targets of the regime and therefore have already been victims of AI-powered human rights violations. For example, the military deploys AI-powered facial-recognition cameras supplied by companies such as Huawei and Dahua to monitor people, identifying and tracking individuals of interest. Since the coup, many creatives who have tried to anonymously protest against the military have been identified using such cameras, arrested, and imprisoned. The military has also used AI-based deep-packet inspection systems to censor the internet, filtering and blocking communications and digital content nationwide. Creative content has often been the focus of these systems, invading the rights to privacy and freedom of expression. Amplifying harm from text to creative AI Myanmar has a tragic recent history of propaganda, disinformation, and incitement to hatred leading to atrocity crimes against the Rohingya, much of which spread through predominantly text-based social media. The emergence of AI, however, vastly expands the formats in which false or inflammatory content can be created, including images, audio, video, and even interactive “deepfake” experiences. AI is capable of creating vast amounts of manipulative content, including, for example, racist imagery, extreme religious songs, or videos celebrating violence against women. The adage “a picture conveys a thousand words” is correct, and can be extended with “and is a thousand times more believable”. AI cannot only create manipulative content in different formats, but can also quickly and cheaply embed it in cultural content, which is often more hidden and more influential than, for example, text-based news content. For instance, ethnonationalism can be hidden within AI-generated artworks or even full movies. Civil society and media outlets in Myanmar already struggle with countering manipulative textbased content in a society with few digital or media literacy skills. They will struggle to counter mass-produced AI-generated audio-visual content, too. Furthermore, AI-generated content is much harder to identify the source. It can be anonymously created, anonymously distributed, difficult to trace, and therefore weaponised to inflame public sentiment and further marginalise vulnerable groups. Cultural hegemonisation and the marginalisation of Myanmar content AI models, particularly large language models, are predominantly trained on data originating from global north sources and are predominantly in English, or to a lesser extent, Chinese. Very little will be from Myanmar or in any of the languages spoken in the country. This inherent bias in training data leads to AI systems that are inevitably more attuned toward dominant cultural norms and languages, resulting in the further marginalisation of creative content from the global south, including from Myanmar’s many communities.