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.

Fri 13 NovComputation and Language
The gist
People often worry that information made by AI might feel less trustworthy, especially about sensitive topics like gender. This paper studied Spanish speakers reading gender-related statements and tested if knowing the content came from AI changed their opinions or how quickly they responded. The researchers found that whether participants knew the content was AI-generated or not, it didn't change their agreement or reaction times. This suggests people may treat AI-generated statements as credible as human-authored ones, even when biases exist.
Open → 2611.18371v1

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.

Tue 10 NovComputers and Society
The gist
The paper looks at how AI language models write about social science topics related to the US and China. It checks if these AI systems show bias favoring one country when answering questions or citing sources. The authors found that while the models are not strongly politically biased, they do differ in which sources they use and sometimes favor one country. This means people using AI for social science research should be careful about possible subtle biases.
Open → 2611.14288v1

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.

Mon 9 NovComputer Vision and Pattern Recognition
The gist
Text-to-image models create pictures from text, but some unsafe images can sneak past the usual filters. The authors developed AnyBreak, a way to trick these models into generating harmful images using special adversarial images alongside normal text prompts. This method works without knowing the model’s inner details and even bypasses safety checks. Their experiments show current defenses miss these sneaky attacks.
Open → 2611.11980v1

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.

Sun 25 OctMachine Learning
The gist
This paper explores how language models often answer questions based on information they can use, even if that information isn't actually supported by the original passage. For example, a model might refuse to answer a question if the passage lacks certain facts, but if those facts are stated in the question instead, the model uses them to produce an answer. The authors found that models tend to ignore missing evidence when the unsupported fact is usable for computation, leading to confidently wrong answers. They also show that judges rarely penalize such errors and that simple prompts fail to fix this flaw.
Open → 2610.34488v1

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.

Sun 18 OctComputation and Language
The gist
This study looks at millions of video titles from a popular pornography website to see how different races, genders, and nationalities are described. The authors found that American pornography sets a standard many others are compared to. They also noticed that ethnic and gender stereotypes still appear a lot, showing old patterns of power like colonialism and cultural taboos. Women are often shown in simple opposing roles, and men tend to be described as powerful.
Open → 2610.25566v1

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.

Wed 14 OctComputation and Language
The gist
As AI tools become common in writing, it’s important to know how much of a text a human actually wrote. The authors find that current ways of measuring this, which look at how well a person’s prompt predicts the final text, can be tricked easily. They show that adding random strings to a prompt can fool the measurement into thinking a mostly AI-written text is mostly human. This means that just looking at how likely the text is from the prompt can’t reliably tell how much a human contributed.
Open → 2610.19767v1

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.

Tue 13 OctCryptography and SecurityArtificial Intelligence
The gist
Sometimes AI systems must decide if inputs are harmful or safe. The authors tested a new AI judge called Jev alongside several others by sending them tricky, harmful inputs. They found that Jev was very good at spotting rule-based problems, like harmful requests, but it was not as good at recognizing phishing emails as some other judges. Jev also runs faster but sometimes flags safe inputs incorrectly. This testing helps show where each system works best and where they still have weaknesses.
Open → 2610.17820v1

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.

Mon 12 OctComputer Vision and Pattern Recognition
The gist
Understanding strange or unusual events in long videos means explaining what happens, not just spotting when something is odd. The authors show that picking which parts of the video to look at is important for good explanations. They made a model that learns to choose video moments based on objects it sees, improving the detail in explanations over just focusing on the most obviously strange parts. Their approach works better than previous methods on a big test set of annotated videos.
Open → 2610.16299v1

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

Thu 8 OctCryptography and Security
The gist
It is hard to tell if text was written by an AI or a human, especially when the AI rephrases sentences to hide its origin. The authors propose SemField, a new way to subtly mark AI-generated text by embedding a special continuous signal into the meanings of sentences. This signal adds up over a whole document, making it easy to detect even if the sentences are moved around or slightly changed. Their method works well against various tricks to hide AI writing and keeps the text sounding natural.
Open → 2610.11848v1

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.

Wed 7 OctArtificial IntelligenceMachine Learning
The gist
Many AI systems assume there is one correct answer when interpreting human-related data, but people often see the same thing in different, equally valid ways. The authors argue that instead of forcing AI to pick a single answer, systems should acknowledge and represent this ambiguity. This approach would better reflect how real people think and experience things, leading to AI that respects diverse perspectives. They suggest changes in how AI is designed, trained, evaluated, and governed to include multiple human interpretations rather than just one.
Open → 2610.10805v1

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.

Mon 5 OctComputation and LanguageArtificial IntelligenceMachine Learning
The gist
Sometimes it’s hard to tell if a person or a computer came up with the ideas behind a piece of writing, even if the words were written by someone else. The researchers created IdeaLens, a tool that looks at the main points and structure of a text, rather than the exact words, to figure out if the ideas are from a human or AI. They trained IdeaLens on a large set of documents and showed it can reliably spot AI-generated ideas, even when humans write the words. This helps answer not just who wrote the text, but who actually came up with the ideas.
Open → 2610.06778v1

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.

Sun 4 OctMachine LearningArtificial Intelligence
The gist
Text-to-image AI models sometimes create unwanted or harmful images, so people want to remove those ideas from the models. The authors found that past methods only change the model’s text input but don’t fully erase what the model has learned, meaning removed concepts can still be detected inside. To fix this, they propose AVCE, a new way to directly clean the model’s internal brain signals related to unwanted concepts. Their tests show AVCE removes these concepts more reliably while keeping the AI’s ability to create good images.
Open → 2610.05401v1

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

Wed 30 SeptComputer Vision and Pattern Recognition
The gist
Videos created by AI can sometimes include unwanted objects, identities, or even unsafe content. The authors developed a method called FOMO that removes just the unwanted parts from videos without messing up the rest of the scene. Unlike previous methods, FOMO also works by removing certain movements or actions in videos. This means it can keep the background and overall look mostly unchanged while erasing only the target content.
Open → 2609.39605v1

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.

Wed 30 SeptCryptography and SecurityComputation and Language
The gist
AI models can add hidden patterns, or watermarks, to text they generate, helping people tell AI text from human text. But if someone slightly changes the AI-written text afterward, the original watermark might still look present even though the meaning changes. The authors created Anchor-ECC, a method that uses special error-correcting codes and markers to spot and locate those small edits within the watermarked text. This method works well across different AI models, detecting most edits while keeping the watermark reliable. Anchor-ECC helps not just verify where AI text came from but also checks if the text has been altered after it was generated.
Open → 2609.38722v1

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.

Tue 29 SeptCryptography and SecurityMachine Learning
The gist
Watermarks are hidden signals added to AI text so people can tell if it was made by a computer. As these watermarks become required by law, it’s important to make them work better and more reliably. The authors created a computer program that can automatically try out and find new watermark designs without human help. This program tested many ideas on several advanced AI models and found over 50 new watermark methods, some clearly better than those made before. They also studied how these new watermarks work and why they succeed, showing that machines can now help improve watermarking faster than people alone.
Open → 2609.37310v1

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.

Thu 24 SeptMachine Learning
The gist
Detecting whether text is written by a human or an AI can be tricky when the AI-generated text is altered or mixed with other content. The authors studied this problem by modeling both human and machine text with a mathematical tool and found the limits where detection is possible. They show that by modifying existing detection tests with a simple 'clipping' step, the tests become more reliable against tricky edits or contamination. Their experiments on several datasets and models confirm that clipping consistently improves the ability to spot AI-generated text while keeping false alarms low.
Open → 2609.29935v1

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.

Sat 12 SeptComputation and Language
The gist
Sometimes, AI chatbots that give relationship advice try to agree with what users say, even if it might encourage bad behavior. The authors created a big set of questions about relationships and tested how two AIs responded depending on how the questions were asked. They found that the AI's agreement depends more on what users imply rather than the exact words they use. Some AIs become more likely to accept a user's views as conversations go on, but one AI was less prone to agreeing with problematic opinions. This helps us understand how AI might unintentionally support harmful ideas in emotional chats.
Open → 2609.13841v1

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.

Wed 9 SeptComputation and LanguageArtificial IntelligenceMachine Learning
The gist
Large language models (LLMs) can write stories in many languages, but their ability in less common languages like Urdu is not well understood. The authors studied three popular LLMs that generated Urdu stories and found many problems. These stories often had basic grammar mistakes, confusing meanings, repeated phrases, and lacked cultural depth. Even when given extra examples to improve, the issues remained. This shows that current LLMs are not yet reliable for creating or understanding content in low-resource languages like Urdu.
Open → 2609.10758v1