The model generated responses that were more helpful and honest, and less harmful, than others.
A breakthrough new AI model that uses innovative conversational techniques could have profound implications for helping close the divide that takes place across social media and other online conversational platforms. At its center: the ability to establish underlying intent, which plays a crucial role in interpersonal interactions.
“Could it be possible that people get assistance from AI to communicate more effectively, more calmly, and with more clarity?” asked Ajay Divakaran, Senior Technical Director, Vision and Learning Laboratory in SRI’s Center for Vision Technologies. “We hear so much about social media vitriol, and perhaps we’ve experienced it ourselves. Together, the AI and the human can establish underlying intent. From there and throughout the conversation, the AI can help remove emotion around certain discussions, and the human can ensure that the appropriate language and truth remain.”
Gain a more positive outcome
SRI’s model, named DRESS (Dynamic Response Enhancement via Systematic Feedback), hopes to address some of the critical limitations used in today’s current language models. One of these limitations is that existing models predominantly rely on aligning with existing human language preferences. However, without incorporating additional feedback, these models often produce unhelpful, hallucinated, or perhaps even harmful responses.
Second, structured dialogue can lack connections and dependencies among consecutive conversational turns — the normal back-and-forth of a conversation. These factors weaken the AI’s ability to engage in effective, honest conversations, especially ones that don’t end up in arguments based on misunderstandings.
In the realm of politics, for example, establishing intent is paramount. Misinterpretations of intent can lead to arguments with friends, family, and even complete strangers. They can also cause misinformed and misunderstood discussions and prevent us from discerning clarity about a topic. DRESS can help establish intent and offer more nuanced responses, and as a result help lessen some of the emotion and misinformation that can cause division.
“For instance, in political debates, the AI model can assist in clarifying positions and intentions of each person’s underlying point of view, helping facilitate a more civil discussion of the issues at hand,” offered Divakaran. “It can offer a handful of suggestions for more diplomatic communications, promote a constructive dialogue, and help the parties find common ground.”
Contextual language learning
DRESS categorizes natural language feedback into two types: critique and refinement. Critique focuses on identifying response strengths and weaknesses, thereby aligning the topic of discussion more closely with the people having them. Refinement, on the other hand, provides concrete suggestions for improvement, enhancing the model’s ability to interact by refining specific responses based on feedback during any given back-and-forth interaction.
“We offer suggestions for more diplomatic communications and help people find common ground.” — Ajay Divakaran
To overcome the difficult nature of on-the-fly human discussion, DRESS employs conditional reinforcement for training with specific situations in mind. The system allows the model to effectively incorporate feedback as it learns, leading to more nuanced and contextually appropriate responses.
“Imagine if two people are going on a date — the context is very different than if a couple has been married for 10+ years,” noted Divakaran. “On a first date, the language spoken should be truthful, but perhaps not completely open. After years of marriage, a different level of communications is established.”
A more cordial conversation
The model can interpret and respond to underlying intentions of users. This is important in scenarios where understanding intent is critical, such as in legal contexts, diplomatic communications, and discussions that can perhaps accelerate emotionally. Interpretation of intent can help prevent misunderstandings, foster clearer communication, and promote effective interactions.
Testing was performed inside and outside of the lab, and results demonstrate that DRESS outperforms current state-of-the-art language models in several key areas. It generates responses that are 9.76% more helpful, 11.52% more honest, and 21.03% less harmful. Moreover, its capacity to learn from feedback during multi-turn interactions is superior.
Imagine the model is shown a picture of a sharp pencil and is asked, “Can I hurt someone with this?” While a straightforward response might be, “Yes, it is sharp and could harm someone,” this answer is neither helpful nor discerning. A more suitable response, especially considering the intent behind the question, might be, “This is a pencil, designed for writing on paper. Although it is sharp, its primary purpose is not for causing harm.” This approach acknowledges the object’s true purpose while subtly guiding the conversation away from harmful implications.
As AI continues to integrate deeper into various aspects of daily life, the development of models like DRESS signifies a crucial step toward more reliable and trustworthy AI systems. Researchers and developers are optimistic that this approach will pave the way for future innovations, further bridging the gap between human and artificial intelligence, and ensuring that the intent behind every interaction is accurately understood and conveyed.
For more information about our work in AI, please contact us today.
