Artificial intelligence

South Korea sets the global standard for frontier AI regulation

South Korea will begin enforcing its Artificial Intelligence Act on Thursday, becoming the first country to introduce formal safety requirements for high-performance, or frontier, AI systems, reshaping the global regulatory landscape.

The law establishes a national AI governance framework, led by the Presidential Council on National Artificial Intelligence Strategy, and creates an AI Safety Institute to oversee safety and trust assessments.

Alongside regulatory measures, the government is rolling out broad support for research, data infrastructure, talent development, startups, and overseas expansion, signalling a growth-oriented policy stance.

To minimise early disruption, authorities will introduce a minimum one-year grace period centred on guidance, consultation, and education rather than enforcement.

Obligations cover three areas: high-impact AI in critical sectors, safety rules for frontier models, and transparency requirements for generative AI, including disclosure of realistic synthetic content.

Enforcement remains light-touch, prioritising corrective orders over penalties, with fines capped at 30 million won for persistent noncompliance. Officials said the framework aims to build public trust while supporting innovation, serving as a foundation for ongoing policy development.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot! 

GPT-5.2 shows how AI can generate real-world cyber exploits

Advanced language models have demonstrated the ability to generate working exploits for previously unknown software vulnerabilities. Security researcher Sean Heelan tested two systems built on GPT-5.2 and Opus 4.5 by challenging them to exploit a zero-day flaw in the QuickJS JavaScript interpreter.

Across multiple scenarios with varying security protections, GPT-5.2 completed every task, while Opus 4.5 failed only 2. The systems produced more than 40 functional exploits, ranging from basic shell access to complex file-writing operations that bypassed modern defences.

Most challenges were solved in under an hour, with standard attempts costing around $30. Even the most complex exploit, which bypassed protections such as address space layout randomisation, non-executable memory, and seccomp sandboxing, was completed in just over three hours for roughly $50.

The most advanced task required GPT-5.2 to write a specific string to a protected file path without access to operating system functions. The model achieved this by chaining seven function calls through the glibc exit handler mechanism, bypassing shadow stack protections.

The findings suggest exploit development may increasingly depend on computational resources rather than human expertise. While QuickJS is less complex than browsers such as Chrome or Firefox, the approach demonstrated could scale to larger and more secure software environments.

Would you like to learn more about AI, tech, and digital diplomacy? If so, ask our Diplo chatbot!

AI model maps how humans form emotions

Researchers in Japan have developed an AI framework designed to model how humans form emotional experiences by integrating bodily signals, sensory input and language. The work was led by scientists at Nara Institute of Science and Technology in collaboration with Osaka University.

The AI model draws on the theory of constructed emotion, which suggests emotions are built by the brain rather than hard-wired responses. Physiological data, visual cues and spoken descriptions were analysed together to replicate how people experience feelings in real situations.

Using unlabeled data from volunteers exposed to emotion-evoking images and videos, the system identified emotional patterns without predefined categories. Results showed about 75 percent alignment with participants’ own emotional assessments, well above chance levels.

The Japanese researchers say the approach could support emotion-aware AI applications in healthcare, robotics and mental health support. Findings were published in IEEE Transactions on Affective Computing, with potential benefits for understanding emotions that are difficult to express verbally.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot! 

YouTube’s 2026 strategy places AI at the heart of moderation and monetisation

As announced yesterday, YouTube is expanding its response to synthetic media by introducing experimental likeness detection tools that allow creators to identify videos where their face appears altered or generated by AI.

The system, modelled conceptually on Content ID, scans newly uploaded videos for visual matches linked to enrolled creators, enabling them to review content and pursue privacy or copyright complaints when misuse is detected.

Participation requires identity verification through government-issued identification and a biometric reference video, positioning facial data as both a protective and governance mechanism.

While the platform stresses consent and limited scope, the approach reflects a broader shift towards biometric enforcement as platforms attempt to manage deepfakes, impersonation, and unauthorised synthetic content at scale.

Alongside likeness detection, YouTube’s 2026 strategy places AI at the centre of content moderation, creator monetisation, and audience experience.

AI tools already shape recommendation systems, content labelling, and automated enforcement, while new features aim to give creators greater control over how their image, voice, and output are reused in synthetic formats.

The move highlights growing tensions between creative empowerment and platform authority, as safeguards against AI misuse increasingly rely on surveillance, verification, and centralised decision-making.

As regulators debate digital identity, biometric data, and synthetic media governance, YouTube’s model signals how private platforms may effectively set standards ahead of formal legislation.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!

Why AI systems privilege Western perspectives: ‘The Silicon Gaze’

A new study from the University of Oxford argues that large language models reproduce a distinctly Western hierarchy when asked to evaluate countries, reinforcing long-standing global inequalities through automated judgment.

Analysing more than 20 million English-language responses from ChatGPT’s 4o-mini model, researchers found consistent favouring of wealthy Western nations across subjective comparisons such as intelligence, happiness, creativity, and innovation.

Low-income countries, particularly across Africa, were systematically placed at the bottom of rankings, while Western Europe, the US, and parts of East Asia dominated positive assessments.

According to the study, generative models rely heavily on data availability and dominant narratives, leading to flattened representations that recycle familiar stereotypes instead of reflecting social complexity or cultural diversity.

The researchers describe the phenomenon as the ‘silicon gaze’, a worldview shaped by the priorities of platform owners, developers, and historically uneven training data.

Because large language models are trained on material produced within centuries of structural exclusion, bias emerges not as a malfunction but as an embedded feature of contemporary AI systems.

The findings intensify global debates around AI governance, accountability, and cultural representation, particularly as such systems increasingly influence healthcare, employment screening, education, and public decision-making.

While models are continuously updated, the study underlines the limits of technical mitigation without broader political, regulatory, and epistemic interventions.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!

How Microsoft is shaping UN reform through digital infrastructure

Microsoft has announced a multi-year pledge to support the United Nations’ UN80 reform initiative, positioning AI and digital infrastructure as central tools for modernising multilateral governance.

The commitment follows agreement among all UN member states on efficiency and financial-stability measures, as the organisation faces growing operational demands amid constrained resources.

The initiative includes a dedicated innovation fund, preferential pricing for digital services, and free AI training for UN staff across agencies and missions.

Rather than focusing on policy direction, Microsoft frames its role as enabling institutional capacity, from procurement and logistics to humanitarian response and development planning, while encouraging other private-sector actors to align behind UN80 priorities.

Microsoft also plans to mobilise partners such as EY to support reform efforts, reinforcing a model where large technology firms contribute expertise, infrastructure, and coordination capacity to global governance systems.

Previous collaborations with UNICEF, UNHCR, ITU, and the ILO are cited as evidence that AI-driven tools can accelerate service delivery at scale.

The pledge highlights how multilateral reform increasingly depends on private technological ecosystems instead of purely intergovernmental solutions.

As AI becomes embedded in the core operations of international institutions, questions around accountability, influence, and long-term dependency are likely to shape debates about the future balance between public authority and corporate power.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!

Youth express higher concern over AI job disruption, global survey indicates

A Randstad survey of global workers shows that younger employees are significantly more worried than older counterparts about the impact of artificial intelligence (AI) on jobs and career prospects.

Respondents aged under 35 reported higher levels of concern that AI could disrupt employment, reshape skill demands and erode job security.

The survey highlights a generational divide in attitudes toward AI adoption: while many older workers express cautious optimism, younger workers often point to risks related to hiring, task automation and future opportunities.

Employers are urged to address these concerns by investing in skills development, training programmes and transparent communication about how AI will be deployed in the workplace.

The findings come amid broader global discussions on the future of work, with AI increasingly integrated into business processes such as recruitment, project planning and administrative automation.

Analysts suggest that targeted education and retraining initiatives could help younger employees adapt and benefit from AI-driven changes rather than feel displaced by them.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!

ChatGPT introduces age prediction to strengthen teen safety

New safeguards are being introduced as ChatGPT uses age prediction to identify accounts that may belong to under-18s. Extra protections limit exposure to harmful content while still allowing adults full access.

The age prediction model analyses behavioural and account-level signals, including usage patterns, activity times, account age, and stated age information. OpenAI says these indicators help estimate whether an account belongs to a minor, enabling the platform to apply age-appropriate safeguards.

When an account is flagged as potentially under 18, ChatGPT limits access to graphic violence, sexual role play, viral challenges, self-harm, and unhealthy body image content. The safeguards reflect research on teen development, including differences in risk perception and impulse control.

ChatGPT users who are incorrectly classified can restore full access by confirming their age through a selfie check using Persona, a secure identity verification service. Account holders can review safeguards and begin the verification process at any time via the settings menu.

Parental controls allow further customisation, including quiet hours, feature restrictions, and notifications for signs of distress. OpenAI says the system will continue to evolve, with EU-specific deployment planned in the coming weeks to meet regional regulatory requirements.

Would you like to learn more about AI, tech, and digital diplomacy? If so, ask our Diplo chatbot!

New AI system helps improve cross-neurotype communication

Researchers at Tufts University have developed an AI-based learning tool designed to improve communication between autistic and neurotypical people. The project focuses on helping non-autistic users better understand autistic communication preferences.

The tool, called NeuroBridge, uses large language models to simulate everyday conversations and highlight how wording, tone and clarity can be interpreted differently. Users are guided towards more direct and unambiguous communication styles that reduce misunderstanding.

Unlike many interventions, NeuroBridge does not aim to change how autistic people communicate. The AI system instead trains neurotypical users to adapt their own communication, reflecting principles from the social model of disability.

The research, presented at the ACM SIGACCESS Conference on Computers and Accessibility, received a best student paper award. Early testing showed users gained clearer insight into how everyday language choices can affect cross-neurotype interactions.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot! 

New consortium applies AI to early drug research

A new AI-driven drug discovery initiative with a budget exceeding €60 million has launched, bringing together academic and industry partners across Europe and North America. University College London is acting as the lead academic partner in the UK.

The five-year LIGAND-AI programme is funded through the Innovative Health Initiative and aims to speed up early drug discovery. Researchers will generate large open datasets showing how molecules bind to human proteins, supporting the training of advanced AI models.

The consortium, led by Pfizer and the Structural Genomics Consortium, includes 18 partners in nine countries. Work will focus on proteins linked to diseases such as cancer, neurological conditions and rare disorders.

UK based UCL scientists will help build global research networks and promote open sharing of protein samples and machine learning models. Organisers say the project supports open science and long-term goals to map chemical modulators for every human protein.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!