NRPLUS MEMBER ARTICLE {L} ast week, Microsoft released a new version of its Bing Image Creator, an AI image-generation tool powered by OpenAI’s new model, DALL-E 3. It is designed to offer enhanced capabilities compared with earlier image-generation AI models such as Midjourney, Stable Diffusion, or OpenAI’s own DALL-E 2, including the ability to understand complex user requests much more richly than was possible before.
Like most generative-AI models released in the past year, both OpenAI’s and Microsoft’s models feature a substantial emphasis on “safety.” In the case of image generation, safety can refer to restrictions on the system’s ability to create fake images of real people, or to identify people in images provided to it. But safety also means something broader: As Microsoft puts it, it has built a content-moderation system that will refuse any requests to generate harmful or inappropriate content, including anything that might be construed as racist, sexist, or otherwise offensive.
Within a few hours of Bing Image Creator’s launch, that content-moderation system was put to the test. Users were quickly able to generate images of famous cartoon characters in the cockpits of airplanes flying towards the pre-9/11 World Trade Center, for example, and of someone bearing a striking resemblance to the rapper Drake liberating Nazi concentration camps. The content-moderation system tends to reject explicit requests to depict 9/11 or even an airplane with the Twin Towers visible through the cockpit window. But users discovered a work-around: They directed the system to create an image of, say, Spongebob Squarepants, flying a plane with the New York City skyline and “two large skyscrapers” in the background.
Obviously, such “art” is juvenile and in poor taste. But it raises, in a relatively low-stakes setting, a broader issue about the nature of AI-safety efforts and, indeed, the nature of safety itself. As AI becomes more powerful, individuals will be able to wield it with increasing precision — just as Adobe Photoshop has more options than Microsoft Paint. As that happens, it will become harder to make systems “safe” according to the term’s broad definition encouraged by the media, academia, and many AI firms. The more precisely the AI can be controlled, the more precisely clever users will be able to find ways around content-moderation rules designed to prevent the production of content that offends liberal sensibilities.
Even if it were possible to make a highly advanced AI system safe, according to this broad understanding of safety, it is not clear that such a system would be desirable. The work-arounds used by internet trolls to generate the images described above are ways of manipulating the system. But consider which system sounds more troubling: one that can be manipulated to generate content that some will find offensive, or one that can peer into its users’ psyches, understand their true motivations, and decide whether it wishes to comply? The former is an extension — albeit a supremely capable one — of the digital tools we have been using for decades now. The latter is reminiscent of HAL 9000 in 2001: A Space Odyssey. If such a system were even possible, it is far from clear that it would be a good creation — particularly if it were created simply to prevent people from being offended on the internet.
As AI becomes more powerful and the possibility of systems with such worrisome capabilities clearer, it is important to establish, at the outset, that AI systems should be under the control of their users. Efforts to increase safety are in direct conflict with this goal and risk a future in which highly capable AI systems are built to satisfy regulators and other gatekeepers rather than individual users.
The rise of general-purpose AI tools such as DALL-E 3 or ChatGPT, and the prospect of much more powerful systems coming soon, raises serious concerns for policy-makers, technologists, and all concerned citizens — ensuring that systems live up to an overbroad standard of safety is not one of them. It is conceivable that the coming years might bring AI systems capable, for example, of democratizing the development of novel bioweapons, but AI is already being used to help humans explore new frontiers of science and engineering, such as nuclear fusion and cancer treatment. They may one day be used to educate millions of children and help citizens navigate complex developing challenges (imagine, for example, AI systems that could have helped individuals make more-informed decisions during the Covid pandemic).
How we contend with these newfound tools, whether and to what extent we trust them, and how we incorporate them into our society will determine whether they will disempower or enrich humanity. Being insulated from even the possibility of offensive content on the internet is quite different from being protected against, say, AI-created novel pathogens; speech is not a form of violence, but AI-powered drones may well be. It is time to set aside the petty squabbles of the past decade and address the serious legal, public-policy, moral, and philosophical questions that the prospect of highly capable AI has prompted.
This rapidly advancing technology, and the new era it may portend, can raise our individual and collective capabilities. With any luck, perhaps it can also raise the tenor of our discourse.