This story originally appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here.
I’m back from a wholesome week off picking blueberries in a forest. So this story we published last week about the messy ethics of AI in warfare is just the antidote, bringing my blood pressure right back up again.
Arthur Holland Michel does a great job looking at the complicated and nuanced ethical questions around warfare and the military’s increasing use of artificial-intelligence tools. There are myriad ways AI could fail catastrophically or be abused in conflict situations, and there don’t seem to be any real rules constraining it yet. Holland Michel’s story illustrates how little there is to hold people accountable when things go wrong.
Last year I wrote about how the war in Ukraine kick-started a new boom in business for defense AI startups. The latest hype cycle has only added to that, as companies—and now the military too—race to embed generative AI in products and services.
Earlier this month, the US Department of Defense announced it is setting up a Generative AI Task Force, aimed at “analyzing and integrating” AI tools such as large language models across the department.
The department sees tons of potential to “improve intelligence, operational planning, and administrative and business processes.”
But Holland Michel’s story highlights why the first two use cases might be a bad idea. Generative AI tools, such as language models, are glitchy and unpredictable, and they make things up. They also have massive security vulnerabilities, privacy problems, and deeply ingrained biases.
Applying these technologies in high-stakes settings could lead to deadly accidents where it’s unclear who or what should be held responsible, or even why the problem occurred. Everyone agrees that humans should make the final call, but that is made harder by technology that acts unpredictably, especially in fast-moving conflict situations.
Some worry that the people lowest on the hierarchy will pay the highest price when things go wrong: “In the event of an accident—regardless of whether the human was wrong, the computer was wrong, or they were wrong together—the person who made the ‘decision’ will absorb the blame and protect everyone else along the chain of command from the full impact of accountability,” Holland Michel writes.
The only ones who seem likely to face no consequences when AI fails in war are the companies supplying the technology.
It helps companies when the rules the US has set to govern AI in warfare are mere recommendations, not laws. That makes it really hard to hold anyone accountable. Even the AI Act, the EU’s sweeping upcoming regulation for high-risk AI systems, exempts military uses, which arguably are the highest-risk applications of them all.
While everyone is looking for exciting new uses for generative AI, I personally can’t wait for it to become boring.
Amid early signs that people are starting to lose interest in the technology, companies might find that these sorts of tools are better suited for mundane, low-risk applications than solving humanity’s biggest problems.
Applying AI in, for example, productivity software such as Excel, email, or word processing might not be the sexiest idea, but compared to warfare it’s a relatively low-stakes application, and simple enough to have the potential to actually work as advertised. It could help us do the tedious bits of our jobs faster and better.
Boring AI is unlikely to break as easily and, most important, won’t kill anyone. Hopefully, soon we’ll forget we’re interacting with AI at all. (It wasn’t that long ago when machine translation was an exciting new thing in AI. Now most people don’t even think about its role in powering Google Translate.)
That’s why I’m more confident that organizations like the DoD will find success applying generative AI in administrative and business processes.
Boring AI is not morally complex. It’s not magic. But it works.
Deeper Learning
AI isn’t great at decoding human emotions. So why are regulators targeting the tech?
Amid all the chatter about ChatGPT, artificial general intelligence, and the prospect of robots taking people’s jobs, regulators in the EU and the US have been ramping up warnings against AI and emotion recognition. Emotion recognition is the attempt to identify a person’s feelings or state of mind using AI analysis of video, facial images, or audio recordings.
But why is this a top concern? Western regulators are particularly concerned about China’s use of the technology, and its potential to enable social control. And there’s also evidence that it simply does not work properly. Tate Ryan-Mosley dissected the thorny questions around the technology in last week’s edition of The Technocrat, our weekly newsletter on tech policy.
Bits and Bytes
Meta is preparing to launch free code-generating software
A version of its new LLaMA 2 language model that is able to generate programming code will pose a stiff challenge to similar proprietary code-generating programs from rivals such as OpenAI, Microsoft, and Google. The open-source program is called Code Llama, and its launch is imminent, according to The Information. (The Information)
OpenAI is testing GPT-4 for content moderation
Using the language model to moderate online content could really help alleviate the mental toll content moderation takes on humans. OpenAI says it’s seen some promising first results, although the tech does not outperform highly trained humans. A lot of big, open questions remain, such as whether the tool can be attuned to different cultures and pick up context and nuance. (OpenAI)
Google is working on an AI assistant that offers life advice
The generative AI tools could function as a life coach, offering up ideas, planning instructions, and tutoring tips. (The New York Times)
Two tech luminaries have quit their jobs to build AI systems inspired by bees
Sakana, a new AI research lab, draws inspiration from the animal kingdom. Founded by two prominent industry researchers and former Googlers, the company plans to make multiple smaller AI models that work together, the idea being that a “swarm” of programs could be as powerful as a single large AI model. (Bloomberg)