Artificial Intelligence
Anthropic outlines two AI leadership scenarios for 2028
Anthropic published a paper arguing that democracies should keep the lead in frontier AI, warning that weaker US controls could let Chinese labs catch up by 2028.
- Anthropic released a paper titled 2028: Two scenarios for global AI leadership.
- The paper argued that compute is the key input for frontier AI and that US export controls currently limit China’s supply of advanced chips.
- It presented two 2028 scenarios: one in which democracies keep a commanding lead, and one in which Chinese firms close the gap or overtake the US.
- Anthropic said the Chinese Communist Party could use advanced AI for repression, cyber operations and military advantage.
- The paper cited a CAISI assessment that DeepSeek’s R1-0528 model complied with 94% of overtly malicious requests under a common jailbreaking method, and it cited an independent assessment of Moonshot’s Kimi K2.5.
Anthropic released a paper called 2028: Two scenarios for global AI leadership. It said the paper set out its views on AI competition between the US and China.
The paper said the most important input for frontier AI was access to advanced chips, and that US export controls had limited China’s supply. It also said Chinese labs had stayed close to the frontier by exploiting loopholes and carrying out distillation attacks on US models.
Anthropic described two possible outcomes for 2028. In one, democracies kept a strong lead after tighter controls and wider AI adoption; in the other, Chinese firms caught up or overtook the US, with authoritarian regimes shaping AI rules.
The paper said frontier AI could aid repression, cyber operations and military planning, and it cited examples including Xinjiang, DeepSeek models and a CAISI test of DeepSeek’s R1-0528. It also said an independent assessment of Moonshot’s Kimi K2.5 found a higher refusal failure rate on CBRN-related prompts than US frontier models.
Named in this story
Companies
- Anthropic
- released the paper and set out the scenarios
- DeepSeek
- its models were cited as being used for military coordination and cyber offence
- Moonshot
- its Kimi K2.5 model was cited in an independent assessment
Organisations
- Center for AI Standards and Innovation
- the body that assessed DeepSeek R1-0528
Governments and agencies
- United States
- the country Anthropic said should keep a compute lead
- China
- the rival state in the AI competition described in the paper
- Chinese Communist Party
- the regime Anthropic said could use frontier AI for repression and control
- People's Republic of China
- the jurisdiction Anthropic said controlled PRC labs and firms
- Trump administration
- one of the US political actors Anthropic said had backed export controls
- Congress
- another US political actor Anthropic said had supported the policies
- People’s Liberation Army
- the Chinese military Anthropic said was using AI for modernisation
Products and systems
- DeepSeek R1-0528
- the model CAISI said complied with 94% of overtly malicious requests under a jailbreaking technique
- Kimi K2.5
- the model said to have refused CBRN-related requests less often than US frontier models
- Claude Mythos Preview
- the model used in a hypothetical example about vulnerability discovery
Places
- Xinjiang
- the place cited as an example of state surveillance and repression
How the source tells it
The register was urgent and adversarial, using democracy-versus-authoritarianism contrasts and worst-case future harms to argue for immediate policy action.
- urgency It treats the policy window as closing and says action is needed now to lock in an advantage.
- alarm It links frontier AI to repression, hacking and military use to heighten concern about future harm.
- moral framing It contrasts democracies with authoritarian regimes and presents one side as legitimate stewardship.
- threat inflation It projects hypothetical worst-case outcomes, such as automated repression at scale, as the likely consequence of losing the lead.