By JoAnne Wadsworth, Communications Consultant, G20 Interfaith Forum
———
On June 25, 2026, the G20 Interfaith Forum, together with the International Academy for Multicultural Cooperation, held a webinar entitled “AI and Translation,” the second in an eight-part series exploring AI and faith. Speakers included Steve Richardson, a computational linguist and professor of computer science at Brigham Young University who has worked on machine translation for roughly fifty years, including at IBM Research and Microsoft Research, and who now leads BYU’s Matrix Lab; Charles Canney, Senior Manager of Digital Media at BYU Brand and Creative and head of the BYU Speeches Translation Initiative; and Ammon Shurtz, a PhD student in AI and translation and the lead research assistant of the Matrix Lab. Marianna Richardson, host of the forum’s AI webinar series, moderated the discussion. Watch the full webinar here.
Steve Richardson
Richardson opened by tracing his half-century in the field, from early “machine translation” in the 1970s through roles at IBM Research and Microsoft Research and a decade applying the technology at The Church of Jesus Christ of Latter-day Saints, to his current work leading the Matrix Lab at BYU. He noted that the neural technology now powering large language models such as ChatGPT, Claude, and Gemini was first developed for translation about a dozen years ago and entered wide use in Google Translate and Microsoft Translator around 2016. Translation, he said, remains one of AI’s primary tasks.
His central theme was a vision of a world without language barriers.
“Imagine a world in which, no matter what language you read, spoke, or even signed, you could communicate with another person anywhere in the world.”
Reaching that vision, Richardson cautioned, is far from finished. Today’s models handle the first couple dozen languages well and a few dozen more adequately, but quality drops off sharply beyond roughly fifty to seventy languages, leaving some 6,900 of the world’s 7,000-plus languages still to address. Half are unwritten and require speech technology, and hundreds of distinct sign languages add further complexity, since sign languages are unrelated to spoken ones—British and American Sign Language users, he noted, cannot even understand one another. Closing these gaps, he argued, could unlock humanitarian aid, emergency and disaster response, medical care for underserved communities, and economic opportunity.
He outlined the working pipeline—gathering paired data, building models, and deploying them in applications—and demonstrated a speech-translation system that recognized his spoken English, translated it into an African click language, and voiced the result aloud. He also described PathSay, an ongoing effort that records BYU Pathway students reading sentences in their native languages; it has gathered roughly 2,200 hours of audio across 30 languages and continues to expand.
Charles Canney
Canney, who has spent 46 years at BYU, framed his remarks around pairing human judgment with artificial intelligence. He described a “treasure”: more than 2,500 BYU devotional and forum addresses collected over decades—often once-in-a-lifetime talks into which speakers poured their best thinking, among them figures such as historian David McCullough, Senator Joseph Lieberman, Archbishop Charles Chaput, and Rabbi Harold Kushner. For years, he said, this library sat behind a single barrier: language.
Although BYU teaches about 85 languages—more than any other U.S. university—that number is dwarfed by the world’s 7,000-plus languages, and requests for translations arrived almost daily. About three years ago his team began translating talks into Spanish, Portuguese, French, and Japanese, but dubbing video naturally proved slow, painstaking, and expensive. In response they built a program called To All, which draws on the best AI voices available across companies and places a natural-sounding voice onto a video in step with the speaker, dubbing from an existing translated script or translating on the fly when none exists. Because it connects directly to Richardson’s low-resource work, Canney explained, To All can reach overlooked languages that most commercial tools ignore, and moving the system to the cloud could soon make it possible to translate into hundreds or thousands of languages at once.
Crucially, he stressed, the tool is meant to relieve translators of mechanical labor, not replace them. The machine is tireless and fast, he observed, but it does not know what is sacred, cannot catch a phrase that would wound, and cannot feel the weight of a speaker’s conviction—that, he said, remains human, even holy, work.
“The AI carries a load, the person carries the meaning.”
Ammon Shurtz
Shurtz explained that the Matrix Lab—short for Machine Translation Research and Interlingual Experimentation—focuses on improving AI translation for under-resourced, underserved languages. Working on the technical side, he develops methods tailored to the distinct properties of individual languages, from Romance and Germanic tongues to the many languages of Asia and to sign languages. His commitment is personal: raised in a multilingual household and of Cambodian heritage, he is acutely aware that everyday tools many English speakers take for granted, such as voice assistants, simply do not exist for most of the world’s languages.
He highlighted AI translation’s potential for humanitarian work, doctor-patient communication, and language preservation, noting that many languages are disappearing without any digital record of the culture embedded in them. The methods differ by case: sign languages require video and computer-vision techniques that isolate a signer’s motion and remain early-stage; low-resource languages related to higher-resource ones—such as certain Indigenous Mexican languages and Spanish—can benefit from transfer learning; and massively multilingual models trained on hundreds of languages at once can share knowledge across them to lift the smaller ones.
Asked about hallucinations, Shurtz called them a serious and still-unsolved problem, but pointed to an active area of research known as automatic quality estimation, in which separate models predict how trustworthy a given translation is likely to be, alongside efforts to build more honest models that report their own confidence.
“It doesn’t fix the hallucination problem, but at least it lets us know when it happens.”
Question and Answer Session
Are translators embracing these tools, or pushing back?
Canney said the translators he works with have been ecstatic, freed from the mechanical drudgery of dubbing to concentrate on questions of meaning and intent. Richardson added historical perspective: thirty years ago translators resented automatic translation because it was poor and they felt forced to use it, but the arrival of neural methods about a decade ago transformed both the quality of the output and translators’ attitudes toward it.
How will you keep To All from being misused, and where is it headed?
Canney explained that the tool’s purpose is not profit—were it, the team would simply chase lucrative high-resource languages—but to carry worthwhile content to communities that are usually left out. He and Richardson jointly vet who receives the technology, which has been licensed to one enterprise-software company and is in a pilot with their own church, and they are working to integrate it with translation-memory tools so it mirrors trusted human work. Richardson emphasized that access is not decided by a group’s faith tradition, describing faith communities as the “boots on the ground” doing good worldwide, and cited disaster response—such as the rapid creation of an English–Haitian Creole system by several major companies after the 2010 earthquake—as an example of translation technology serving humanity across cultures.
How are you acquiring data for low-resource languages?
Shurtz first clarified the term: a low-resource language is simply one that lacks enough data to reach good-quality translation. Richardson then described the scale of the challenge—only an estimated 400 to 500 languages have more than a million speakers—and the data appetite involved, from hundreds of thousands of text examples to hundreds of hours of recorded speech. He detailed PathSay’s goal of gathering balanced voice samples across many speakers and flagged open questions around dialect coverage and the privacy and fair compensation of those who lend their voices.
Concluding Remarks
Looking ahead, the panelists shared their hopes for the field. Canney suggested that technology “migrates to the creator,” steadily lowering the cost and barriers to creative work and placing powerful tools directly in people’s hands. Shurtz wished for private, on-device AI that keeps users’ data off corporate servers. Richardson voiced a hope aimed squarely at the forum’s audience: that faith communities might contribute carefully prepared translation data—perhaps through an IF20 data consortium—so that future AI models represent every religion accurately and without bias, in any language. Marianna Richardson noted that the series’ next webinar, on AI and benchmarks, would take up exactly that question.
In their closing statements, Canney said his team simply wants “all that’s lovely” to reach the whole world in every language; Shurtz said he loves how deeply the field is grounded in humanity; and Richardson returned to his opening vision, stressing that it can only be realized through collaboration and mutual respect, especially on matters of faith.
“It’s not going to happen unless we work together.”
Marianna Richardson thanked the speakers and invited viewers to the next session in the series, “AI and Benchmarks,” which will examine how AI models are evaluated—including efforts to measure and correct religious bias and to ensure different faiths are represented accurately across languages.
———
JoAnne Wadsworth is a Communications Consultant for the G20 Interfaith Forum Association and Editor of the Viewpoints Blog, IF20’s platform for commentary at the intersection of faith and global policy. She has covered the work of the Forum across multiple G20 cycles, producing summaries, articles, and editorial content that bring the voices of interfaith leaders and scholars to a broader audience. Her writing engages with topics spanning environmental governance, religious freedom, anti-racism, and the role of faith in multilateral policy processes. She is based in the United States.