New research reveals that while humans excel in empathy-based tasks, AI models rapidly adapt to instructions, blurring the line between machine and human-written content.
Research: Trying to be human: Linguistic traces of stochastic empathy in language models
*Important notice: arXiv publishes preliminary scientific reports that are not peer-reviewed and, therefore, should not be regarded as definitive, used to guide development decisions, or treated as established information in the field of artificial intelligence research.
In an article recently submitted to the arXiv preprint* server, researchers tested how empathy and the incentive to appear human influence the ability to differentiate between artificial intelligence (AI) and human-written content. In two studies, participants provided relationship advice or descriptions, while others judged whether the text was human or AI-generated. The results showed that humans excelled in empathetic tasks, but the ability of AI models to mimic human-like writing significantly improved when instructed to appear human. AI used conversational, informal language to mimic human empathy.
Related Work
Past work suggested that humans need help differentiating AI-generated content from human-written text, relying on flawed cues. Studies showed that AI-generated content, such as bios or profiles, often appeared more human-like than content people wrote. However, previous research did not consider whether human authors were incentivized to appear human. Additionally, the limited context of self-presentations may have constrained the expression of true human language ability.
AI vs. Human Perception
In study 1, researchers explored how task awareness impacts the perception of human versus AI-written relationship advice. Participants were split into two groups: the adversarial condition, where they were instructed to write as human-like as possible, and the naïve condition, where no guidance was given.
A total of 530 human-generated texts were analyzed alongside generative pre-trained transformer 4 (GPT-4)-generated texts, and a separate group evaluated these on a 5-point scale to determine if they were AI or human-written. The results showed that human-written texts were consistently rated more human-like, but AI-generated texts significantly improved human-like perception under adversarial instructions.
The adversarial condition led to a 39.3% increase in human-likeness for AI-generated texts, while it had little effect on human writers, increasing their perceived humanness by only 2.6%. Interestingly, the study also found that humans became worse at distinguishing between human and AI-generated texts under adversarial conditions.
Despite AI's boost in human-likeness, human-written advice still held an overall advantage. The study highlights AI's capacity to adapt its writing style when prompted but also underscores the inherent challenge humans face in differentiating between human and AI-generated content, especially when AI is instructed to appear more human.
Empathy Enhances AI Text Generation
In study 2, researchers introduced a manipulation that required participants to either provide empathetic relationship advice or describe a relationship, resulting in a 2x2x2 between-subjects design. The team evaluated 523 human-written texts and an equal number of GPT-4-generated texts.
Participants were randomly assigned to one of four conditions: relationship advice/control, relationship advice/adversarial, relationship description/control, or relationship description/adversarial. The human-generated and GPT-4-generated texts were rated on a 5-point scale, with human-written texts consistently rated more human-like.
Results showed a significant interaction between task type and adversarial conditions. In empathy-driven tasks, adversarial instructions improved the perceived human-likeness of large language model (LLM)-generated texts by 41.0%, while human-written texts remained largely unaffected (-1.4%).
For non-empathy tasks like relationship descriptions, the gap between human and AI-written texts narrowed, with adversarial instructions having less impact. This study highlights AI's ability to mimic human writing more effectively in empathetic tasks when provided with specific guidance, while humans showed minimal ability to increase their perceived humanness.
Linguistic Strategies for Human-Likeness
In study 3, researchers examined the linguistic strategies employed by humans and LLMs to enhance their human-like qualities, utilizing data from study 2 as a baseline. The analysis focused on relationship advice in both naive and adversarial conditions, assessing differences in linguistic features through n-gram differentiation and psycholinguistic metrics provided by the linguistic inquiry and word count (LIWC) software.
The study found that LLMs were particularly effective at modifying their advice to appear more human-like, while humans struggled to make such adjustments. Notable differences included LLMs using phrases like "dear friend" and informal language, while human texts featured more common verbs and first-person singular pronouns.
The findings revealed that human-written relationship advice showed minimal variation between conditions, indicating difficulty adapting to the adversarial context. In contrast, LLMs significantly shifted their style when aiming for a more human presentation, using informal expressions like "hey" and conversational markers such as "really sorry" and "yeah." Analysis of spelling mistakes and vocabulary frequency further demonstrated the LLM's proficiency, consistently producing texts with fewer errors.
Compassion in Text Generation
The study examined the ability of LLMs to generate human-like text, particularly in tasks requiring empathy, where human participants outperformed LLMs. However, LLMs successfully adapted their language style when prompted to appear more human, narrowing the perceived gap in humanness.
Despite LLMs mimicking empathetic writing, they lacked genuine understanding, operating instead on learned statistical patterns, often referred to as "stochastic empathy." Future research could explore empathy's role across various writing tasks and languages, enhancing the knowledge of LLM capabilities in generating authentic human-like content.
Conclusion
To sum up, LLMs have been claimed to produce text indistinguishable from human writing. However, the experiments revealed that LLMs, rather than humans, significantly adjusted their writing style when prompted to be more human-like. LLMs effectively utilized representations of empathetic language characterized by informal, simple, self-referential, and present-focused expressions, while humans struggled to adapt their writing style, showing only minimal improvement.
*Important notice: arXiv publishes preliminary scientific reports that are not peer-reviewed and, therefore, should not be regarded as definitive, used to guide development decisions, or treated as established information in the field of artificial intelligence research.
Journal reference:
- Preliminary scientific report.
Kleinberg, B., Zegers, J., Festor, J., Vida, S., Präsent, J., Loconte, R., & Peereboom, S. (2024). Trying to be human: Linguistic traces of stochastic empathy in language models. ArXiv. https://arxiv.org/abs/2410.01675