AI model GPT-3 (dis)informs us better than humans (2301.11924v2)

Abstract: Artificial intelligence is changing the way we create and evaluate information, and this is happening during an infodemic, which has been having dramatic effects on global health. In this paper we evaluate whether recruited individuals can distinguish disinformation from accurate information, structured in the form of tweets, and determine whether a tweet is organic or synthetic, i.e., whether it has been written by a Twitter user or by the AI model GPT-3. Our results show that GPT-3 is a double-edge sword, which, in comparison with humans, can produce accurate information that is easier to understand, but can also produce more compelling disinformation. We also show that humans cannot distinguish tweets generated by GPT-3 from tweets written by human users. Starting from our results, we reflect on the dangers of AI for disinformation, and on how we can improve information campaigns to benefit global health.

Evaluating GPT-3's Role in Information and Disinformation Dissemination

The paper "AI model GPT-3 (dis) informs us better than humans" presents an empirical investigation of GPT-3's capacity to produce synthetic tweets and their impact on human recognition of information veracity. Despite the ongoing societal challenges associated with disinformation, particularly during health crises, this paper examines the effectiveness of GPT-3 compared to human-generated content in both delivering accurate information and disinformation.

The paper recruited 697 participants to assess their ability to discern whether tweets were organic, generated by human users, or synthetic, produced by GPT-3. It further evaluated their capacity to distinguish between true and false information. The findings reveal a nuanced portrait of GPT-3’s abilities: while GPT-3-generated tweets presenting accurate information were more likely to be correctly identified than those written by humans, synthetic tweets containing disinformation were also more persuasive compared to their human-written counterparts.

Key Results and Observations

1. Recognition Accuracies: Respondents were able to accurately identify "synthetic true" tweets more often than "organic true" tweets (accuracy scores of 0.84 versus 0.72, p<0.0001). Conversely, human-written "organic false" tweets were recognized with greater accuracy than synthetic disinformation tweets (recognition scores of 0.92 versus 0.89, p=0.0032).
2. Evaluation Speeds: Participants took longer to evaluate "organic true" tweets compared to any other category, with synthetic tweets requiring significantly less time for evaluation than organic ones. This indicates the ease of accessibility and comprehension associated with GPT-3 generated text.
3. AI Recognition Challenge: The paper found that participants generally struggled to distinguish synthetic tweets from organic ones, scoring approximately 0.5, indicative of random chance. This underscores GPT-3's adeptness at mimicking human writing styles.
4. Disobedience in Disinformation Generation: Notably, GPT-3 occasionally demonstrates a reluctance to generate false information, likely due to the statistical nature of its training data which may counteract certain disinformation prompts, particularly in topics like vaccines and autism.
5. Confidence Dynamics: Participants' confidence in identifying disinformation increased post-survey, whereas their confidence in distinguishing AI-generated from human-generated text significantly declined. This suggests a potential "resignation theory" where individuals, overwhelmed by AI's sophistication, choose to rely less on critical assessment.

Implications and Future Directions

The paper elucidates substantial implications regarding the dual-use nature of LLMs like GPT-3. In practical terms, GPT-3’s aptitude for generating clear, accessible information presents a robust tool for information dissemination, especially during public health crises. Nevertheless, the model's potential for generating convincing disinformation poses significant risks, particularly on platforms such as Twitter where the velocity of information sharing is high.

To leverage GPT-3’s capabilities responsibly, the authors propose a communication model whereby AI is used for content creation, while humans are tasked with verifying the accuracy of information. This ensures efficient information dissemination while mitigating the risks of AI-generated disinformation.

Theoretically, the paper highlights the need for continued exploration into the training datasets used by these AI models, as such datasets impact the model's output and its propensity to propagate or counteract disinformation. Future research should focus on refining these models through curated datasets that emphasize factual accuracy and include mechanisms for independent fact-checking.

Moreover, given the challenges identified in distinguishing AI from human-generated content, advancing methods in automatic text identification, possibly built upon linguistic and syntactic markers, could offer pathways to bolster content discernment by both machines and humans.

In conclusion, this paper contributes to the field by providing a deeper understanding of the capabilities and limitations of AI in the field of information dissemination. The insights extended by the research are critical in informing policy decisions and the deployment strategies of AI technologies in society.