Large Language Models (LLMs) perform impressively well in various applications. However, the potential for misuse of these models in activities such as plagiarism, generating fake news, and spamming has raised concern about their responsible use. Consequently, the reliable detection of AI-generated text has become a critical area of research. AI text detectors have shown to be effective under their specific settings. In this paper, we stress-test the robustness of these AI text detectors in the presence of an attacker. We introduce recursive paraphrasing attack to stress test a wide range of detection schemes, including the ones using the watermarking as well as neural network-based detectors, zero shot classifiers, and retrieval-based detectors. Our experiments conducted on passages, each approximately 300 tokens long, reveal the varying sensitivities of these detectors to our attacks. Our findings indicate that while our recursive paraphrasing method can significantly reduce detection rates, it only slightly degrades text quality in many cases, highlighting potential vulnerabilities in current detection systems in the presence of an attacker. Additionally, we investigate the susceptibility of watermarked LLMs to spoofing attacks aimed at misclassifying human-written text as AI-generated. We demonstrate that an attacker can infer hidden AI text signatures without white-box access to the detection method, potentially leading to reputational risks for LLM developers. Finally, we provide a theoretical framework connecting the AUROC of the best possible detector to the Total Variation distance between human and AI text distributions. This analysis offers insights into the fundamental challenges of reliable detection as language models continue to advance. Our code is publicly available at https://github.com/vinusankars/Reliability-of-AI-text-detectors.
The rise in malicious usage of large language models, such as fake content creation and academic plagiarism, has motivated the development of approaches that identify AI-generated text, including those based on watermarking or outlier detection. However, the robustness of these detection algorithms to paraphrases of AI-generated text remains unclear. To stress test these detectors, we build a 11B parameter paraphrase generation model (DIPPER) that can paraphrase paragraphs, condition on surrounding context, and control lexical diversity and content reordering. Using DIPPER to paraphrase text generated by three large language models (including GPT3.5-davinci-003) successfully evades several detectors, including watermarking, GPTZero, DetectGPT, and OpenAI's text classifier. For example, DIPPER drops detection accuracy of DetectGPT from 70.3% to 4.6% (at a constant false positive rate of 1%), without appreciably modifying the input semantics. To increase the robustness of AI-generated text detection to paraphrase attacks, we introduce a simple defense that relies on retrieving semantically-similar generations and must be maintained by a language model API provider. Given a candidate text, our algorithm searches a database of sequences previously generated by the API, looking for sequences that match the candidate text within a certain threshold. We empirically verify our defense using a database of 15M generations from a fine-tuned T5-XXL model and find that it can detect 80% to 97% of paraphrased generations across different settings while only classifying 1% of human-written sequences as AI-generated. We open-source our models, code and data.
Advancements in model algorithms, the growth of foundational models, and access to high-quality datasets have propelled the evolution of Artificial Intelligence Generated Content (AIGC). Despite its notable successes, AIGC still faces hurdles such as updating knowledge, handling long-tail data, mitigating data leakage, and managing high training and inference costs. Retrieval-Augmented Generation (RAG) has recently emerged as a paradigm to address such challenges. In particular, RAG introduces the information retrieval process, which enhances the generation process by retrieving relevant objects from available data stores, leading to higher accuracy and better robustness. In this paper, we comprehensively review existing efforts that integrate RAG technique into AIGC scenarios. We first classify RAG foundations according to how the retriever augments the generator, distilling the fundamental abstractions of the augmentation methodologies for various retrievers and generators. This unified perspective encompasses all RAG scenarios, illuminating advancements and pivotal technologies that help with potential future progress. We also summarize additional enhancements methods for RAG, facilitating effective engineering and implementation of RAG systems. Then from another view, we survey on practical applications of RAG across different modalities and tasks, offering valuable references for researchers and practitioners. Furthermore, we introduce the benchmarks for RAG, discuss the limitations of current RAG systems, and suggest potential directions for future research. Github: https://github.com/PKU-DAIR/RAG-Survey.
Our work addresses the critical issue of distinguishing text generated by Large Language Models (LLMs) from human-produced text, a task essential for numerous applications. Despite ongoing debate about the feasibility of such differentiation, we present evidence supporting its consistent achievability, except when human and machine text distributions are indistinguishable across their entire support. Drawing from information theory, we argue that as machine-generated text approximates human-like quality, the sample size needed for detection increases. We establish precise sample complexity bounds for detecting AI-generated text, laying groundwork for future research aimed at developing advanced, multi-sample detectors. Our empirical evaluations across multiple datasets (Xsum, Squad, IMDb, and Kaggle FakeNews) confirm the viability of enhanced detection methods. We test various state-of-the-art text generators, including GPT-2, GPT-3.5-Turbo, Llama, Llama-2-13B-Chat-HF, and Llama-2-70B-Chat-HF, against detectors, including oBERTa-Large/Base-Detector, GPTZero. Our findings align with OpenAI's empirical data related to sequence length, marking the first theoretical substantiation for these observations.
Recently, ChatGPT, along with DALL-E-2 and Codex,has been gaining significant attention from society. As a result, many individuals have become interested in related resources and are seeking to uncover the background and secrets behind its impressive performance. In fact, ChatGPT and other Generative AI (GAI) techniques belong to the category of Artificial Intelligence Generated Content (AIGC), which involves the creation of digital content, such as images, music, and natural language, through AI models. The goal of AIGC is to make the content creation process more efficient and accessible, allowing for the production of high-quality content at a faster pace. AIGC is achieved by extracting and understanding intent information from instructions provided by human, and generating the content according to its knowledge and the intent information. In recent years, large-scale models have become increasingly important in AIGC as they provide better intent extraction and thus, improved generation results. With the growth of data and the size of the models, the distribution that the model can learn becomes more comprehensive and closer to reality, leading to more realistic and high-quality content generation. This survey provides a comprehensive review on the history of generative models, and basic components, recent advances in AIGC from unimodal interaction and multimodal interaction. From the perspective of unimodality, we introduce the generation tasks and relative models of text and image. From the perspective of multimodality, we introduce the cross-application between the modalities mentioned above. Finally, we discuss the existing open problems and future challenges in AIGC.
To address the challenges of digital intelligence in the digital economy, artificial intelligence-generated content (AIGC) has emerged. AIGC uses artificial intelligence to assist or replace manual content generation by generating content based on user-inputted keywords or requirements. The development of large model algorithms has significantly strengthened the capabilities of AIGC, which makes AIGC products a promising generative tool and adds convenience to our lives. As an upstream technology, AIGC has unlimited potential to support different downstream applications. It is important to analyze AIGC's current capabilities and shortcomings to understand how it can be best utilized in future applications. Therefore, this paper provides an extensive overview of AIGC, covering its definition, essential conditions, cutting-edge capabilities, and advanced features. Moreover, it discusses the benefits of large-scale pre-trained models and the industrial chain of AIGC. Furthermore, the article explores the distinctions between auxiliary generation and automatic generation within AIGC, providing examples of text generation. The paper also examines the potential integration of AIGC with the Metaverse. Lastly, the article highlights existing issues and suggests some future directions for application.
As AI-generated text continues to evolve, distinguishing it from human-authored content has become increasingly difficult. This study examined whether non-expert readers could reliably differentiate between AI-generated poems and those written by well-known human poets. We conducted two experiments with non-expert poetry readers and found that participants performed below chance levels in identifying AI-generated poems (46.6% accuracy, χ 2 (1, N = 16,340) = 75.13, p < 0.0001). Notably, participants were more likely to judge AI-generated poems as human-authored than actual human-authored poems (χ 2 (2, N = 16,340) = 247.04, p < 0.0001). We found that AI-generated poems were rated more favorably in qualities such as rhythm and beauty, and that this contributed to their mistaken identification as human-authored. Our findings suggest that participants employed shared yet flawed heuristics to differentiate AI from human poetry: the simplicity of AI-generated poems may be easier for non-experts to understand, leading them to prefer AI-generated poetry and misinterpret the complexity of human poems as incoherence generated by AI.
Abstract Recent advances in generative pre-trained transformer large language models have emphasised the potential risks of unfair use of artificial intelligence (AI) generated content in an academic environment and intensified efforts in searching for solutions to detect such content. The paper examines the general functionality of detection tools for AI-generated text and evaluates them based on accuracy and error type analysis. Specifically, the study seeks to answer research questions about whether existing detection tools can reliably differentiate between human-written text and ChatGPT-generated text, and whether machine translation and content obfuscation techniques affect the detection of AI-generated text. The research covers 12 publicly available tools and two commercial systems (Turnitin and PlagiarismCheck) that are widely used in the academic setting. The researchers conclude that the available detection tools are neither accurate nor reliable and have a main bias towards classifying the output as human-written rather than detecting AI-generated text. Furthermore, content obfuscation techniques significantly worsen the performance of tools. The study makes several significant contributions. First, it summarises up-to-date similar scientific and non-scientific efforts in the field. Second, it presents the result of one of the most comprehensive tests conducted so far, based on a rigorous research methodology, an original document set, and a broad coverage of tools. Third, it discusses the implications and drawbacks of using detection tools for AI-generated text in academic settings.
Purpose With the advent of ChatGPT, a sophisticated generative artificial intelligence (AI) tool, maintaining academic integrity in all educational settings has recently become a challenge for educators. This paper discusses a method and necessary strategies to confront this challenge. Design/methodology/approach In this study, a language model was defined to achieve high accuracy in distinguishing ChatGPT-generated essays from human written essays with a particular focus on “not falsely” classifying genuinely human-written essays as AI-generated (Negative). Findings Via support vector machine (SVM) algorithm 100% accuracy was recorded for identifying human generated essays. The author discussed the key use of Recall and F2 score for measuring classification performance and the importance of eliminating False Negatives and making sure that no actual human generated essays are incorrectly classified as AI generated. The results of the proposed model's classification algorithms were compared to those of AI-generated text detection software developed by OpenAI, GPTZero and Copyleaks. Practical implications AI-generated essays submitted by students can be detected by teachers and educational designers using the proposed language model and machine learning (ML) classifier at a high accuracy. Human (student)-generated essays can and must be correctly identified with 100% accuracy even if the overall classification accuracy performance is slightly reduced. Originality/value This is the first and only study that used an n-gram bag-of-words (BOWs) discrepancy language model as input for a classifier to make such prediction and compared the classification results of other AI-generated text detection software in an empirical way.
Advancements in artificial intelligence (AI) over the last decade demonstrate that machines can exhibit communicative behavior and influence how humans think, feel, and behave. In fact, the recent development of ChatGPT has shown that large language models (LLMs) can be leveraged to generate high-quality communication content at scale and across domains, suggesting that they will be increasingly used in practice. However, many questions remain about how knowing the source of the messages influences recipients' evaluation of and preference for AI-generated messages compared to human-generated messages. This paper investigated this topic in the context of vaping prevention messaging. In Study 1, which was pre-registered, we examined the influence of source disclosure on young adults' evaluation of AI-generated health prevention messages compared to human-generated messages. We found that source disclosure (i.e., labeling the source of a message as AI vs. human) significantly impacted the evaluation of the messages but did not significantly alter message rankings. In a follow-up study (Study 2), we examined how the influence of source disclosure may vary by the adults’ negative attitudes towards AI. We found a significant moderating effect of negative attitudes towards AI on message evaluation, but not for message selection. However, source disclosure decreased the preference for AI-generated messages for those with moderate levels (statistically significant) and high levels (directional) of negative attitudes towards AI. Overall, the results of this series of studies showed a slight bias against AI-generated messages once the source was disclosed, adding to the emerging area of study that lies at the intersection of AI and communication.
Abstract The question of how generative AI tools, such as large language models and chatbots, can be leveraged ethically and effectively in education is ongoing. Given the critical role that writing plays in learning and assessment within educational institutions, it is of growing importance for educators to make thoughtful and informed decisions as to how and in what capacity generative AI tools should be leveraged to assist in the development of students’ writing skills. This paper reports on two longitudinal studies. Study 1 examined learning outcomes of 48 university English as a new language (ENL) learners in a six-week long repeated measures quasi experimental design where the experimental group received writing feedback generated from ChatGPT (GPT-4) and the control group received feedback from their human tutor. Study 2 analyzed the perceptions of a different group of 43 ENLs who received feedback from both ChatGPT and their tutor. Results of study 1 showed no difference in learning outcomes between the two groups. Study 2 results revealed a near even split in preference for AI-generated or human-generated feedback, with clear advantages to both forms of feedback apparent from the data. The main implication of these studies is that the use of AI-generated feedback can likely be incorporated into ENL essay evaluation without affecting learning outcomes, although we recommend a blended approach that utilizes the strengths of both forms of feedback. The main contribution of this paper is in addressing generative AI as an automatic essay evaluator while incorporating learner perspectives.
Innovation in design is increasingly crucial to satisfy the growing demand for user-centric products and mitigate the effects caused by design uncertainties. Recently, the integration of AI technologies into product design processes has ignited unprecedented opportunities for design innovation, offering novel opportunities to augment processes and enhance design outcomes. AI-generated content (AIGC) represents emerging AI technologies that can efficiently generate quality design data across various modalities such as texts, images, audio, video and simulated sensor data. With such technology, designers are enabled to simplify and expedite the design process while ensuring the creation and production of high-quality design objects. This study aims to report the role of state-of-the-art AIGC technologies, how they will drive design innovation and identify their promising future applications. It delves into the vision, challenges, and opportunities presented by AIGC technologies, offering insights to inspire further research agenda. By leveraging various AIGC technologies, designers can unlock new avenues for creativity and productivity, ultimately shaping the future of design innovation.
Generative AI (GAI) has emerged as a significant advancement in artificial intelligence, renowned for its language and image generation capabilities. This paper presents “AI-Generated Everything” (AIGX), a concept that extends GAI beyond mere content creation to real-time adaptation and control across diverse technological domains. In networking, AIGX collaborates closely with physical, data link, network, and application layers to enhance real-time network management that responds to various system and service settings as well as application and user requirements. Networks, in return, serve as crucial components in further AIGX capability optimization through the AIGX lifecycle, i.e., data collection, distributed pre-training, and rapid decision-making, thereby establishing a mutually enhancing interplay. Moreover, we offer an in-depth case study focused on power allocation to illustrate the interdependence between AIGX and networking systems. Through this exploration, the article analyzes the significant role of GAI for networking, clarifies the ways networks augment AIGX functionalities, and underscores the virtuous interactive cycle they form. It is hoped that this article will pave the way for subsequent future research aimed at fully unlocking the potential of GAI and networks.
With advancements in generative artificial intelligence (AI) technology, there is a growing concern about its ethical implications, transparency, and consumer reactions to AI-generated content. Building on the persuasion knowledge model and algorithm aversion literature, this study explores the effects of AI disclosure in prosocial advertising on consumer attitudes and donation intentions. The findings of Study 1 indicate that the initial disclosure of AI-generated content leads to unfavourable attitudes towards ads, with perceived ad credibility serving as a mediating factor. In Study 2, participants who perceived AI as more human-like rather than machine-like tended to experience a diminished negative impact of AI disclosure. Study 3 also highlights the crucial role of perceived ad credibility in influencing donation intentions following the disclosure of AI-generated content. The theoretical and practical implications of our findings for social marketers and nonprofit organizations are discussed further.
Abstract Peer feedback is introduced as an effective learning strategy, especially in large-size classes where teachers face high workloads. However, for complex tasks such as writing an argumentative essay, without support peers may not provide high-quality feedback since it requires a high level of cognitive processing, critical thinking skills, and a deep understanding of the subject. With the promising developments in Artificial Intelligence (AI), particularly after the emergence of ChatGPT, there is a global argument that whether AI tools can be seen as a new source of feedback or not for complex tasks. The answer to this question is not completely clear yet as there are limited studies and our understanding remains constrained. In this study, we used ChatGPT as a source of feedback for students’ argumentative essay writing tasks and we compared the quality of ChatGPT-generated feedback with peer feedback. The participant pool consisted of 74 graduate students from a Dutch university. The study unfolded in two phases: firstly, students’ essay data were collected as they composed essays on one of the given topics; subsequently, peer feedback and ChatGPT-generated feedback data were collected through engaging peers in a feedback process and using ChatGPT as a feedback source. Two coding schemes including coding schemes for essay analysis and coding schemes for feedback analysis were used to measure the quality of essays and feedback. Then, a MANOVA analysis was employed to determine any distinctions between the feedback generated by peers and ChatGPT. Additionally, Spearman’s correlation was utilized to explore potential links between the essay quality and the feedback generated by peers and ChatGPT. The results showed a significant difference between feedback generated by ChatGPT and peers. While ChatGPT provided more descriptive feedback including information about how the essay is written, peers provided feedback including information about identification of the problem in the essay. The overarching look at the results suggests a potential complementary role for ChatGPT and students in the feedback process. Regarding the relationship between the quality of essays and the quality of the feedback provided by ChatGPT and peers, we found no overall significant relationship. These findings imply that the quality of the essays does not impact both ChatGPT and peer feedback quality. The implications of this study are valuable, shedding light on the prospective use of ChatGPT as a feedback source, particularly for complex tasks like argumentative essay writing. We discussed the findings and delved into the implications for future research and practical applications in educational contexts.
Human-AI interaction in text production increases complexity in authorship. In two empirical studies (n1 = 30 & n2 = 96), we investigate authorship and ownership in human-AI collaboration for personalized language generation. We show an AI Ghostwriter Effect : Users do not consider themselves the owners and authors of AI-generated text but refrain from publicly declaring AI authorship. Personalization of AI-generated texts did not impact the AI Ghostwriter Effect , and higher levels of participants’ influence on texts increased their sense of ownership. Participants were more likely to attribute ownership to supposedly human ghostwriters than AI ghostwriters, resulting in a higher ownership-authorship discrepancy for human ghostwriters. Rationalizations for authorship in AI ghostwriters and human ghostwriters were similar. We discuss how our findings relate to psychological ownership and human-AI interaction to lay the foundations for adapting authorship frameworks and user interfaces in AI in text-generation tasks.
Recently, Artificial Intelligence Generated Content (AIGC) has gained significant attention from society, especially with the rise of Generative AI (GAI) techniques such as ChatGPT, GPT-4 [ 165 ], DALL-E-3 [ 184 ], and Sora [ 137 ]. AIGC involves using AI models to create digital content, such as images, music, and natural language, with the goal of making the content creation process more efficient and accessible. Large-scale models have become increasingly important in AIGC as they provide better intent extraction and generation results. This survey provides a comprehensive review of the history of generative models and recent advances in AIGC, focusing on both unimodal and multimodal interaction. From the perspective of unimodality, we introduce the generation tasks and relative models of text and image. From the perspective of multimodality, we introduce the cross-application between the modalities mentioned above. Finally, the survey discusses the existing open problems and future challenges in AIGC. Overall, this survey serves as a valuable resource for individuals interested in understanding the background and secrets behind the impressive performance of AIGC techniques.
Importance: Timely tests are warranted to assess the association between generative artificial intelligence (GenAI) use and physicians' work efforts. Objective: To investigate the association between GenAI-drafted replies for patient messages and physician time spent on answering messages and the length of replies. Design, Setting, and Participants: Randomized waiting list quality improvement (QI) study from June to August 2023 in an academic health system. Primary care physicians were randomized to an immediate activation group and a delayed activation group. Data were analyzed from August to November 2023. Exposure: Access to GenAI-drafted replies for patient messages. Main Outcomes and Measures: Time spent (1) reading messages, (2) replying to messages, (3) length of replies, and (4) physician likelihood to recommend GenAI drafts. The a priori hypothesis was that GenAI drafts would be associated with less physician time spent reading and replying to messages. A mixed-effects model was used. Results: Fifty-two physicians participated in this QI study, with 25 randomized to the immediate activation group and 27 randomized to the delayed activation group. A contemporary control group included 70 physicians. There were 18 female participants (72.0%) in the immediate group and 17 female participants (63.0%) in the delayed group; the median age range was 35-44 years in the immediate group and 45-54 years in the delayed group. The median (IQR) time spent reading messages in the immediate group was 26 (11-69) seconds at baseline, 31 (15-70) seconds 3 weeks after entry to the intervention, and 31 (14-70) seconds 6 weeks after entry. The delayed group's median (IQR) read time was 25 (10-67) seconds at baseline, 29 (11-77) seconds during the 3-week waiting period, and 32 (15-72) seconds 3 weeks after entry to the intervention. The contemporary control group's median (IQR) read times were 21 (9-54), 22 (9-63), and 23 (9-60) seconds in corresponding periods. The estimated association of GenAI was a 21.8% increase in read time (95% CI, 5.2% to 41.0%; P = .008), a -5.9% change in reply time (95% CI, -16.6% to 6.2%; P = .33), and a 17.9% increase in reply length (95% CI, 10.1% to 26.2%; P < .001). Participants recognized GenAI's value and suggested areas for improvement. Conclusions and Relevance: In this QI study, GenAI-drafted replies were associated with significantly increased read time, no change in reply time, significantly increased reply length, and some perceived benefits. Rigorous empirical tests are necessary to further examine GenAI's performance. Future studies should examine patient experience and compare multiple GenAIs, including those with medical training.
Seven preregistered experiments demonstrate that when consumers believe emotional marketing communications are written by an AI (vs. a human), positive word of mouth and customer loyalty are reduced. Drawing from authenticity theory, we show that this “AI-authorship effect” is attenuated for factual (vs. emotional) messages (Study 2); when an AI only edits the communication (Study 3); when a communication is signed directly by an AI (Study 4); and when consumers believe that most marketing communications are written by AI (Study WA1). Importantly, when consumers believe a communication is reused (i.e., not originally written by the sender), the effect is reversed (Study 6). This “AI-authorship effect” is serially mediated by perceived authenticity (Studies 5 and 6) and moral disgust (Studies 1–6 and WA1). These findings are evidenced using both personalized and mass communications, different emotions, businesses and organizational employees, and both hypothetical and behavioral measures.
Can large language models, a form of artificial intelligence (AI), generate persuasive propaganda? We conducted a preregistered survey experiment of US respondents to investigate the persuasiveness of news articles written by foreign propagandists compared to content generated by GPT-3 davinci (a large language model). We found that GPT-3 can create highly persuasive text as measured by participants' agreement with propaganda theses. We further investigated whether a person fluent in English could improve propaganda persuasiveness. Editing the prompt fed to GPT-3 and/or curating GPT-3's output made GPT-3 even more persuasive, and, under certain conditions, as persuasive as the original propaganda. Our findings suggest that propagandists could use AI to create convincing content with limited effort.