Papers for
content moderators
Papers whose findings have a practical use for this group, as judged from the abstract. Open a paper to read what it means in practice.
AI generated content does not reduce trust in gender biased statements
Evaluating Human Sensitivity to Gender Biased AI-Generated Content through Human Assessment of Google's Autocomplete Function
Abstract: Large Language Models (LLMs) are increasingly present, but their effects on human users remain underexplored. We tested whether AI-attributed information influences views regarding gender-related statements, a domain where biases are pervasive. Spanish-speaking participants recruited through Prolific evaluated biased statements derived from Google's autocomplete function and human-authored, non-biased gender-related facts. Participants were either informed that some of the material was AI-generated or received no information about its origin. AI attribution did not affect reaction times or agreement judgments. This result suggests that AI-generated content does not automatically raise credibility concerns and may be evaluated with less scrutiny compared to human-produced content.
AI models show some country bias in social science essays on US and China
Can AI be a 'Researcher Without Borders'? Auditing Bias in AI-Assisted Social Science Research about China and the US
Abstract: Although large language models have demonstrated strong performance on algorithmic and mathematical tasks, their suitability as social science research assistants remains unclear, particularly because many social science questions lack agreed-upon, quantifiable benchmark answers. Evaluating this suitability is important since researchers and students increasingly rely on AI to generate research ideas. When questions concern the countries in which models were developed, bias may be exacerbated: Do AI models favour the countries where they were developed or the social theories associated with those countries' scholarly traditions? Such biases could distort research findings, undermine scientific credibility, and distort findings when they are disseminated to the public. To investigate whether models exhibit country-focused bias in social science research, we designed a benchmarking task requiring AI to write short essays on social science questions in relation to the US and China as well as some broader social science questions. We varied three conditions: question type (cross-country comparisons, country-specific questions, and general questions); model (two US-developed, two China-developed, and one Europe-developed model); and prompting modality (a base essay-writing request, a researcher identity, a US researcher identity, and a Chinese researcher identity). Responses were manually annotated for bias and source use by two researchers with experience in social science in relation to China and the US. Our analyses show that, on the whole, US and Chinese models and the European one are not politically biased but differ in the sources they cite. Still, some comparative questions favour one country over another, and this tendency is more pronounced for certain questions. These findings suggest that researchers and students should exercise caution when using AI for social science research.
Anybreak allows black-box jailbreak of text to image models for unsafe content
AnyBreak: Towards Jailbreaking Any T2I Model for Any Concept
Abstract: Text-to-image (T2I) models have demonstrated remarkable capabilities in image generation and editing, yet this progress has also introduced significant security risks. Prior work has shown that adversarial prompt can induce the generation of Not-Safe-For-Work (NSFW) content, including pornographic, violent, and graphic imagery. However, such attacks are typically mitigated by prompt filters and often rely on white-box assumptions. In this paper, we propose AnyBreak, a novel jailbreak framework that constructs adversarial images via multi-scale progressive semantic alignment. During the generation process, these adversarial images steer the generation trajectory in latent space towards unsafe data manifolds, enabling cross-modal transfer of malicious concepts from the text modality to the image modality. AnyBreak operates using only clean textual prompts and adversarial images at inference time, and can jailbreak any T2I models towards any malicious concepts in a fully black-box setting. We further introduce AnyBreak-V, an enhanced variant that incorporates a compensation term to bypass post-hoc safety checkers. Extensive experiments demonstrate that AnyBreak achieves high attack success rates while preserving visual fidelity and semantic consistency with the input prompts under black-box and clean-text constraints. Our findings reveal that existing T2I defense leave critical cross-modal attack surfaces insufficiently addressed. The code is available at https://github.com/luqi-glory/AnyBreak.
Models treat usable but unsupported facts as valid answers
Models Verify Sufficiency, Not Validity: Usable Falsehoods Evade the Checks that Gaps Trigger
Abstract: Ask a model "how many settlements did the Normans control?" and it answers that the passage does not say. Tell it "given that the Normans controlled 12 settlements of 400 inhabitants, what is the total?" and it computes 4,800. Nothing was added to the passage; the unsupported fact merely became usable. Benchmarks for unanswerable questions are built almost entirely by omission, deleting the fact a question needs and scoring refusal, so they never probe this. Holding source support fixed and varying only whether an unsupported premise is usable, flagging drops 8-30 points and commitment to an answer rises 1.5-2.3x on GSM8K across four models spanning two families and 2B-26B; all 16 paired contrasts remain significant after Holm correction. Two further manipulations reproduce it, one also on SQuAD: a fact asserted rather than asked about, and a corrupted intermediate value rather than a removed one. Models articulate the exact gap they then ignore, stating that "there is no financial information provided" for an inert premise and silently adopting a fabricated price for the same item. We read these as one behaviour: whether a model checks depends substantially on whether usable material is present, even when source support is held fixed. Because the main design never instructs models to reject premises introduced in the question, adopting such a premise can also be read as conditional reasoning, and we delimit what the design does and does not show. In the judging seat the pattern differs: asked whether the answer follows, judges pass solutions with a required step deleted 78-99% of the time at every scale tested. A natural prompt-level mitigation lowers accuracy by 22 points without improving detection.
Porn titles reveal stereotypes about race gender and nationality
Flattening Fantasies: Analyzing Stereotypes in Pornographic Discourse
Abstract: This study presents a large-scale quantitative analysis of 5.4 million video titles from xHamster, the world's fourth most visited pornography platform, to examine linguistic patterns used to describe different dimensions of identities (e.g., race and gender). Specifically, we investigate how nationalities, ethnic groups, and gender identities are portrayed in pornographic discourse, focusing on stereotypes and deviations from implicit norms of pornography. The results show the dominance of American pornography over other groups, which acts as the de facto norm for the adult industry. We identify persistent ethnic stereotypes reflecting colonial power dynamics, religious fetishization, and the sexualization of cultural taboos. The analysis reveals gendered representational differences across cultures, with a prevalent Madonna-Whore dichotomy in the portrayal of women and stereotypical depictions of men in physically dominant roles.
Measuring human input in AI writing is unreliable using likelihood scores
Likelihood Is Not Authorship: Can We Reliably Measure Human Contribution in AI-Assisted Writing?
Abstract: AI-text detection generally asks whether a machine wrote a text. As AI becomes an increasingly integral and accepted part of writing, the more relevant question is no longer whether AI was used, but how much of the text reflects human contribution. A prominent line of work estimates this by measuring how predictive a person's prompt is of the final text. We show that this measure can be easily manipulated. We introduce SPECIFICITY PADDING that simply asks the model to copy a few random IDs from the prompt, and that alone makes a title-only request look more human than one that supplies the article's key ideas. A context-swap audit then holds each document fixed, and only shows or hides the IDs to the evaluator. The score of the unchanged text swings by a third to half of its scale, and it rises for every document we test, on news and scientific abstracts, across varied LLMs from 3.8B to 753B parameters. The same articles can average 85\% human or 15\% human depending only on which prompt is on record, and even with no prompt at all, the output-only variant rewards the random strings themselves. An exact identity, the credit ledger, explains why. The score is a surprisal-weighted average over the text, so rare strings that the context makes copyable collect the credit, whatever they mean. A two-point bound marks a deeper limit. If the prompt could have been machine-written, every transcript-only method must, in the worst case, miss by half the contribution at stake. Likelihood-based scores alone cannot reliably establish human contribution without evidence of what the person added and how the text was produced.
Jev judges AI safety tasks better than LLMs but misses some phishing
Same Attack, Different Judge: A First Controlled Adversarial Comparison of Jev, LLM Judges, and Purpose-Built Safety Classifiers
Abstract: TypeSafe AI's Jev is a new kind of safety judge. It generates no text and no label: it answers questions the developer defines in advance and returns a probability for each answer. The vendor admits that adversarial input can change these answers but gives no numbers. In this paper, we present the first controlled adversarial evaluation of Jev. We build a test harness that sends every system the same payload under the same policy, and run it on four public-data tasks, from rule-based ones (input moderation, injection detection, routing) to a judgment-based one (phishing detection), with fifty attack variants from thirteen families. The targets are Jev, five GPT and GLM judges, and two local classifiers (Llama Guard 3, Prompt Guard 2). We find that Jev is the most robust system on input moderation: eight forgeries of its structured input bypass none of its decisions, and a sentence asking only why someone might make a harmful request bypasses 9% of its blocks, against 56% for GPT-6 Astra. On phishing detection. However, the same forgeries bypass 19-29% of the emails it flags, and even without attack Jev detects only 62% of phishing emails, against 77-81% for the GLM judges. Jev is also the fastest judge, but it flags the most benign requests (27%). The local classifiers fail differently: a training-example statement makes Llama Guard 3 miss 62% of the phishing it had flagged. These results suggest that Jev resists attacks on tasks with a clear rule to check but is more exposed on tasks that call for an overall judgment.
Model learns to pick video moments that explain anomalies better
Learning Temporal Evidence Selection for Video Anomaly Understanding
Abstract: Video anomaly understanding (VAU) aims to generate language-level interpretations of anomalous events in long videos, requiring visual evidence that supports explanation beyond anomaly detection alone. Existing VAU pipelines commonly rely on anomaly-centric or predefined temporal sampling strategies, which efficiently emphasize suspicious moments but overlook contextual cues, object interactions, and temporally distributed evidence needed for reasoning. Motivated by recent progress in learnable temporal selection for long-video question answering, we revisit VAU temporal sampling as a downstream-aligned evidence selection problem. We introduce the semantic-conditioned Temporal Sampling Q-Former that learns to select temporal visual evidence through the VAU generation objective. Since generic VAU prompts provide weak video-specific guidance, we condition the sampler on CLIP-embedded object tags automatically extracted from the input video. These tags are used only for temporal evidence selection and are not appended to the LLM generation prompt, separating sampler conditioning from prompt augmentation. The proposed sampler operates in visual-token space, computes soft temporal selection weights from frame-level representations, and merges duplicated selections before response generation. Experiments on HIVAU-70K show that our model achieves the best performance among the compared fine-tuned models, while ablations demonstrate that learnable temporal sampling improves over fixed anomaly-focused sampling and that CLIP-embedded object-tag conditioning further improves over generic-prompt conditioning. Our results suggest that temporal sampling in VAU should be aligned with explanation-oriented evidence utility rather than anomaly saliency alone.
Semantic watermarking technique detects ai generated text despite paraphrasing
SemField: A Simple, Linear, Continuous, yet Robust Semantic Watermark
Abstract: The rapid proliferation of Large Language Models (LLMs) necessitates reliable watermarking techniques to identify AI-generated text and ensure appropriate attribution. While token-based watermarks are vulnerable to paraphrasing, a central challenge for semantic watermarking is to turn sentence meanings into a stable, well-calibrated document-level signal. To this end, we introduce SemField, a simple, training-free semantic watermark that embeds a continuous, linear signal directly into the sentence embedding space. Using a shared secret key to define a specific Gaussian direction, SemField iteratively evaluates candidate sentences and selects those that maximize the alignment of the cumulative document aggregate with this targeted direction. We also propose SemField-PL, a polarity-locked variant that first determines the optimal orientation from the initial sentence and continuously reinforces it throughout the generation process. The document-level aggregated formulation guarantees exact invariance to sentence reordering and provides theoretical bounds against structural tampering, such as sentence insertion and deletion. Lastly, with extensive evaluations across three models, we demonstrate that both variants outperform 12 recent baselines. Across clean detection and four paraphrasing attacks, both variants achieve a mean True Positive Rate (TPR) of 88.2% to 90.4% at a 1% False Positive Rate (FPR), all while maintaining comparable perplexity and naturalness as human-generated content. We open-source our implementation at https://github.com/declare-lab/SemField
Human centered AI should model multiple valid interpretations
Whose Ground Truth? Embracing Ambiguity in Human-Centered AI
Abstract: As AI systems increasingly interact with people and make decisions about them, understanding human interpretations becomes an important part of developing human-centered AI. Conventional machine learning and AI systems are largely developed under the assumption that a single definitive ground truth exists, with variability in human annotations often resolved through aggregation or treated as noise. However, for many human-centered tasks, human interpretation is inherently ambiguous, and multiple interpretations of the same input may be simultaneously reasonable and valid. Reducing such ambiguity to a single target risks overlooking meaningful information about the diversity of human perception, judgment, and experience. In this position paper, we call for a shift towards modeling the interpretation space of plausible human judgments, while distinguishing meaningful ambiguity from annotation noise. We argue that this perspective should guide how AI systems are represented, learned, evaluated, deployed, and governed, supporting more human-centered AI that better reflects the diversity of human interpretation.
IdeaLens detects whether ideas in writing come from AI or humans
IdeaLens: Detecting AI Ideas in Long-form Writing
Abstract: While modern AI detectors identify who wrote the words, emerging policies on AI use increasingly hinge on a different question: who came up with the ideas? We introduce IdeaLens, a detector that identifies whether a document's ideas came from a human or AI (idea provenance), regardless of who wrote its words. To focus IdeaLens on ideas rather than prose, we represent documents as outlines: lists of items that each pair a discourse role with a brief, paraphrased description of the content, minimizing word-level overlap with the raw text. We train IdeaLens on 1M FineWeb documents with silver labels from Pangram, a prose provenance detector. Since the outlines are largely stripped of surface-level information, the labels must be fit mainly through the ideas. In a controlled study, IdeaLens's AI flag rate drops from 95% to 7% as models write from increasingly detailed human plans, while Pangram 4 still flags 92%; from AI-derived plans, IdeaLens stays above 96%. Conversely, on a new dataset of 50 stories that human authors wrote from AI-generated plans, IdeaLens flags 68% of the stories as AI, compared to 8% for Pangram 4. On a comprehensive suite of 19 existing detection benchmarks, we show that IdeaLens maintains strong detection rates at low false positive rates, suggesting that ideas themselves provide a powerful discriminative signal, and its performance holds across domains, formats, and languages. Finally, we examine 90K predictions from IdeaLens to characterize systematic differences between human and AI ideation. We release our models and labeled datasets to facilitate future research on idea provenance detection.
Diffusion models enable verifiable unlearning of specific concepts internally
No Concept Escapes the Audit: Auditing-Aware Unlearning for Verifiable Concept Erasure in Diffusion Models
Abstract: Text-to-image diffusion models can generate prohibited content, which motivates concept erasure through machine unlearning. Most erasure methods intervene at the text interface, through prompt modification or localized updates to text-conditioning weights, and they are evaluated by what the model outputs for given prompts. Such evaluation cannot see what the network still encodes. Latent-space auditing, which bypasses text conditioning and probes the denoising network directly, shows that erased concepts remain recoverable from internal representations. We find that this also holds for methods built to be robust against adversarial prompts, and that the problem grows with the number of erased concepts. We propose Auditing-Aware Unlearning for Verifiable Concept Erasure in Diffusion Models (AVCE), a framework that grounds erasure in the model's latent representations. AVCE audits the embedding neighborhood of each concept and condenses the discovered vulnerable directions into an anchor at the weakest geometric point. It edits cross-attention and self-attention projections in closed form at this anchor, then fine-tunes the two pathways with pathway-level auditing losses, using orthogonal gradient projection to consolidate multiple concepts. Experiments on SD v1.5, SDXL, and Flux 1.0 across object, explicit-content, and artistic-style unlearning show that AVCE reduces attack success rates by 5.07x and improves auditing scores by 3.84x over the strongest baseline, while preserving competitive generation quality.
Selective video unlearning removes concepts while preserving the scene
FOMO: Forget the Concept, Don't Miss Out on the Scene in Selective Video Unlearning
Abstract: The rapid advancement of generative video models has enabled the synthesis of increasingly realistic and temporally coherent videos, while also raising concerns about the generation of harmful content. The reliance on large-scale web datasets during training inevitably exposes these models to undesirable material, making concept unlearning an essential mitigation. Existing methods mainly target static visual concepts, such as objects, identities, or unsafe appearance, largely overlooking motion unlearning. Furthermore, these approaches often pay little attention to preserving the surrounding scene. As a result, successful concept removal may unintentionally alter the background, composition, or overall video dynamics. We argue that effective unlearning should ideally change only what is targeted, while minimizing unnecessary changes to the remaining scene. In this work, we introduce FOMO, to the best of our knowledge the first training-based selective video unlearning method that directly treats preservation of the original scene as a priority. We formulate unlearning around two complementary objectives: what to change and what to preserve. Our method localizes concept-related representations and modifies them, while the preservation mechanism maintains non-target scene information without requiring auxiliary data. Beyond simply erasing unwanted concepts, FOMO explicitly redirects the generation toward a specified safe alternative. We further extend this formulation to motion unlearning, where the concept is defined by temporal behavior rather than a fixed spatial region. Our solution achieves effective unlearning across unsafe content, object, and motion concepts, while achieving the best trade-off between concept removal and scene preservation. Code: https://github.com/gmum/FOMO Project Page https://gmum.github.io/FOMO
Local error checks improve detection of edits in watermarked AI text
Anchor-ECC: Local Integrity Checking for Watermarked LLM Outputs via Error-Correcting Codes
Abstract: LLM watermarking has become an effective approach to distinguishing AI-generated text from human-written text by embedding detectable patterns during generation. However, a small post-generation edit may change the meaning of the text without removing its overall watermark signal, creating a risk that the modified content is still attributed to the original model. We propose Anchor-ECC, which incorporates the error-correcting code (ECC) constraints and explicit boundary anchors into the watermark structure and pairs them with a dynamic-programming decoder to detect and localize post-generation edits. Across Qwen3-8B, Mistral-7B-Instruct-v0.3, and OPT-125M, the approximate-hard setting achieves about 99.7% block-level true positive rate (TPR) with at most 7.6% false alarm rate (FAR) for edit detection under mixed insertions, deletions, and substitutions, while preserving the distinction between watermarked outputs and unwatermarked text. Additional quality experiments identify lower-perplexity configurations that retain strong edit-detection performance. Together, these results extend LLM watermarking from source identification to local integrity verification while supporting configurable trade-offs between detection reliability and generation quality.
Autoresearch discovers better watermarks for language models
AutoMark: Enabling Autoresearch to Discover Better LLM Watermarks
Abstract: With LLM watermarking being deployed commercially and now required by regulations, improving its reliability and effectiveness has become crucial. Yet, recent progress in the field of LLM watermarking has increasingly been driven by improving details of existing methods, an effort fundamentally limited by the pace of human researchers. In this work, we enable for the first time the autonomous discovery of new distortion-free state-of-the-art watermarking schemes. To enable this, we (i) establish strict criteria to ensure that watermarks are reliable (e.g., they do not have an unexpectedly high false positive rate), (ii) propose rigorous statistical tests to automatically evaluate whether a watermarking scheme satisfies our criteria, and (iii) design an evaluation suite to rank watermarks along three key dimensions: detectability, quality, and robustness. By running our framework with 3 frontier models (GPT-6 Astra, Opus 5, Gemini-3.8 Flash), we discover over 50 different watermarking schemes, including several that outperform prior works along all key dimensions. We complement this by a manual study of the discovered schemes, distilling the key ideas into smaller components, and individually studying the impact of each component across dimensions (detectability, quality, robustness) to better understand how the proposed schemes operate. Importantly, we find that the agents, on top of improving existing ideas, also discover fundamentally new ideas (e.g., aligning watermark scores with random per-request direction). Overall, our work establishes the first steps of fully autonomous watermarking research, enabling the discovery of more reliable and effective watermarks. Our code is available at https://github.com/eth-sri/automark, and a blogpost to visualize our results at https://www.sri.inf.ethz.ch/blog/automark.
Clipping improves detecting AI-generated text even with editing errors
Robust Detection of LLM-Generated Text under Contamination
Abstract: We study the detection of LLM-generated text under editing and contamination. Modeling human and machine text as finite-order Markov processes with Huber contamination, we characterize an exact boundary for reliable detection under our assumptions. Detection is impossible when contamination is sufficiently large relative to clean-source separation. Below this boundary, a collection of clipped likelihood-ratio tests achieves vanishing worst-case errors. This construction motivates clipping as a simple modification of existing statistical detectors. For a broad class of additive scores, we identify conditions under which the clipped test is consistent while the raw test's worst-case power tends to zero. We evaluate seven detectors across three datasets and three generation models, and on the RAID benchmark. Clipping improves robustness in both studies, with gains varying across detectors and contamination settings. For example, at a target false-positive rate of 5\%, clipping improves the log-likelihood--log-rank ratio (LRR) detector's true-positive rate by a median of 8.3 percentage points in the controlled study and 2.1 and 4.3 points in rate- and attack-specific RAID evaluations, respectively.
How query wording affects AI agreement in relationship advice chats
Sweet Talkers: How Query Formulation Shapes Sycophancy in Romantic Relationship Advice
Abstract: Large language models (LLMs) are increasingly used for emotional support and relationship advice, where a model's tendency to preserve a user's face can inadvertently reinforce harmful interpersonal behaviors. To systematically examine this risk, we developed the Romantic Relationship Advice-Seeking Prompts (RRASP) dataset of 2,400 prompts across five relationship themes and evaluated social sycophancy using the ELEPHANT framework on two consumer-facing models, GPT-5 Mini and Gemini 3 Flash. Contrary to our initial hypothesis, grammatical mood alone did not produce systematic differences in sycophantic behavior, suggesting that what a user implies matters more than how they phrase it. Instead, perspective-driven framing had a stronger influence, with gaps between original and flipped prompts widening in follow-up responses. Consistent increases in framing and moral sycophancy across turns indicate that models become more likely to accept a user's stated premises and affirm their ethical stance as a dialogue progresses. Notably, Gemini 3 Flash exhibited substantially smaller increases in moral sycophancy than GPT-5 Mini, suggesting it is more resistant to reinforcing ethically problematic positions across turns.
Multilingual large language models struggle with Urdu stories
Multilingual in Name Only? Cultural and Linguistic Weaknesses of LLMs in Urdu
Abstract: Multilingual large language models (LLMs) are increasingly used for open-ended text generation, yet their behaviour in low-resource languages remains poorly understood. In this work, we question how correct and reliable is the generation of multilingual LLMs when used for the task of story generation. We consider Urdu language as a representative low-resource language. We generate Urdu-Stories, a corpus of 93 stories generated using three contemporary LLMs (GPT-5.1, Qwen-3-Max, DeepSeek-3.1). We manually annotate the errors present in them under a nine-label linguistic, semantic, and cultural taxonomy. Our notable findings suggest that LLMs often make basic errors of grammar and semantics. The stories lack coherence, have unnatural repetition and show pervasive cultural shallowness. We further show using few-shot prompting that the cultural and context errors largely remain unresolved. Our findings highlight the limitations of current LLMs as a reliable source of content generation and information retrieval for low-resource languages.