By: Alfonso Gonzalez-Cano ‘27
Artificial Intelligence has exploded in popularity in past years, bringing millions of users to different LLMs (Large Language Models) such as Snapchat’s My AI, Google’s Gemini, and OpenAI’s ChatGPT. Part of AI’s appeal stems from its production of seemingly original text based on any prompt. However, does this text feature a unique voice?
In this study, I prompted the three different LLMs — ChatGPT, Gemini, and Snapcaht’s My AI — with twelve complex topics and recorded each AI’s opinion. Topics included RAM, derivatives, constellations, programming, temperature, value, AI learning, global warming, copyright, entropy, chance of death, and nuclear weapons. After each LLM responded, I analyzed the transcripts and recorded the most common words, bigrams (two-word phrases), and trigrams (three-word phrases) to determine if there was a change in vocabulary.
First and most surprisingly, all three LLMs had different opinions on the prompted topics. For example, when prompted about “value,” all three AIs gave different definitions: ChatGPT decided “value” is multifaceted and consists of personal, economic, and philosophical value. Gemini said “value” can be interpreted from both an economic and personal standpoint but excluded philosophy. Finally Snapchat’s My AI said “value” is completely dependent on context and varies between people.
As for their vernacular, each AI expressed a different vocabulary and synthesis of their response. Chat GPT was the most formal, utilizing sophisticated language and noun phrases. Further, it first provided an overview of the topic before defining the specifics and ending with a summary.
Table: Displays ChatGPT’s frequency of using the top 10 words, bigrams, and trigrams.
Contrastingly, Gemini used semi-formal vocabulary that one would say to an acquaintance. As for its structure, Gemini quickly defined the topic, elucidated more facts, and ended with its own opinion. Additionally, Gemini often encouraged the user to conduct their own research, even providing search links.
Finally, Snapchat’s My AI used friendly and informal vocabulary. The AI first defined the topic and then simply provided its opinion. Interestingly, My AI showed more emotion through frequent exclamations and questions in its writing and was also the only LLM to address the user by their name.
Table: Displays Gemini’s frequency of using the top 10 words, bigrams, and trigrams.
From the points enumerated above, different LLMs clearly have their own voice. But what does this mean for us users? Understanding the style of writing and structure of responses from each LLM helps us decide which program is best suited for our prompts. For example, My AI might be more apt to converse about day-to-day activities while ChatGPT might instead provide more precise information about complex topics. Further, since each LLM has a different viewpoint and thus bias, it is important to keep this in mind as we explore the diverse applications of AI.
