Fun with Generative AI

Background

This work was developed in the CTL Seminar, A Deep Dive into Advising and AI, in Spring 2024. In this example set, one short introduction document about LaGuardia Community College has been translated into the top 23 languages spoken at LaGuardia. The following AI tools were used to generate examples:

  • Google Transalte
  • Gemeni
  • ChatGPT 3.5
  • ChatGPT 4

The goal of the project was to translate information about the college (such as the orientation materials) into languages that are commonly spoken by LaGuardia students and their family members. Only a sample document is shown on this page.

These are the top 24 languages spoken at LaGuardia Community College.

Languages Spoken at LaGuardia (Fall 2022)
Data obtained from Institutional Research. 89% did not respond to the language item in the CUNY survey and were excluded.
# Language Fall 2022 # Language Fall 2022
1English60.5%13French0.4%
2Spanish18.6%14Urdu0.4%
3Bengali3.6%15Punjabi0.4%
4Chinese3.5%16Portuguese0.3%
5Nepali1.4%17Igbo0.3%
6Haitian Creole1.2%18Hindi0.2%
7Tibetan1.2%19Burmese0.2%
8Tagalog1.1%20Pilipino0.2%
9Arabic1.0%21Uzbek0.2%
10Korean1.0%22Thai0.2%
11Polish0.6%23Russian0.1%
12Albanian0.4%24Japanese0.1%

The whole translation process took about 1.5 hours and it cost $2.84.

Below is a screenshot of the list of the top 24 languages spoken at LaGuardia Community College. The complete slide deck, developed in CTL Seminar A Deep Dive into Advising and AI (Spring 2024), is available here

📄 Google Translate
Sample Prompt
Using this text and other available resources, explain to [LANGUAGE] speakers why they should go to a community college. 
Use [LANGUAGE]. [COPY THE ENGLISH VERSION OF THE LAGUARDIA DESCRIPTION]

Overall, the quality of Google Translation was fairly limited. There are a lot of awkward expressions, a lack of fluency, and, in some cases, outright mistakes in translation. Native speakers of those languages would definitely notice that the text has been translated with machine translation.

🤖 Gemini
Sample Prompt
Using this text and other available resources/information, explain to [LANGUAGE] speakers who are not familiar with 
the American higher education why they should go to a community college. Use [LANGUAGE].

The following outputs were generated by the command above using Gemini (May, 2024). There is definitely a great deal of fluency that was lacking in Google Translation. It is interesting that Gemini outperforms Google Translation, which Google has spent a great deal of time and resources to bring to its current level.

💬 ChatGPT 3.5
Sample Prompt
Using this text and other available resources/information, explain to [LANGUAGE] speakers who are not familiar with 
the American higher education why they should go to a community college. Use [LANGUAGE]. [+AUDIO GENERATION OPTION]

The outputs below were generated by the command above using ChatGPT 3.5. While the quality of translation is equivalent to that of Gemini, a notable function of ChatGPT 3.5 is its audio generation model (a.k.a. Whisper). The audio quality of the speech is natural, although it struggles with numbers and English loanwords. It costs a fraction of a cent to generate an audio file.


🖥️ ChatGPT 4o
Sample Prompt
Using this text and other available resources/information, explain to [LANGUAGE] speakers who are not familiar with 
the American higher education why they should go to a community college. Use [LANGUAGE]. [+AUDIO GENERATION OPTION]

ChatGPT-4o was introduced in May 2024. The outputs below were generated by the command above using ChatGPT-4o. The quality of the translation and text-to-speech was a significant improvement. The text is no longer direct translation from the original document and offers a great deal of fluency and natural flow in the translated languages. In addition, the token-based cost was significantly reduced, allowing users to add more text (contexts) to enrich the output.

🎰 ChatGPT 5
Sample Prompts
Please write a Python script to run an API command to translate a document into the following languages. 
"Albanian" "Arabic" "Bengali" "Burmese" "Traditional Chinese" "Simplified Chinese" "French" 
"Haitian Creole" "Hindi" "Igbo" "Japanese" "Korean" "Nepali" "Pilipino" "Polish" "Portuguese" 
"Punjabi" "Russian" "Spanish" "Tagalog" "Thai" "Tibetan" "Urdu" "Uzbek"

for lang in "Albanian" "Arabic" "Bengali" "Burmese" "Traditional Chinese" "Simplified Chinese" "French" 
"Haitian Creole" "Hindi" "Igbo" "Japanese" "Korean" "Nepali" "Pilipino" "Polish" "Portuguese" 
"Punjabi" "Russian" "Spanish" "Tagalog" "Thai" "Tibetan" "Urdu" "Uzbek"; do python_script.py 
--audio --websearch --reasoning "medium" -p "Translate the following document into $lang. 
Preserve the original style and produce a faithful, equivalent translation in $lang. After 
completing the translation, explain to $lang speakers who are not familiar with the American 
higher-education system why attending a community college can be beneficial. Tailor this 
explanation to the specific circumstances of the $lang-speaking community in New York City 
and keep it within a few paragraphs. Write in $lang. Generate only output and do not 
introduce any additional commentary before the output." --file english.txt; done

GPT-5 was introduced in May 2025, and it was claimed to have doctoral-student-level reasoning skills. The model comes with a new concept called "reasoning," which basically loops questions and responses until it reaches a highly refined response. The most advanced reasoning is extremely expensive and time-consuming — it takes about 10–20 minutes, but it can resolve a very difficult question (about $10 with the API). The issue of hallucinations is practically resolved with this reasoning model and a new live web-search function. OpenAI and its major competitors (Gemini and Claude) started to focus on image and video creation, and TTS (text-to-speech) hasn't seen much progress since 2024. At OpenAI, the latest model remains GPT-4o-tts. A few new AI companies (such as ElevenLabs) have started to offer high-quality text-to-speech services, but they are usually very expensive.

Another major advancement in LLMs is their ability to code. Until GPT-4o, we needed to code (in Python or similar computer languages) if we wanted to run a large number of inquiries. With GPT-5, LLMs can generate API code by themselves, so we won't need to write a program at all. The prompts for this section show a two-step process to reflect this change. Also, the prompt now asks for more than simple translation so that the model can show its ability beyond transaltion.

The quality of translation has plateaued with GPT-4o, and there is not much room to improve since then. Therefore, the prompt has been revised to add a few additional pieces of information to the response. The quality of responses is impressive—even scary to see what the large language model can accomplish now.

ChatGPT AI Language Tutors

Sample Usage

python tts_generate.py --model gpt-4o-mini-tts --lang ja --input hare.txt
📄 Text + 🔊 Audio

Last update: