Unlocking the future of translation for local journalism: Lessons from CPI’s AI experiment
In December 2023, Centro de Periodismo Investigativo (CPI), an award-winning newsroom in Puerto Rico, received support from the American Journalism Project’s Product & AI Studio to expose more English-speaking audiences to their work. While AI translation tools held some promise, they didn’t always capture cultural nuances — an important ingredient in CPI’s reporting. CPI’s evolving, meticulously documented experiments with AI-driven translations offer a masterclass in how local news organizations can embrace innovation without sacrificing integrity.
Bridging languages, expanding impact
CPI’s mission to serve both Puerto Rican residents and the Puerto Rican diaspora demanded a solution to a longstanding challenge: how to faithfully translate investigative reports from Spanish to English while maintaining the depth, nuance, and accuracy of the original stories. Enter Noel Algarín Martínez, CPI’s English editor, who spearheaded a months-long journey to harness AI as a translation partner.
For Algarín Martínez, the stakes were high. Beyond overcoming linguistic barriers, CPI needed to ensure every translated article met their high standards of community service and investigative journalism, adhered to AP Style and preserved cultural context.
From simple prompts to a translation powerhouse
Algarín Martínez’s journey began with cautious experimentation. When he tested five translation tools, including Google Translate, DeepL and ChatGPT, they all revealed sometimes glaring gaps in reliability. ChatGPT stood out as the better tool, but not without caveats. Early translations were riddled with unpredictable omissions and stylistic inconsistencies, he said.
Liam Andrew, AJPs Product and AI Studio technology lead, and Algarin Martínez worked together to refine prompts with precision. A breakthrough finally came when they designed a detailed, context-aware prompt:
“Translate the following Spanish news article into English, adhering to AP Style guidelines. Maintain cultural nuance while ensuring clarity and proper grammar. Adapt references for an English-speaking audience.”
Algarín Martínez’s work sought to find thoughtful ways to use AI to help elevate storytelling in English, not just provide workflow shortcuts. This nuanced approach, coupled with iterative feedback loops, transformed the quality of translations. CPI developed a “Bilingual Bridge” AI assistant, blending AI capabilities with human expertise to tackle the complex task of cultural adaptation.
Developing this AI assistant, however, was not an easy task. Algarín Martínez had to try varied prompting techniques like attaching the full AP Stylebook or putting the prompt into a GPT rather than the context window.
Ultimately they found that more detailed prompts were confusing ChatGPT and it was being “too creative” with translations. That’s when Andrew suggested they try OpenAI’s API playground, rather than ChatGPT. Working in the API, Algarín Martínez reduced the “temperature,” i.e. the level of randomness/creativity in its answers, and created more complex prompts.
They were then able to build a strong product by lowering the temperature to zero while still prompting the model. This helped to avoid being too literal and instead prioritized readability and fluency to “feel as though the article was originally written in English.” The prompt also included safeguards against hallucinations, encouraged localization, and promoted dynamic collaboration with the human editor.
Lessons for every newsroom
Near the end of 2024, Algarín Martínez published a document detailing each experiment and lesson. Here are his top recommendations for newsrooms implementing AI-generated translations.
- Iteration is key: Careful testing and refining underscores the importance of iteration. AI tools are not plug-and-play; they require consistent feedback to meet journalistic standards.
- Human oversight is non-negotiable: AI won’t replace human translators at CPI, Algarín Martínez said. It enhanced productivity and preserved quality. Their translator is now serving as more of an editor, which is serving CPI better.
- Cultural sensitivity matters: By emphasizing the preservation of Puerto Rican nuances, CPI ensured their stories resonated with English-speaking audiences without losing their authenticity.
- Transparency builds trust: As part of CPI’s workflow, they are implementing full disclosure of AI involvement to help maintain trust with their readers.
A playbook for the future
This experiment is more than a success story — it’s a blueprint. Algarín Martínez crafted an AI-translation playbook, which offers actionable insights for local newsrooms eager to expand their reach through multilingual content. From tailoring prompts to integrating AI into workflows, the insights he shares in the playbook have been shared with AJP portfolio newsrooms and are now available for anyone to utilize.
Local news leaders often face the challenge of innovating without compromising journalistic integrity and ethics. CPI’s work thoughtfully explores how to tackle this challenge. Their journey demonstrates that when technology and human expertise converge, the result can be transformative.
Click here to download CPI’s playbook.
—
Dorrine Mendoza is the product and partnerships lead for the American Journalism Project’s Product & AI Studio, which explores the smart application of generative AI and other technology within local news. This Studio, made possible by OpenAI with additional support from the Patrick J. McGovern Foundation, shares regular updates of its learnings on AJP’s website.
