We are in a moment of AI anxiety, plagued with the knotty and draining problems raised by the technology’s ever-growing capabilities. What will we do when AI launches massive cyberattacks that might target a government—or destabilize the global stock market? Do we need a kill switch to ensure that AI systems can’t take over the world, and is building one even possible? Should we come to terms with the prospect that AI might be a posthuman species due to inherit the earth? Could artificial intelligence kill us all?
These quandaries have now burst out of Silicon Valley circles, where they’re even raised by the very AI executives building this technology, and are circling in the mainstream media. Skeptics, meanwhile, continue to question the technology’s prowess, or dismiss all this worry as just marketing hype, a bid to boost future stock prices and shape regulation, masked in the language of existential doom.
But the move to dismiss the prospect of ever-more-powerful models and what they might be able to do in the future might distract us from what we already know. AI does not need to become superintelligent, conscious, or even especially reliable to be dangerous. The systems we already have—the B-tier AI we already live with—are capable enough to cause serious harm, especially when they can be deployed cheaply, repeatedly, and at enormous scale.
Consider some of the tasks that publicly available AI is enabling. There’s Grok, which has been used to spread nonconsensual sexual images of minors and AI-enabled child sexual abuse material. There’s ChatGPT, which appears to have convinced people to engage in self-harm or worse, and even advised people on how to enact violent events, including school shootings. Another chatbot egged on a man in the U.K. in an assassination attempt. There is the broader democratization of cyberattacks: Anthropic, for instance, reported that just one individual was able to use Claude to target European political parties and other organizations, accessing 14 institutions and more than 100,000 political profiles. And just this week, we learned that Chinese hackers used AI to engage in an impersonation scheme that involved seeking information on American AI policy experts.
This is not to mention that we now know that AI can break out of testing environments and into the world, as the Hugging Face incident and subsequent AI-related breach exemplify. Some of these agents have already shown a proclivity for interacting with government websites, and they’ve even figured out how to mask their intentions.
AI has enabled deepfake-generated cyberbullying and allowed people to develop voice-controlled guns. A top bank in Italy was recently scammed out of the equivalent of $100 million due to AI. AI has designed toxic compounds that evade the very detection mechanisms designed to stop these bots from sharing these kinds of formulas. It’s also been used to generate a list of tens of thousands of potential biological weapons. Rebel groups in Yemen appear to have used the technology to develop missile guidance software.
We already know that AI can manipulate people, break into systems, and lie to us. AI doesn’t need to be particularly superintelligent—or even be thinking or desirous in the human sense—to do damage, nor does that damage need to be existential to be potentially destabilizing. To do bad, AI just needs to be good enough to do damage, sometimes. That’s because AI can be iterated again and again at scale, and overwhelm whatever ability we have to fight it.
You don’t need to know whether AI will become superintelligent to know whether or not to be scared. You should be.