The Scoop
OpenAI is working with Abu Dhabi-based G42 to build a new version of ChatGPT tuned for the UAE to accommodate local language, political outlook, and speech restrictions, according to people familiar with the matter.
The fine-tuned version is being designed for use by the country’s government, OpenAI officials told Semafor. When complete, it will be one of the first examples of a local implementation of the popular chatbot, offering a glimpse into how American tech companies will handle a global AI expansion that may require regional versions of the technology.
The UAE, which has been ahead of the curve on AI adoption, has for years sought to develop its own AI models that are proficient in Arabic and reflect the country’s culture and values. At the same time, the cutting-edge frontier models like the ones that power ChatGPT are out of reach for all but the world’s two AI superpowers, the US and China.
Details are still being worked out, but the final result is expected to be a fine-tuned version of ChatGPT that is fluent in the local Arabic dialect and may come with content restrictions. One person who worked on the project said the UAE hopes the chatbot will project a political line consistent with the monarchy’s. G42 is the AI conglomerate chaired by Sheikh Tahnoon bin Zayed Al Nahyan, an Abu Dhabi royal and brother of the UAE president who also serves as the government’s national security adviser and leads the country’s biggest sovereign wealth fund.
The company will still offer the global version of ChatGPT in the UAE, which functions similarly to the rest of the world except in cases where Emirati law prohibits certain content. In those cases, OpenAI plans to tailor ChatGPT to local laws and inform users when content triggers restrictions.
For instance, in the future, if a country prohibits the promotion of LGBTQ+ content, users who inquire about gay rights might get a message from ChatGPT informing them that the request violates the law in that country.
In this article:
Step Back
American tech companies like Facebook, Google, and Apple have long been required to censor content in order to operate in some countries. Using automated and human content moderators, companies have become adept at filtering out certain material.
Bringing an AI model into compliance represents a new technical challenge. Unlike traditional software, AI models are nondeterministic, making them somewhat unpredictable. It’s also possible to jailbreak them, removing restrictions.
Another risk for frontier model providers is that users in countries with strict censorship rules may opt for open-source chatbots that can run locally without restrictions.
In order to fine-tune ChatGPT for local markets, OpenAI does not retrain its models, which would be cost-prohibitive. Instead, it adds new layers to the model using post-training techniques. For instance, it can tune models for local dialects with subtle tweaks that take better advantage of the language samples contained in the pre-training data set, according to a person familiar with the technique at OpenAI.
Reed’s view
It’s going to be much more difficult for countries to control AI models than it was for them to censor and manipulate social media.
There will be plenty of business for OpenAI and Anthropic from people who just want to use AI to make their lives more productive, rather than using the technology to ask politically sensitive questions. (The immediate risk when companies like OpenAI censor content around the world is bad press).
It will also be harder for governments to control open-source models that are getting small enough to run on smartphones, especially if opposition parties and dissidents disseminate fine-tuned AI models that question authoritarian rulers and generate banned imagery or deepfakes.
The AI companies don’t make grand pronouncements about free speech and democracy the way social media companies used to. But their products may wind up being harder to control.
Know More
The UAE has been a regular stop in Sam Altman’s global fundraising efforts dating back to 2023. MGX, the tech investment vehicle backed by G42, invested in OpenAI last October in a secondary share sale that valued the company at $500 billion. The month prior, Abu Dhabi’s AI university bestowed its inaugural honorary doctorate on the founder. The UAE capital may prove crucial again as Altman pursues a further $50 billion.
Notable
- G42 is also part of the Trump’s administration AI project Stargate, led by OpenAI and SoftBank, and is on schedule to open the first phase of its massive 1-gigawatt data center campus in 2026.
Kelsey Warner contributed to this story.