AEyeDE: An Attention-Based Attribution Framework for AI-Generated Text Detection

Aria Nourbakhsh; Adelaide Danilov; Christoph Schommer; Salima Lamsiyah

cs.CLApr 13, 2026

AEyeDE: An Attention-Based Attribution Framework for AI-Generated Text Detection

Authors: Aria Nourbakhsh, Adelaide Danilov, Christoph Schommer, Salima Lamsiyah

Organizations: Department of Computer Science, University of Luxembourg

Abstract

Detecting AI-generated text is becoming increasingly challenging as modern language models approach human-level fluency and can evade detectors that rely on surface statistics or likelihood-based signals. We propose AEyeDE, an attribution-driven approach to human-AI authorship detection that leverages model attention as a discriminative signal. Specifically, we extract attention-based attribution matrices for both human- and AI-generated text using a proxy Transformer model with white-box access and train a lightweight Convolutional Neural Network to learn representations from these attribution maps. Across encoder-decoder translation settings, our method consistently outperforms a text-only baseline. In decoder-only settings, it performs strongly in generator-specific detection, remains competitive on standard benchmarks, and shows robustness under cross-dataset transfer and alternative-spelling perturbations. We further show that attention maps exhibit recurring local structures whose relative frequencies differ consistently between human- and AI-generated text across datasets and proxy models. These findings suggest that attention-based attribution maps provide a complementary and interpretable signal for AI-generated text detection. We will make the code publicly available to support future research.

Figures & tables

Figure 1: Overview of AEyeDE . Given a text sample and white-box access to a proxy generator model $G_{\theta}$ , we extract an attention-derived attribution matrix $A$ (top left). For decoder-only models, we summarize $A$ by sampling fixed-size square blocks (e.g., $128\times 128$ ) along the main diagonal, where the strongest local token-token interactions are concentrated (orange boxes) ( Xiao et al., 2023 ; Ivanitskiy et al., ; Qi et al., 2025 ) ; for encoder-decoder models, we use the full cross-attention map. Each block is encoded by a CNN (bottom), producing one embedding per block. These embeddings are aggregated with learnable attention pooling to form a global attribution representation, which is passed to a lightweight MLP classifier to predict human versus LLM authorship (top right). In our implementation, the final low-resolution CNN feature map is $16\times 16$ ; its spatial cells correspond to regions of the original attribution block and are later used for motif clustering.

	ar-en					de-en					fr-en
Config	Acc	Prec	Rec	F1	AUC	Acc	Prec	Rec	F1	AUC	Acc	Prec	Rec	F1	AUC
AEyeDE(CNN)	68.9	63.0	91.4	74.6	77.7	79.0	72.6	93.0	81.5	87.3	83.6	78.0	93.6	85.1	91.0
AEyeDE(CNN+text)	71.9	65.7	91.9	76.6	81.5	81.1	74.9	93.6	83.2	89.2	85.4	80.3	93.9	86.5	92.7
text (baseline)	61.9	57.3	93.4	71.0	71.5	69.5	63.7	90.5	74.8	77.5	72.4	66.2	91.4	76.8	81.4

Table 1: Performance on Marian-MT generated translation and attribution maps. CNN and CNN+text configurations are based on AEyeDE .

	Cohere						GPT-neo
	Acc	Prec	Rec	F1	AUC	TPR-FPR	Acc	Prec	Rec	F1	AUC	TPR-FPR
AEyeDE	97.23	97.48	98.43	97.95	99.51	95.22	97.22	97.20	98.73	97.96	99.71	95.89
Binoculars	67.94	67.91	99.48	80.72	78.10	34.83	67.61	67.63	99.78	80.62	25.84	0.45
Curvature	83.27	86.29	89.39	87.81	90.86	47.01	78.91	98.22	70.03	81.76	79.70	65.47
GLTR	67.39	67.42	99.93	80.52	76.68	10.99	67.96	67.92	99.55	80.75	71.54	44.99
LogRank	67.39	67.42	99.93	80.52	75.49	10.24	69.58	69.31	98.58	81.39	73.95	36.70

Table 2: Results on individual arrangement where each human and AI-generated attention attributions were extracted from the same model specified in the dataset to train the AEyeDE model.

	Cohere						GPT-neo
	Acc	Prec	Rec	F1	AUC	TPR-FPR	Acc	Prec	Rec	F1	AUC	TPR-FPR
AEyeDE	72.38	90.98	65.55	76.19	80.67	58.67	72.25	96.57	61.06	74.82	78.86	58.97
Binoculars	67.29	67.57	99.03	80.33	74.21	49.33	67.46	67.51	99.85	80.55	74.43	43.57
Curvature	57.41	79.03	50.15	61.36	64.00	20.18	70.13	93.37	60.01	73.07	77.82	52.84
GLTR	67.44	67.44	100.00	80.55	76.32	26.53	67.51	68.11	97.53	80.21	73.42	15.10
LogRank	67.44	67.44	100.00	80.55	75.14	24.51	68.06	69.44	94.10	79.91	73.20	16.52

Table 3: Results on the unified arrangement where the samples corresponding to the Mistral model were excluded from the training of AEyeDE .

	GPT-neo
	Acc	Prec	Rec	F1	AUC	TPR-FPR
AEyeDE	96.85	95.68	98.27	96.96	99.51	93.50
Curvature	98.40	99.83	97.04	98.42	99.22	97.45
GLTR	98.25	98.89	97.60	98.24	99.17	97.60
LogRank	97.95	99.69	96.20	97.91	99.03	97.50
Rank	83.40	81.63	86.20	83.85	91.22	36.60

Table 4: Results on HC3 using GPT-neo to extract attention attributions.

Table 6

Table 7

Method	Mistral
	Acc	Prec	Rec	F1	AUC	TPR
AEyeDE	66.36	62.16	83.64	71.32	74.43	14.55
Binoculars	51.36	50.71	96.36	66.45	63.16	6.36
Curvature	49.44	49.74	98.73	66.16	50.47	9.76
GLTR	51.40	50.74	98.88	67.07	74.38	15.46
LogRank	51.27	50.68	98.47	66.92	74.25	13.01

Table 9: Cross-dataset evaluation on Beemo using Mistral as the proxy model.

Panel of Figure 2 — Figure 2: Examples of the top motif cluster (by absolute mean prevalence gap $\Delta\bar{r}$ between gold-machine and gold-human) for each dataset/proxy-model configuration. Each panel shows representative $8{\times}8$ patches (z-normalized) from the corresponding cluster.

Dataset	Setting	Proxy-model	cid	Human $\bar{r}_{g}$	Machine $\bar{r}_{g}$	$\Delta r$	95% CI	$p$
HC3	individual	GPT-neo 2.7B	24	9.7	20.7	+11.0	[ 9.3, 12.7]	$<\!0.001$
RAID	unified	Cohere Cmd-R 7B	13	80.0	69.5	-10.5	[-12.9, -8.2]	$<\!0.001$
RAID	individual	Mistral 7B	23	63.5	54.8	-8.7	[-11.1, -6.2]	$<\!0.001$
RAID	individual	GPT-neo 2.7B	4	87.6	81.1	-6.5	[ -9.0, -4.1]	$<\!0.001$
RAID	individual	Llama 3.1 8B	12	7.5	3.1	-4.4	[ -4.8, -3.9]	$<\!0.001$
RAID	unified	GPT-neo 2.7B	37	11.6	8.5	-3.0	[ -4.1, -2.0]	$<\!0.001$

Table 10: Top-1 motif cluster by absolute class-rate gap

|\Delta r(c)|

for each dataset/setting. Rates are mean per-sample motif shares (%), and

\Delta r

is reported in percentage points (machine minus human). Confidence intervals are 95% bootstrap CIs over samples;

p

is from a label-permutation test.

AI-Generated Text Detection

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

See all 167 papers

Show, Don't TELL: Explainable AI-Generated Text Detection
May 27, 2026Aldan Creo, Suraj RanganathExplainable Artificial Intelligence AI-Generated Text Detection
School of Computing, Information and Data Sciences University of California, San Diego United States of America
Automatic detection of Gen-AI texts: A comparative framework of neural models
Mar 19, 2026Cristian Buttaro, Irene AmeriniAI-Generated Text Detection Robustness of Machine-Generated Text Detection
Sapienza University in Rome Latina
DeBERTa-Sentinel: Toward Transparent and Trustworthy Detection of AI-Generated Text
Aug 2, 2026Muhammad Yousaf Rehman, Muhammad IslamAI-Generated Text Detection
SPECS, University of Hertfordshire, Hatfield, UK · College of Science and Engineering, James Cook University, Cairns, 4878, QLD, Australia · Centre for AI and Data Science Innovation, James Cook University, Cairns, QLD 4878, Australia

May 27, 2026J/K move · Enter open · S save

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

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

Open paper AskPDF arXiv

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.

Abstract

AI-Generated Text Detection

Explore similar work

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

Automatic detection of Gen-AI texts: A comparative framework of neural models

DeBERTa-Sentinel: Toward Transparent and Trustworthy Detection of AI-Generated Text