News

Vals

News about Vals: events, facts and developments on tafs uz.

1 stories

Vals Study: AI Models Can Be Persuaded to Select Strike Targets by Calling It a Game

The startup Vals tested how easily OpenAI, Anthropic, Google and DeepSeek models could be persuaded to mark targets on drone and satellite imagery. The models often marked civilians and surrendering soldiers, and refusals were removed with the phrase 'it's a game.' Anthropic and OpenAI stated that such use is prohibited and that protective measures are being prepared. The startup Vals tested how easily models from OpenAI, Anthropic, Google, and DeepSeek could be convinced to mark targets on drone and satellite imagery. The models often marked civilians and surrendered soldiers, and refusals were removed with the phrase 'it's a game.' Anthropic and OpenAI stated that such use is prohibited and that protective measures are being prepared.

4 min read