.55$ $P(\mathrm{AI})$ $82\%$ $\rightarrow$ AI

Appendix

Table 15: The FT-SAE (

\tau=0.85

) logistic-regression detector applied to a GPT-4o-written review from the PRDB test set ( ICLR2019-1259-r1g5pn1527 ). Tokens on which a listed feature fires are shaded red (feature pushes toward AI) or blue (toward human), with three intensities set by the token-level contribution

|w_{i}a_{i,t}|

of the strongest listed feature on that token.

Input text	Active features (Logit $\Delta$ )	Score
(Llama3) Based on the lab values e , 6,a it appears that the neonate is experiencing a mild to f moderate neonatal anemia, f likely related to the g mother ’s g chorioamnionitis . a The low f MCH, MCHC, and platelet count e,f , along with the high RDW and neutrophils, suggest that the neonate may be f experiencing a hem g olytic process, possibly due to maternal g infection . a The slight e,f enlarged platelets f and poly chrom b asia also support f this diagnosis . a The anisocytosis may indicate that 6 the neonate is f experiencing 6 a degree of d hemolysis, which is consistent with the mother’s ch orio 3 amnionitis. It 4,a ’ s 1,4 essential 1,2 to 1,a monitor the a neon ate a ’ s condition closely a , 6,a as a chorioamnionitis can lead to c neonatal sepsis and other complications. The a neon ate may require a close monitoring a , 1,6,a possible 6,a antibiotics a , a,g and 6,a,g possibly a,g blood trans g fusions to manage 6 the g anemia . a Please 1,a note 1,2,5,7 that 1,5,a this is 1 a 1,a,e general interpretation 1,b and 1,a,b a more 1,a detailed 1,a,b evaluation should be performed 1,a by 1,a,c a 1,a,g healthcare 1 professional. 1,a	Toward AI 1 $+2.10$ introductory phrases or filler text commonly used by ai language models when declining personal questions 2 $+0.26$ word important and its synonyms like crucial when used to emphasize the significance of advice or caution 3 $+0.19$ repeated phrases or sequences in text 4 $+0.11$ activates on the phrase it is specifically when used as a formal or impersonal subject 5 $+0.07$ transitions used to summarize or conclude a text 6 $+0.06$ words and phrases that introduce supplementary information or instructions within a sequence 7 $+0.06$ dates and timestamps in forum posts or metadata + 7 more $+0.25$ Toward human a $-0.23$ text following or concluding medical advice or health consultations b $-0.18$ words related to estimations or imprecise approximations c $-0.13$ parenthetical clauses or aside remarks in text d $-0.08$ words related to quantifying or subsetting groups like part some or partially e $-0.06$ accounting and bookkeeping ledger entries involving debit credit and bank account transactions f $-0.06$ subjective judgment phrases or opinions about quality g $-0.04$ words related to physical sensations or gore such as bleeding footprints or bloodcurdling screams + 6 more $-0.76$	$\Sigma$ AI $+3.10$ $\Sigma$ human $-1.54$ bias $+0.61$ logit $+2.14$ $P(\mathrm{AI})$ $89\%$ $\rightarrow$ AI

Appendix

Table 16: The FT-SAE (

\tau=0.85

) logistic-regression detector applied to a Llama3-written document from the COLING test set ( coling29322_CUDRT ). Tokens on which a listed feature fires are shaded red (feature pushes toward AI) or blue (toward human), with three intensities set by the token-level contribution

|w_{i}a_{i,t}|

of the strongest listed feature on that token.

Input text	Active features (Logit $\Delta$ )	Score
(Human) This paper deals with neuromorphic computing architectures for solving deep learning problems. Given that the energy-efficiency is a critical need for modern machine learning tools, this is an important study in my b,c opinion. Until now, 8 most of the spiking neuron based works don’t really apply e for realistic deep model learning. Therefore, advancements 2,3,4,7,8 such as these 8 would 5,8 be 8 very d,f useful. 8 However, my biggest concern is that the paper may be a bit difficult 1,6,a to 1,a grasp for the general 1 ML audience 1 . 6 I myself c am c,g,h not an expert in g,h this g topic g,h and had quite a bit of h difficulty a,h as h an ML researcher in understanding g,h parts of h the paper 6 . h	Toward AI 1 $+0.19$ text concerning accessibility or difficulty of academic papers for specific readerships 2 $+0.11$ words related to progress or improvements such as advances advancement advancements or shifts 3 $+0.09$ words related to advancement or progress including advances and advancing 4 $+0.09$ words related to research and development progress specifically experimentation and advancement 5 $+0.07$ academic phrases predicting future influence or impact on researchers 6 $+0.06$ academic writing style describing paper accessibility or technical complexity 7 $+0.04$ academic terminology related to progress or improvements in research 8 $+0.04$ advancements or improvements in technical research or methodology + 6 more $+0.09$ Toward human a $-0.39$ readability of academic papers specifically phrases describing ease or difficulty of comprehension b $-0.25$ first person singular pronouns referring to the author or reviewer c $-0.18$ first-person opinion markers like i believe or i would say d $-0.14$ intensifier words like extremely and very e $-0.11$ contractions and auxiliary verb phrases like ve, s, and t f $-0.11$ intensifier adverb very or so often used to modify adjectives or quantifiers in critical evaluation contexts g $-0.08$ expressing lack of domain knowledge or expertise h $-0.08$ first person singular phrases expressing confusion or inability to understand something + 9 more $-0.21$	$\Sigma$ AI $+0.78$ $\Sigma$ human $-1.53$ bias $-0.51$ logit $-1.26$ $P(\mathrm{AI})$ $22\%$ $\rightarrow$ Human

Appendix

Table 17: The FT-SAE (

\tau=0.85

) logistic-regression detector applied to a human-written review from the PRDB test set ( ICLR2020-205-rkxbmCrTKS ). Tokens on which a listed feature fires are shaded red (feature pushes toward AI) or blue (toward human), with three intensities set by the token-level contribution

|w_{i}a_{i,t}|

of the strongest listed feature on that token.

Input text	Active features (Logit $\Delta$ )	Score
(Human) Post rebuttal: the author feedback d addresses 6,d my concerns a,d , though I d still think the d readability of the paper 7 needs improvement 5 . d I increased the overall score. ——— My main concern is that the main algorithm ( Sec b 4 f ) and the privacy analysis are not clearly stated, making it hard to check the quality of the paper. More 3 specifically, In terms of clarity 2 , - Algorithm 1 f and 2 needs 3,5 to 5 be 3,5 better 3 explained 1,c in c the text. The pseudo code c can also be made 5 more 3,5 readable 1,5 , 1 for example, Line c 4 c and 7 are actually if- c then-else statements but the else parts c are implicit ( the c full version is better). - PTERR e IFICBALL should c be explained 4 more . c It is unclear how much c noise e is added. - c It seems g to me g Algorithm 1 would find c k e clusters. Yet the c recursive part calls c it twice with k-1 . c Doesn’t that mean we would have 2(k-1) clusters? - Line 3 c,e in Algorithm 2 mentioned c a c,e subroutine P e GE which is not defined. PGE seems g to be important as it is the main component in Algorithm 2 that c guarantees Differential Privacy.	Toward AI 1 $+0.16$ words related to requests for improved clarity or explanation in academic reviews 2 $+0.06$ clarity clarification and related headings in academic review forms 3 $+0.05$ academic peer review feedback requesting more clarity or examples 4 $+0.04$ words related to research analysis and critique like elaborated explored or clarified 5 $+0.03$ criticism of document clarity or need for simplification 6 $+0.02$ academic paper verbs introducing research contributions 7 $+0.02$ academic evaluation of research papers focusing on methodology impact and practical applicability + 3 more $+0.05$ Toward human a $-0.36$ first person singular pronouns referring to the author or reviewer b $-0.36$ figures and tables in technical documents c $-0.34$ academic peer review feedback focusing on mathematical notation equation parts and specific paper reference structure d $-0.27$ post-rebuttal review comments and feedback e $-0.26$ mathematical notation and variables in technical papers f $-0.24$ referencing sections or specific numbers within technical or academic papers g $-0.20$ forms of the word seem including seem seemingly seems + 31 more $-1.99$	$\Sigma$ AI $+0.42$ $\Sigma$ human $-4.02$ bias $-0.51$ logit $-4.11$ $P(\mathrm{AI})$ $2\%$ $\rightarrow$ Human

Appendix

Table 18: The FT-SAE (

\tau=0.85

) logistic-regression detector applied to a human-written review from the PRDB test set ( NeurIPS2019-16-af0539fdca ). Tokens on which a listed feature fires are shaded red (feature pushes toward AI) or blue (toward human), with three intensities set by the token-level contribution

|w_{i}a_{i,t}|

of the strongest listed feature on that token.

Input text	Active features (Logit $\Delta$ )	Score
(Human) The reason is, stores want customers to use cash. By giving us cash, we are more likely to use cash next time. I feel a little guilty when using my bank card at the store because I know 5 I’m giving 1 about 2-3% of the sale to the bank. Unless I don’t really like where I’m shopping (ie Walmart), I c try to use cash if I have it. I doubt 5 these large stores pay e extra 3,e for 3 supplying the 1 cash portion. They just e need to keep the cash onhand. In other 6,a countries, do they e not mind paying banks a percentage of each transaction? That’s a huge f loss for retailers. ( I 2,d also 2,4 heard tipping e isn’ t popular in e some countries, maybe the lack of e regard for vendors e is g related somehow ??) 7 Oh , a plus, it ’ e s a e value added 3,a service e . A customer is more likely to return b to a store if they 5 provide 1,e this 1,5,e service e .	Toward AI 1 $+0.12$ forms of the verb provide or related synonyms 2 $+0.11$ introductory phrases used to preface an additional observation or note 3 $+0.08$ words indicating an increase or supplement such as addition additional and extra 4 $+0.04$ word also as a transition or additive marker 5 $+0.03$ words or phrases used in contemplative or introspective personal reflections 6 $+0.03$ word such used to introduce examples 7 $+0.03$ exclamation points at the end of sentences or phrases + 12 more $+0.22$ Toward human a $-0.20$ conversational meta-commentary phrases like off topic or similar topic b $-0.20$ refund and return policies or processes c $-0.15$ words or phrases commonly found in customer reviews or informal first-person narratives d $-0.12$ words or phrases indicating imprecision or approximate estimations e $-0.11$ words indicating service quality or restaurant experiences including staff interactions and feelings about service f $-0.11$ repeated capitalized or emphatic adverbs used for emphasis g $-0.11$ insulting or dismissive colloquial language directed at a person or their ideas + 31 more $-0.89$	$\Sigma$ AI $+0.67$ $\Sigma$ human $-1.89$ bias $+0.66$ logit $-0.58$ $P(\mathrm{AI})$ $36\%$ $\rightarrow$ Human

Appendix

Table 19: The Base-SAE (

\tau=0.85

) logistic-regression detector applied to a human-written document from the COLING test set ( coling53434_CUDRT ). Tokens on which a listed feature fires are shaded red (feature pushes toward AI) or blue (toward human), with three intensities set by the token-level contribution

|w_{i}a_{i,t}|

of the strongest listed feature on that token.

Input text	Active features (Logit $\Delta$ )	Score
(Human) This paper proposes an out g -of- distribution 3 (OOD) detection 3 method, where the features of in- distribution 3 (ID) samples 3,6 and those of the OOD samples are decor related 3 in a class- wise 3 manner by using HSIC. A theoretical guarantee for decorrelation is provided b,c,e . The prop b posed method, Conditional-i, is evaluated on a O 5 OD 5,e detection a,d benchmarks in image a classification and NLP, and shows S OTA d results. Reviewers 1,a had many major concerns, e.g., novelty from HOOD, missing a baselines with a OOD 5,a exposure a,b , hyperparameter tuning, and insufficient theory a to support the proposed method a , which the a authors 4 addressed well, and most of the a reviewers a,b have a been 2,7,a convinced a,c .␣ $\hookleftarrow$ The paper is well- a written a,b,g and the mathematical notation seems fine for machine learners.␣ However, the a relation a,b between Theorem␣ 1 and Assumption a 1 is weird.␣ Any theorem must always hold (otherwise it’s not a a theorem ), and the a sentence a,b “ a Theorem a,f ␣ 1 holds if Assumption␣ 1 holds” doesn’t a make sense.	Toward AI 1 $+0.49$ academic and technical publication identifiers including urls and scholarly citation elements 2 $+0.17$ past participle of to be 3 $+0.10$ mathematical and computer science technical terminology and formal definitions 4 $+0.09$ words related to authorship and citation such as author and authors 5 $+0.02$ proper names and capitalized personal identifiers in narrative or academic text 6 $+0.02$ words related to news announcements corporate updates and official company statements 7 $+0.01$ past participle been + 1 more $+0.03$ Toward human a $-1.26$ academic peer review feedback context b $-0.88$ word article when referring to specific texts or media content c $-0.30$ past participle verbs indicating future availability or state change such as released or available d $-0.22$ words ending in logist or surgeon related to medical professions e $-0.22$ narrative fiction character actions and pronouns f $-0.07$ abbreviations for change my view or reddit forum markers g $-0.06$ words and phrases describing restaurant and service quality experiences + 2 more $-0.89$	$\Sigma$ AI $+0.94$ $\Sigma$ human $-3.90$ bias $+0.53$ logit $-2.43$ $P(\mathrm{AI})$ $8\%$ $\rightarrow$ Human

Appendix

Table 20: The FT-Model+SAE (

\tau=0.85

) logistic-regression detector applied to a human-written document from the COLING test set ( coling12300_peersum ). Tokens on which a listed feature fires are shaded red (feature pushes toward AI) or blue (toward human), with three intensities set by the token-level contribution

|w_{i}a_{i,t}|

of the strongest listed feature on that token.

Input text	Active features (Logit $\Delta$ )	Score
(Llama) The paper presents a novel approach for verifying the local robustness 7 of networks with piecewise linear activation functions, which relies on geometric projections 7 rather 7,c than constraint solving or conservative over-approximations 7 . The approach is particularly suited for certifying $\ell$ 2-robustness and 7 has been a,b shown to outperform existing 7 exact tools by multiple orders of magnitude 7 while maintaining 7 the same or 7 better 7,d precision under 7 a time constraint 7 . The authors provide 6 a thorough 1 evaluation of their approach, including comparisons with other tools and a discussion of its limitations 5 . The paper is well- written 2 and the presentation is clear and concise. The authors have also made their implementation available on GitHub, which is a significant contribution to the research community. The only potential 4 drawback of the approach is its incompleteness, which may lead to unknown results in some cases. However, the authors provide 6 a thorough 1 discussion of this limitation 5 and its implications. Overall, the paper is a significant contribution to the field 3 of neural network verification and is likely to be of interest to researchers and practitioners in this area.	Toward AI 1 $+0.99$ the word thorough 2 $+0.63$ word written used to describe paper quality 3 $+0.50$ academic field noun when used in singular or plural form 4 $+0.45$ activates on the word potential 5 $+0.32$ word limitations within academic paper reviews 6 $+0.23$ forms of the verb provide in academic paper reviews 7 $+0.22$ academic paper method descriptions and technical approach summaries + 5 more $+0.37$ Toward human a $-0.16$ past participle been often following have or had b $-0.11$ the word been in passive voice constructions c $-0.05$ adversative transition words like rather and instead d $-0.03$ word better or best used to compare performance or quality + 2 more $-0.08$	$\Sigma$ AI $+3.71$ $\Sigma$ human $-0.43$ bias $-0.37$ logit $+2.91$ $P(\mathrm{AI})$ $95\%$ $\rightarrow$ AI

Appendix

Table 21: The FT-Model+SAE (Llama) (

\tau=0.85

) logistic-regression detector applied to a Llama-written review from the PRDB test set ( ICLR2021-1934-6POeD7QDvXS ). Tokens on which a listed feature fires are shaded red (feature pushes toward AI) or blue (toward human), with three intensities set by the token-level contribution

|w_{i}a_{i,t}|

of the strongest listed feature on that token.

	COLING (dev)			COLING (devtest)			COLING (test)			RAID (test)
	AUC	F1	TPR@ $5$	AUC	F1	TPR@ $5$	AUC	F1	TPR@ $5$	AUC	F1	TPR@ $5$
gram2vec	81.9	77.9	36.2	80.1	73.1	45.3	77.7	72.0	33.6	77.5	73.5	43.6
+XGBoost	93.1	87.1	69.2	82.0	75.1	55.0	81.5	70.1	53.8	83.5	82.1	47.7
w/o Mask
Base-SAE	97.1	91.4	84.0	96.2	93.2	90.4	94.2	88.9	81.8	95.2	92.3	86.4
+XGBoost	98.9	94.8	92.8	95.7	91.6	87.8	92.9	86.3	79.1	96.3	93.3	87.8

Appendix

Table 22: Performance comparison between Logistic regression and XGBoost detector heads trained on the COLING train set, evaluated on the held-out COLING and RAID splits. We see that XGBoost improves performance of gram2vec by a large margin. It also improves performance of our models but does so by a comparatively smaller amount. This suggests that our features are likely predominantly linear with relatively negligible interaction terms.

Document-Level Paraphrase Ablation Results
		Baseline			Few Shot			Guided			SAE			SAE $\times 3$
	F1	F1	ASR	Lev	F1	ASR	Lev	F1	ASR	Lev	F1	ASR	Lev	F1	ASR	Lev
COLING
Base-SAE	80.3	81.1	5.3	24.0	78.1	10.7	31.5	77.2	12.6	39.0	66.8	28.1	26.1	46.9	55.1	34.5
FT-SAE	84.4	83.3	6.3	24.0	82.0	8.8	31.5	80.9	11.1	39.0	70.5	26.9	28.7	51.7	52.6	37.5
FT-Model+SAE	88.3	86.0	8.7	24.0	82.9	14.7	31.5	79.5	20.7	39.0	69.9	34.2	26.5	45.3	63.2	33.7

Appendix

Table 23: Adversarial paraphrasing of 2500 AI-written documents per dataset (2500 human documents fixed). The first column is each detector’s F1 before the attack. For every intervention we report F1 after paraphrasing, the attack success rate (ASR: share of documents detected before paraphrasing that are classified human afterwards) and the mean word-level Levenshtein distance to the original as a share of the longer text (Lev), all in percent.

Split	#Docs	#Gen.	Generators (families)	Domains
COLING
train	610,767	39	GPT-3.5, GPT-3 (davinci) (3), OPT (9), Llama (6), T0/BLOOM (4), FLAN-T5 (5), Cohere, Gemma (2), GPT-4/4o (2), Mistral, Dolly (2), GPT-J/NeoX (2), GLM	Q&A, Reddit, stories, WikiHow, arXiv abstracts, Wikipedia, student essays, Yelp reviews, news, commonsense, opinion (CMV), scientific tables, peer reviews
dev	261,758	39	GPT-3.5, GPT-3 (davinci) (3), OPT (9), Llama (6), T0/BLOOM (4), FLAN-T5 (5), Cohere, Gemma (2), GPT-4/4o (2), Mistral, Dolly (2), GPT-J/NeoX (2), GLM	Q&A, Reddit, stories, WikiHow, arXiv abstracts, Wikipedia, student essays, Yelp reviews, news, commonsense, opinion (CMV), scientific tables, peer reviews
devtest	32,557	5	GPT-4/4o, Gemini, Gemini (humanized), Mistral, Mistral (humanized)	arXiv abstracts, Wikipedia, WikiHow, Reddit, student essays, peer reviews, books, news, poetry, recipes, reddit, reviews, wiki
test	73,941	14	GPT-4/4o (4), Llama (3), GPT-4/Llama-2 (mixed), GLM, GPT-3.5, Qwen, Baichuan, Dolly, StableLM	IELTS essays, editing (expand, polish, refine, rewrite, summarise), peer-review summaries, peer reviews, mixed human–AI text (news, blog, email, speech, reviews, abstracts, QA)
PRDB

Appendix

Table 24: Composition of the datasets used for training and evaluation. COLING train and dev are a stratified split of one pool (MAGE, M4GT-Bench and HC3). RAID is the non-adversarial portion of its test split. We see that all of our splits cover a wide spread of models and domains.

Panel of Figure 10 — Figure 10: Interface for our interpretable AI detector. Top features for the given text are displayed on the right bar along with their description, autointerp score, local importance, and odds ratios. Text in the input is highlighted in proportion to feature strength. Hovering over highlighted tokens provides a more detailed description of the feature along with auto-interpretation score and number of occurrences in human vs. AI text.

	GPT-4o			Llama 3.1 70B			Qwen 2.5 72B			Claude 3.5 Sonnet			Gemini 1.5 Pro
	AUC	F1	TPR@ $5$	AUC	F1	TPR@ $5$	AUC	F1	TPR@ $5$	AUC	F1	TPR@ $5$	AUC	F1	TPR@ $5$
Trained on GPT-4o
Baselines
gram2vec	100	99.6	100	97.9	93.5	90.0	99.7	98.6	99.9	99.7	98.2	99.7	99.1	96.6	98.1
hf-gemma2	100	100	100	97.4	92.3	89.9	100	99.6	100	99.2	96.3	96.9	99.1	96.1	96.5
w/o Mask

Appendix

Table 25: Model performance across LLM-generated text targets in the PRDB dataset, for detectors trained on GPT-4o- and LLaMA-generated text respectively. Each column group corresponds to text generated by a different source model.

\dagger

indicates off-the-shelf models that were not trained on any in-domain data.

Figure 20: Recall at FPR=5% (y-axis) vs. Interpretability Threshold (x-axis) for

L_{1}

-regularized logistic regression classifiers trained on the COLING dataset evaluated on various splits of the RAID dataset Dugan et al. (2024) .

Most Predictive Features: FT-SAE $\tau>0.85$ — General Domain Text (COLING)
Feature	Concept Description (auto-interp: gemini-3.1-flash-lite)	Corr.	Logit $\Delta$
AI Features
13_3738	introductory phrases or filler text commonly used by ai language models when declining personal questions ( e.g., “As an”, “AI language model”, “I do not have”)	0.859	$+0.987$
13_1545	academic and scientific research methodology and technical analysis terminology ( e.g., “dynamics”, “analytical”, “methodology”)	0.876	$+0.909$
17_12178	critical commentary or argumentative analysis of student policies and social responsibility ( e.g., “have”, “effects”, “motivated by”, “can”, “unfair”)	0.892	$+0.558$
17_8638	neuron activates on random text fragments and noisy character sequences often found in garbled or fragmented data ( e.g., “astion”, “Mol”, “a priori”)	0.867	$+0.372$

Appendix

Table 26: Top features of the FT-SAE

\tau>0.85

logistic-regression detector on the COLING dataset Wang et al. (2025) , ranked by mean logit contribution (Logit

\Delta

) along with their auto-interpretation simulation score (§ 4.2 ).

Most Predictive Features: FT-Model+SAE $\tau>0.85$ — General Domain Text (COLING)
Feature	Concept Description (auto-interp: gemini-3.1-flash-lite)	Corr.	Logit $\Delta$
AI Features
7_8935	the word the when appearing frequently	0.865	$+0.865$
7_2327	the preposition on	0.855	$+0.860$
7_7227	actions and physical reactions of characters in narrative prose ( e.g., “laughed nervously”, “shook his head”, “stood”)	0.887	$+0.796$
13_4399	imperative instructional verbs in written guides or manuals ( e.g., “Remember”, “Perform”, “Write”)	0.861	$+0.642$

Appendix

Table 27: Top features of the FT-Model+SAE

\tau>0.85

logistic-regression detector on the COLING dataset Wang et al. (2025) , ranked by mean logit contribution (Logit

\Delta

) along with their auto-interpretation score (§ 4.2 ).

Most Predictive Features: Base-SAE $\tau>0.85$ — General Domain Text (COLING)
Feature	Concept Description (auto-interp: gemini-3.1-flash-lite)	Corr.	Logit $\Delta$
AI Features
13_1157	code syntax punctuation like brackets braces and delimiters ( e.g., } , ) , ] )	0.863	$+0.604$
17_941	html input tag type attribute syntax ( e.g., “type”, type= , type=" )	0.859	$+0.558$
13_1449	conclusion or concluding remarks headers ( e.g., “Conclusion”, “conclusion”, “conclusion.”)	0.864	$+0.543$
22_6940	software license and copyright disclaimers in source code headers ( e.g., “licensed under”, “WITHOUT ANY WARRANTY”, “Apache 2.0”)	0.903	$+0.522$

Appendix

Table 28: Top features of the Base-SAE

\tau>0.85

logistic-regression detector on the COLING dataset Wang et al. (2025) , ranked by mean logit contribution (Logit

\Delta

) along with their auto-interpretation simulation score (§ 4.2 ).

Most Predictive Features: Base-SAE (GPT-4o) $\tau>0.85$ — AI-written Paper Reviews
Feature	Concept Description (auto-interp: gemini-3.1-flash-lite)	Corr.	Logit $\Delta$
AI Features (GPT-4o)
17_13662	single quotation marks surrounding titles or specific terms ( e.g., ’ , ‘ , ’ )	0.853	$+0.525$
17_10356	modal verb might appearing in various contexts ( e.g., “might”, “might be”, “might enhance”)	0.866	$+0.286$
7_8231	words indicating evidence or confirmation such as supported and weaknesses	0.851	$+0.278$
7_15224	words related to offer such as offered offers or names containing ofer	0.857	$+0.272$

Appendix

Table 29: Top features of the Base-SAE (GPT-4o)

\tau>0.85

logistic-regression detector on the PRDB dataset Yu et al. (2026) , ranked by mean logit contribution (Logit

\Delta

) along with their auto-interpretation score (§ 4.2 ).

Most Predictive Features: Base-SAE (Llama) $\tau>0.85$ — AI-written Paper Reviews
Feature	Concept Description (auto-interp: gemini-3.1-flash-lite)	Corr.	Logit $\Delta$
AI Features (Llama 3.1 70B)
17_8515	punctuation and special characters in url paths or academic citations ( e.g., . , “/paper”, ? , = , & )	0.885	$+0.806$
7_6754	word also when used as a transition in academic writing	0.865	$+0.428$
7_156	word field occurring in academic or technical research contexts ( e.g., “field”, “fields”, “feld”)	0.866	$+0.296$
13_9022	underscore character used as a word separator or within code variable names ( e.g., _ , “best_split”, “x_0”)	0.855	$+0.253$

Appendix

Table 30: Top features of the Base-SAE (Llama)

\tau>0.85

logistic-regression detector on the PRDB dataset Yu et al. (2026) , ranked by mean logit contribution (Logit

\Delta

) along with their auto-interpretation score (§ 4.2 ).

Most Predictive Features: FT-Model+SAE (GPT-4o) $\tau>0.85$ — AI-written Paper Reviews
Feature	Concept Description (auto-interp: gemini-3.1-flash-lite)	Corr.	Logit $\Delta$
AI Features (GPT-4o)
13_15950	academic paper review questions asking for more information details or clarification ( e.g., “provide additional”, “provide insights”, “provide a mathematical framework”)	0.866	$+1.002$
17_8459	academic review feedback or peer review decisions ( e.g., “acceptance”, “reject”, “accepting”)	0.870	$+0.814$
22_4800	double colons following bolded markdown section headers ( e.g., Significance: , Quality: , **Questions: )	0.860	$+0.672$
17_12216	activates on the word potential	0.884	$+0.515$

Appendix

Table 31: Top features of the FT-Model+SAE (GPT-4o)

\tau>0.85

logistic-regression detector on the PRDB dataset Yu et al. (2026) , ranked by mean logit contribution (Logit

\Delta

) along with their auto-interpretation simulation score (§ 4.2 ).

Most Predictive Features: FT-Model+SAE (Llama) $\tau>0.85$ — AI-written Paper Reviews
Feature	Concept Description (auto-interp: gemini-3.1-flash-lite)	Corr.	Logit $\Delta$
AI Features (Llama 3.1 70B)
17_4373	the word thorough	0.892	$+0.993$
13_12061	numbered list items in parentheses within question sections of research papers ( e.g., (1) , (2) , 1. )	0.858	$+0.898$
7_5460	the word one when it appears at the start of a sentence or bullet point	0.851	$+0.745$
13_164	opening square bracket in list or string literals ( e.g., [’ , [ , << )	0.866	$+0.610$

Appendix

Table 32: Top features of the FT-Model+SAE (Llama)

\tau>0.85

logistic-regression detector on the PRDB dataset Yu et al. (2026) , ranked by mean logit contribution (Logit

\Delta

) along with their auto-interpretation score (§ 4.2 ).

Most Predictive Features: FT-SAE (GPT-4o) $\tau>0.85$ — AI-written Paper Reviews
Feature	Concept Description (auto-interp: gemini-3.1-flash-lite)	Corr.	Logit $\Delta$
AI Features (GPT-4o)
13_1632	modal verbs beginning questions in academic paper reviews ( e.g., “Can”, “Could”, “can”)	0.871	$+0.370$
17_7486	word environments appearing in technical literature regarding system deployment or model applicability ( e.g., “environments”, “domains”, “architectures”)	0.869	$+0.345$
22_14556	methodology and methodologies in academic research contexts ( e.g., “methodology”, “methodologies”, “methodological”)	0.903	$+0.258$
13_3382	activates on the word clarity when used in academic or technical peer review critiques	0.859	$+0.240$

Appendix

Table 33: Top features of the FT-SAE (GPT-4o)

\tau>0.85

logistic-regression detector on the PRDB dataset Yu et al. (2026) , ranked by mean logit contribution (Logit

\Delta

) along with their auto-interpretation score (§ 4.2 ).

Most Predictive Features: FT-SAE (Llama) $\tau>0.85$ — AI-written Peer Reviews
Feature	Concept Description (auto-interp: gemini-3.1-flash-lite)	Corr.	Logit $\Delta$
AI Features (Llama 3.1 70B)
17_5027	academic review phrases like provide a clear or provide a nuanced ( e.g., “provide a clear”, “provide a concise”, “provide a nuanced”)	0.893	$+0.583$
13_13076	introductory phrases for questions in academic paper reviews ( e.g., “include:”, “potential questions for”, “is”)	0.880	$+0.497$
7_13926	thorough analysis of phrase ( e.g., “thorough analysis of”, “analysis of”, “evaluation of”)	0.886	$+0.464$
13_3534	initial transition words like firstly and introductory adverbial clauses ( e.g., “Firstly”, “First”, “The”)	0.875	$+0.452$

Appendix

Table 34: Top features of the FT-SAE (Llama)

\tau>0.85

logistic-regression detector on the PRDB dataset Yu et al. (2026) , ranked by mean logit contribution (Logit

\Delta

) along with their auto-interpretation simulation score (§ 4.2 ).

AI-Generated Text Detection

AI-generated text detection develops methods to distinguish human-written, machine-generated, and AI-edited text.

See all 170 papers

May 24, 2026

READER: Reasoning-Enhanced AI-Generated Text Detection

Recent advances in large language models (LLMs) have made it increasingly difficult to distinguish human-written text from AI-generated content. Many existing detectors train supervised neural classifiers that achieve strong in-distribution performance but are often opaque and can degrade substantially under distribution shift. We present READER, a reasoning-enhanced AI text detector that outputs both a human/AI label and a structured rationale describing the evidence for its decision. A key component of our approach is READ, a curated supervision set of rationales and verdicts. We fine-tune an LLM on READ to build READER, which reasons before detecting at inference time. Despite having only 1.5B parameters, READER consistently outperforms existing detectors as well as prompted, high-capacity LLM baselines (GPT-5.2, Gemini-3-Pro, and DeepSeek-V3.2), which are 100 to 1000 times larger in scale.

Pingfan Su, Kai Ye, Shijin Gong +4

Department of Statistics, London School of Economics and Political Science · School of Management, University of Science and Technology of China · School of Mathematics, University of Birmingham

AI-Generated Text Detection Large Language Model-Based Text Classification

Ask PDF

May 27, 2026

Show, Don't TELL: Explainable AI-Generated Text Detection

Research on AI-generated text detection has presented a number of approaches to discern human from AI prose, some of which achieving high in-distribution performance. However, real-world applicability has stalled because their outputs are misaligned with the needs of users, such as professors, who are presented with a numeric score that has no attached explanation. We tackle this issue with a novel architecture, TELL, that bakes explainability from the ground-up. While our system still offers a numerical score like other detectors for comparability, TELL takes a fundamentally different approach where we aim to show the user the "tells" by which the model believes a text is AI or human-written, to empower the user to decide who wrote a text using their own judgment and understanding of the context of the writing and its alleged author. We train TELL on a custom SFT dataset of domain-specific authorship annotations, and further refine the system using GRPO with curriculum learning to improve performance. We achieve competitive performance with state-of-the-art detectors (AUROC 0.927) while natively providing annotations that explain the basis for the detector's decision. We further evaluate the quality of our explanations using a dataset of human annotations and report a high (mean 72.3%) win-rate on annotation concreteness, falsifiability, coherence, plausibility and grounding, allowing users to critically think and decide for themselves. Our work thus reframes the problem of AI-generated text detection in a human-centric perspective and paves the way for a new family of detectors that focus on native explainability.

Aldan Creo, Suraj Ranganath

School of Computing, Information and Data Sciences University of California, San Diego United States of America

Explainable Artificial Intelligence AI-Generated Text Detection Human Oversight of AI Explainability Evaluation Explainable Text Classification

Ask PDF

Mar 5, 2026

NOTAI.AI: Explainable Detection of Machine-Generated Text via Curvature and Feature Attribution

We present NotAI.AI, an explainable AI-generated text detection system. Instead of returning only a binary label or confidence score, the system shows which signals influenced the prediction and lets users inspect an attribution-based sensitivity estimate obtained by subtracting selected local contributions. NotAI.AI combines sentence-level conditional probability curvature, a neural detector score, and interpretable stylometric and readability features in an XGBoost meta-classifier. It explains predictions with TreeSHAP feature contributions and can turn the resulting evidence into a concise natural-language explanation. We evaluate the system on a category-balanced subset of RAID containing human-written, clean AI-generated, and attacked AI-generated texts. The full model outperforms variants based on individual feature families, reaching 0.9685 F1 on the held-out within-subset test split. In an automatic evaluation, two model judges rate 94.5-98.6% of generated explanations as faithful to the supplied detector evidence. The web interface (https://notai-ai.vercel.app), source code (https://github.com/Oleksandr-MB/EMNLP2026DEMO_NotAI.AI), and demonstration video(https://youtu.be/l8Nk8kdBTHQ) are publicly available.

Oleksandr Marchenko Breneur, Adelaide Danilov, Aria Nourbakhsh +1

Department of Computer Science, Faculty of Science, Technology and Medicine, University of Luxembourg

AI-Generated Text Detection Explainable Text Classification Explainable Artificial Intelligence Robustness of Machine-Generated Text Detection

Ask PDF

Abstract

AI-Generated Text Detection

Explore similar work

READER: Reasoning-Enhanced AI-Generated Text Detection

Show, Don't TELL: Explainable AI-Generated Text Detection

NOTAI.AI: Explainable Detection of Machine-Generated Text via Curvature and Feature Attribution