Klaus-Gerd Giesen<p>"Unchecked <a href="https://chaos.social/tags/AI" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>AI</span></a> agency poses significant <a href="https://chaos.social/tags/risks" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>risks</span></a> to public <a href="https://chaos.social/tags/safety" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>safety</span></a> and <a href="https://chaos.social/tags/security" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>security</span></a> ,ranging from misuse by malicious actors to a potentially irreversible loss of <a href="https://chaos.social/tags/human" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>human</span></a> control. [...] Indeed, various <a href="https://chaos.social/tags/scenarios" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>scenarios</span></a> and <a href="https://chaos.social/tags/experiments" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>experiments</span></a> have demonstrated the possibility of AI agents engaging in <a href="https://chaos.social/tags/deception" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>deception</span></a> or pursuing goals that were not specified by human operators and that conflict with human interests, such as self-preservation."</p><p><a href="https://arxiv.org/abs/2502.15657" rel="nofollow noopener" translate="no" target="_blank"><span class="invisible">https://</span><span class="">arxiv.org/abs/2502.15657</span><span class="invisible"></span></a></p>