Limitations of Tonality with GPT

Tonality Limitations with GPT

A colleague of mine once told me about a time that someone had asked him on a lunch date. As the day of the date arrived, he texted the person that he was to meet, “I’m loking forward to seeing you today.” Almost immediately after he sent the text, he noticed his misspelling and sent another text: “And by loking I meant looking.” His date sent back a rather surprising reply: “I am way too busy to be someone’s eye candy!” He was surprised by her response to say the least! Ever the diplomat, he responded, “Let me clarify. I anticipate that it will be a pleasant experience to be in your company at lunch today.”

the audience and their age as well as their technical proficiency, and the piece’s style. While GPT can be given general instructions in the prompt as to the tone of the content to be generated, Textmetrics offers the ability to apply a detailed set of rules once, across all communications, so that the preferred tone of the generated content is consistent from one communication to the next.

As of the last update in January 2022, GPT 3.5 does not have explicit tonality or emotion modeling features. It generates text based on patterns learned during training and doesn’t understand or generate text with specific emotional tones. While users may achieve a sense of tonality by shaping their inputs and contexts in a way that guides the model to respond in a desired manner, the model lacks an understanding of emotions or tonality; it is generating responses based on statistical patterns present in the data it was trained on. We think that Textmetrics’ capabilities are the perfect compliment to GPT: Textmetrics provides the consistency in tone that GPT -generated content often lacks. Textmetrics, being a rules-based NLP, applies the business rules with which it has been programmed to shape GPT composition output, yielding consistent, reliable content.

We have found that GPT Tonality expression is limited in the following ways:

• Lack of Genuine Understanding: GPT doesn’t truly understand emotions or tonality. It relies on patterns learned during training and doesn’t have an inherent understanding of emotional nuance. Textmetrics applies a detailed set of rules once, across all communications, so that the preferred tone of the generated content is consistent and clear.
• Sensitive to Input Phrasing: The model’s output can be sensitive to slight changes in input phrasing. Small alterations might lead to different responses, and it may not consistently maintain a specific tonality across varied inputs.
• Prone to Bias and Inappropriate Content: The model can generate biased or inappropriate content based on the biases present in the training data. Users need to carefully craft inputs to guide the model away from unintended tonalities or biases.
• Contextual Dependence: The model’s responses are heavily dependent on the context provided by the user. Without proper context, it might not generate responses with the desired tonality.
• Inability to Adapt to User Feedback: GPT lacks the ability to adapt its tonality based on explicit user feedback. Users have to iteratively refine inputs to achieve the desired tone.
• Difficulty in Handling Complex Tonal Shifts: Generating text with consistent tonality over long or complex interactions can be challenging. The model will struggle to maintain a specific tone across a conversation.
• Generic Responses: In certain situations, the model may provide generic or neutral responses rather than expressing a specific tonality. This can limit its effectiveness in applications requiring nuanced emotional expression. Additionally, GPT is prone to over reliance on certain transitional phrases and is unable to cite its sources.

Textmetrics can bring consistency to the tone of all of your written content, whether generated by an AI or a human. If you would like to learn more, please feel free to reach out; we anticipate that it will be a pleasant experience to hear from you!