New research reveals how easily even cutting-edge AIs can be nudged to say dangerous or inappropriate things; the Washington Post explores how hospitals are scrambling to implement AI; and more, in this week's roundup of AI-related news in healthcare from Advisory Board's Thomas Seay.

As hospitals adopt AI, hope and fear are surging. The Washington Post offers this (somewhat meandering) overview of AI in healthcare, covering everything from conversational AIs to mortality prediction algorithms. Two themes: First, the healthcare industry's commitment to “do no harm” meshes uncomfortably with AI tools that often improve care but also misfire unpredictably. And second, healthcare workers are much like their counterparts in other industries: hopeful AI could make their jobs better —  and worried it could lead some jobs to disappear.

Speaking of the AI jobs threat … is it already over? Well, we seem to have reached the disillusionment stage of the AI hype cycle. Slate  argues that AI's impact on employment has been wildly overstated, citing one venture capitalist who claims, “The demise of industries due to AI is just not going to be a thing.”

• Analysis: The sky-high expectations for generative AI were always going to collapse. Still, it feels way  premature for postmortems. ChatGPT launched less than a year ago, and many workplaces still prohibit its use — plus, each new generation of AI models will be more powerful than the last. Personally, I suspect the impact of generative AI on employment has scarcely begun.

There's still no way to stop AI 'jailbreaks.' Most large language models (LLMs), including ChatGPT, include safeguards to prevent them from giving dangerous, offensive, or otherwise objectionable replies. Yet these protections can be bypassed by often-trivial “jailbreaks,” such as claiming a question is purely hypothetical. This paper reviews current jailbreaks, identifies some with up to 99% success rates, and argues that “current LLMs and safeguards cannot adequately defend jailbreak prompts in all scenarios.”

• Analysis: As healthcare organizations implement LLMs, especially in public-facing arenas (such as via online chatbots), they'll have to consider that some users may act maliciously — whether to harvest information the organization didn't intend to expose, or simply to troll their AIs into saying inappropriate things.

AI makes bad students better — and good students worse. To quote this fascinating study of law-school exams, “[S]tudents at the bottom of the class saw huge performance gains with AI assistance, while students at the top of the class saw performance declines.”

• Analysis: While AI is already better at some tasks than most humans, it still lags behind the best humans. Yet even highly capable people may feel tempted to use AI to quickly and easily produce good-enough results. To avoid this trap, we need to carefully consider where AI is better than we are (and, in those cases, let it do its thing) and where it’s worse (and, in those instances, use AI sparingly or not at all).

Last week's 'AI Roundup': What 1,684 leaders told McKinsey about AI's future. In last week's column, we explored what McKinsey Global Survey found about the benefits — and risks — of generative AI, how major websites are trying to prevent AIs from scraping their content, and more.