An official website of the United States government
Recently, there have been significant advances and wide-scale use of generative AI in natural language generation. Models such as OpenAI’s GPT3 and Meta’s LLaMA are widely used in chatbots, to summarize documents, and to generate creative content. These advances raise concerns about abuses of these models, especially in social media settings, such as large-scale generation of disinformation, manipulation campaigns that use AI-generated content, and personalized scams. We used stylometry (the analysis of style in natural language text) to analyze the style of AI-generated text. Specifically, we applied an existing authorship verification (AV) model that can predict if two documents are written by the same author on texts generated by GPT2, GPT3, ChatGPT and LLaMA. Our AV model was trained only on human-written text and was effectively used in social media settings to analyze cases of abuse. We generated texts by providing the language models with fanfiction snippets and prompting them to complete the rest of it in the same writing style as the original snippet. We then applied the AV model across the texts generated by the language models and the human written texts to analyze the similarity of the writing styles between these texts. We found that texts generated with GPT2 had the highest similarity to the human texts. Texts generated by GPT3 and ChatGPT were very different from the human snippet, and were similar to each other. LLaMA-generated texts had some similarity to the original snippet but also has similarities with other LLaMA-generated texts and texts from other models. We then conducted a feature analysis to identify the features that drive these similarity scores. This analysis helped us answer questions like which features distinguish the language style of language models and humans, which features are different across different models, and how these linguistic features change over different language model versions. The dataset and the source code used in this analysis have been made public to allow for further analysis of new language models.
more »
« less
Human heuristics for AI-generated language are flawed
Human communication is increasingly intermixed with language generated by AI. Across chat, email, and social media, AI systems suggest words, complete sentences, or produce entire conversations. AI-generated language is often not identified as such but presented as language written by humans, raising concerns about novel forms of deception and manipulation. Here, we study how humans discern whether verbal self-presentations, one of the most personal and consequential forms of language, were generated by AI. In six experiments, participants (N = 4,600) were unable to detect self-presentations generated by state-of-the-art AI language models in professional, hospitality, and dating contexts. A computational analysis of language features shows that human judgments of AI-generated language are hindered by intuitive but flawed heuristics such as associating first-person pronouns, use of contractions, or family topics with human-written language. We experimentally demonstrate that these heuristics make human judgment of AI-generated language predictable and manipulable, allowing AI systems to produce text perceived as “more human than human.” We discuss solutions, such as AI accents, to reduce the deceptive potential of language generated by AI, limiting the subversion of human intuition.
more »
« less
- Award ID(s):
- 1901151
- PAR ID:
- 10463192
- Date Published:
- Journal Name:
- Proceedings of the National Academy of Sciences
- Volume:
- 120
- Issue:
- 11
- ISSN:
- 0027-8424
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
AI language technologies increasingly assist and expand human communication. While AI-mediated communication reduces human effort, its societal consequences are poorly understood. In this study, we investigate whether using an AI writing assistant in personal self-presentation changes how people talk about themselves. In an online experiment, we asked participants (N=200) to introduce themselves to others. An AI language assistant supported their writing by suggesting sentence completions. The language model generating suggestions was fine-tuned to preferably suggest either interest, work, or hospitality topics. We evaluate how the topic preference of a language model affected users’ topic choice by analyzing the topics participants discussed in their self-presentations. Our results suggest that AI language technologies may change the topics their users talk about. We discuss the need for a careful debate and evaluation of the topic priors built into AI language technologies.more » « less
-
As text generated by large language models proliferates, it becomes vital to understand how humans engage with such text, and whether or not they are able to detect when the text they are reading did not originate with a human writer. Prior work on human detection of generated text focuses on the case where an entire passage is either human-written or machine-generated. In this paper, we study a more realistic setting where text begins as human-written and transitions to being generated by state-of-the-art neural language models. We show that, while annotators often struggle at this task, there is substantial variance in annotator skill and that given proper incentives, annotators can improve at this task over time. Furthermore, we conduct a detailed comparison study and analyze how a variety of variables (model size, decoding strategy, fine-tuning, prompt genre, etc.) affect human detection performance. Finally, we collect error annotations from our participants and use them to show that certain textual genres influence models to make different types of errors and that certain sentence-level features correlate highly with annotator selection. We release the RoFT dataset: a collection of over 21,000 human annotations paired with error classifications to encourage future work in human detection and evaluation of generated text.more » « less
-
As text generated by large language models proliferates, it becomes vital to understand how humans engage with such text, and whether or not they are able to detect when the text they are reading did not originate with a human writer. Prior work on human detection of generated text focuses on the case where an entire passage is either human-written or machine-generated. In this paper, we study a more realistic setting where text begins as human-written and transitions to being generated by state-of-the-art neural language models. We show that, while annotators often struggle at this task, there is substantial variance in annotator skill and that given proper incentives, annotators can improve at this task over time. Furthermore, we conduct a detailed comparison study and analyze how a variety of variables (model size, decoding strategy, fine-tuning, prompt genre, etc.) affect human detection performance. Finally, we collect error annotations from our participants and use them to show that certain textual genres influence models to make different types of errors and that certain sentence-level features correlate highly with annotator selection. We release the RoFT dataset: a collection of over 21,000 human annotations paired with error classifications to encourage future work in human detection and evaluation of generated text.more » « less
-
Artificial Intelligence (AI) is a transformative force in communication and messaging strategy, with potential to disrupt traditional approaches. Large language models (LLMs), a form of AI, are capable of generating high-quality, humanlike text. We investigate the persuasive quality of AI-generated messages to understand how AI could impact public health messaging. Specifically, through a series of studies designed to characterize and evaluate generative AI in developing public health messages, we analyze COVID-19 pro-vaccination messages generated by GPT-3, a state-of-the-art instantiation of a large language model. Study 1 is a systematic evaluation of GPT-3's ability to generate pro-vaccination messages. Study 2 then observed peoples' perceptions of curated GPT-3-generated messages compared to human-authored messages released by the CDC (Centers for Disease Control and Prevention), finding that GPT-3 messages were perceived as more effective, stronger arguments, and evoked more positive attitudes than CDC messages. Finally, Study 3 assessed the role of source labels on perceived quality, finding that while participants preferred AI-generated messages, they expressed dispreference for messages that were labeled as AI-generated. The results suggest that, with human supervision, AI can be used to create effective public health messages, but that individuals prefer their public health messages to come from human institutions rather than AI sources. We propose best practices for assessing generative outputs of large language models in future social science research and ways health professionals can use AI systems to augment public health messaging.more » « less
- Free Publicly Accessible Full Text
-
Accepted Manuscript1.0
- Journal Article:
- https://doi.org/10.1073/pnas.2208839120
