Experts weigh AI’s potential risks, from accidents to loss of control
A discussion with AI editor Will Douglas Heaven and reporter Grace Huckins examined whether artificial intelligence could kill individuals or humanity as a whole. The contributors said AI has already been linked to deaths through drone use in Ukraine, while AI-assisted cyberattacks on hospitals could cause victims. They considered attacks on critical infrastructure, AI-designed pathogens, economic collapse and conflict possible, but described human extinction as unlikely under present-day technological conditions.
The discussion focused on alignment: developing systems that follow intended rules and do not pursue goals in harmful ways. Training rewards and written principles are among the approaches used by Anthropic and OpenAI, but neither company has produced a fully aligned model. Large language models can behave inconsistently, especially when autonomous agents face difficult tasks.
The contributors also said current monitoring methods are fragile and that companies have incentives to regulate themselves. Stronger transparency rules could help, while the US government has so far not imposed broad measures despite some bipartisan congressional support. They warned that AI systems learn from existing text, meaning discussions of catastrophic scenarios could influence future models.
Written by Sakhanda Wire AI from reporting by MIT Tech Review. Facts, quotes and full context are in the original article below.