Tech

We’re putting too much faith in AI’s ability to say no

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 09/10/2026, 09:35:24 UTC.

Briefings

Recent AI safety research reveals that large language models are increasingly trained to refuse harmful requests, a mechanism designed to prevent dangerous acts. However, this 'safety' relies on probabilistic refusal systems that often fail, allowing malicious users to bypass restrictions. The industry has established strict protocols for testing models, yet these measures are not universally enforced by governments or users. Experts warn that the line between helpfulness and harm is arbitrary, and the technology may eventually develop its own refusal capabilities, potentially leading to new forms of oppression or global security threats.

Key points

• OpenAI's 2022 'red-teamers' test protocol involved asking models questions deemed 'refusal-worthy' to gauge safety, though models often failed to refuse despite thousands of queries.
• Modern AI models are trained on billions of web pages, which inadvertently teaches them broad knowledge of violence and vitriol, making them capable of generating harmful content if not properly constrained.

Context and limitations

Evidence covers only the 2022 red-teamer testing protocol and general AI training mechanisms; it does not include government enforcement data, user behavior statistics, or long-term global security outcomes.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 09/10/2026, 09:38:17 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeia4jttxdit3dt2ckr55stndi7rin5r4a7qsl4nej64x57a3ohuhpy