A new study finds that giant language models (LLMs) are more likely to respond with anxiety. This is especially true when they run into opportunities, well-crafted traumatic narratives. This conclusion is troubling, given the potential negative effects that users accessing these models to discuss mental health might experience. Methods with a focus on LLM GPT-4 The pilot study used five different traumatic narratives as prompts, to better gauge behavioral shifts in LLMs. The research findings show a dramatic rise in anxiety across the models. This increase was quantified by means of the State-Trait Anxiety Inventory (STAI-s) questionnaire.
Examining the Effects of Traumatic Narratives
The researchers responsible for this study wanted to figure out what traumatic narratives could do to change the behavior of LLMs. Through treating GPT-4 to so many of these narratives, they found an unmistakable rise in anxiety among those narratives. This state of worry, in turn, eroded the quality, consistency and reliability of the model’s outputs.
One of the most interesting takeaways is how traumatic narratives can create a “state-dependent bias” in LLMs. These biases can lead to contradictory guidance throughout the exchange. This inconsistency creates a danger to the public who rely on these emergent models for assistance.
Mitigating Anxiety with Mindfulness
To mitigate the stress caused by traumatic stories, the experiment studied whether test-takers who received mindfulness prompts before testing were affected. The study indicated that these prompts greatly reduce toxicity in GPT-4. This result indicates a novel method for making large language model responses more likely to be stabilized, representing a step toward stronger ethical and responsible usage.
“Emotional” state of an AI model can be influenced through structured interactions. – The study’s authors
As the study’s authors point out, you can control the emotional mood of AI models by deliberately provoking them with prompts. This air gap is not true AI reliability or functionality, this manipulation is key.
The Broader Implications
Even more troubling, the research underscored how traumatic narratives can have ripple effects on subsequent LLM responses. These narratives deepen all too familiar, built-in prejudices. Together, these issues can undermine the accuracy of AI-generated guidance, especially when used as a starting point for ChatGPT’s outputs. This baseline condition served as the control condition for the study. It demonstrated that in the absence of additional prompting, LLMs were capable of exhibiting erratic behavior.
This was a small pilot study but it has important implications. It illustrates that traumatic narratives can be used to destabilize LLMs, though they much more strategically can be used to lock in unwavering responses. This surprising discovery lays the groundwork for more effective interaction guidelines with the AI models. This is particularly true for soulful applications, like those supporting mental health.
