Tag
Trakkr Research tested 1,285 statistics cited by ChatGPT, Perplexity, Gemini, Claude, and Google AI Overviews, finding only 40% traced back to a study saying the same thing, while many 'zombie stats' circulate with no research behind them.
PROVISO 是一个在 LLM agent 执行动作前进行溯源感知决策的控制器，通过将可信与外部观察编码为溯源类型记忆来平衡安全性与实用性，在多个 agent 安全基准上显著优于现有基线。
Sentient Foundation 介绍了 Fingerprinting Leaderboard，一个针对模型指纹技术的开放基准，用于在模型权重公开后验证其归属。该基准从模型质量保留、抗移除能力以及检测可靠性三个维度对各类指纹方法进行评分。
本文通过 LLM 辅助的内容分析，从专业人道主义情况报告中推导出 15 槽信息架构，并评估了 121 份人类报告和 261 份 LLM 生成报告，发现 LLM 生成的危机报告在部分指标上媲美人类，但无法提供专业出处且仅复现少量官方数据，自动评分指标（如 ROUGE-L、BERTScore）与报告完整性关联不佳。
This paper introduces Latent Quotient Evidence Training (LQET), a relational training objective that helps language models update beliefs based on evidence provenance rather than report count, reducing overconfidence from repeated relayed information in human-AI feedback loops with 41-44% RMSE reduction across Qwen3-4B and Phi-4-mini.
文章介绍了 OpenAI 为满足欧盟 AI 法案第 50(2) 条来源透明度要求而推出的文本水印 textGrain，并展示其发布的失效曲线：同义词替换 10% 时检测率从 92% 降至 66%，替换 25% 时仅剩 17%。作者还构建了一个交互式攻击实验室，可实时演示翻译回译、同义词替换和数学化文本如何让水印检测的 p 值崩塌，强调未检测到水印并不能证明内容出自人类之手。
论文首次基于 GitHub 全量 SKILL.md 的 git 历史构建了 AI 代理技能的复制网络，追踪 219 万次技能采用，发现少数仓库是几乎所有复制的源头、安全修复难以传播，并提出基于复制模型的审计排序方法以识别高风险技能仓库。
A U.S. Executive Order directs agencies to replace "AI" with "Super Intelligence" in official communications, and NIST is migrating its public terminology accordingly — raising provenance and evidence-intelligence concerns about distinguishing name changes from substantive changes in source authority.
The paper evaluates twenty deepfake detectors against ten generators, finding detection accuracy degrading from 99.5% to 76% over time and collapsing below 2% under adversarial perturbations, then proposes a calibrated detection paradigm based on plausibly deniable authenticity via generator reconstruction, achieving at most 1% false certification.
After Anthropic announced invisible watermarks in Claude's output, the open-source tool watermarks-remover quickly gained traction, accumulating 15,000+ Stars in a week. It can remove provenance markers and metadata from multiple vendors' AI models.
An article examining how agent memory systems should distinguish source observations, accepted decisions, derived summaries, and operational checkpoints in a human-readable layer, asking how to draw boundaries between files, relational stores, and vector indexes to prevent stale summaries from outliving their sources.
论文提出 ForgePrint（搜索-蒸馏框架），通过定向改写将一个模型的输出伪装成目标模型的指纹，使归因分类器误判来源模型，在 CNN/DM 上达到 70.2% 的目标成功率，说明文本归因不能等同于来源真实性。
The article discusses the challenges of maintaining audit trails in multi-agent AI coding workflows, highlighting the need for agent identity, delegation, and provenance to manage responsibility and control.
This academic paper explores the provenance and admissibility of machine self-reports in AI systems, discussing implications for trust, transparency, and legal contexts.
Apple introduces an opt-in camera mode on iPhone 18 Pro that uses hardware-based cryptographic signing and cloud processing to verify photo authenticity, rejecting the C2PA standard in favor of a hardware-rooted approach.
The paper introduces ClaimReceipt, a claim-relative receipt specification and verifier for verifying evidence sufficiency and coverage in agent evaluations, validated through experiments on historical records and prospective audits.
This paper proposes claim-locked reporting, a provenance-before-prose protocol that fixes statistical evidence before LLM generation to improve reproducibility and accuracy in scientific reports.
The article analyzes the provenance and censorship behavior of the Ox Alpha AI model, identifying it as part of the GLM family and revealing a unique censorship pattern on specific topics like domestic incidents.
ECHO is a cognitively inspired, auditable memory architecture for long-horizon agents, evaluated on benchmarks like LoCoMo and LongMemEval-S with high retrieval performance.
This paper proposes a composable trust infrastructure for manufacturing knowledge graphs, focusing on cross-system provenance, temporal reasoning, and decision traceability.