AI-powered applications are exciting because of their potential to support people in unprecedented ways but they are also particularly challenging to design right: How does one design a useful AI-enabled product if the underlying AI is occasionally wrong? How can we make these applications feel predictable even though the AI technology is complex and can appear unpredictable?  While some specialized design knowledge related to Human-AI Interaction already exists, the production of this knowledge is not keeping up with the pace at which new AI-powered applications are invented. Consequently, without much fanfare or deliberation (or recognition of the fact!), some critical knowledge gaps are getting filled with reasonable-sounding but unverified assumptions. I will present a series of experiments (related to predictive text entry and AI-supported decision making) demonstrating that several of the key assumptions, upon which a lot of research projects and products rest, are wrong. I will then describe recent projects that build on corrected knowledge foundations and share some early promising results. I conclude with two calls to action for our field. First, we need to engage in critical technical practice, i.e., explicitly name, assess and correct (if necessary) the hidden assumptions of our field. Second, with Human-AI Interaction being a relatively new field but one that many people depend on, we need a greater investment in systematic production, synthesis and dissemination of reliable design knowledge for Human-AI Interaction.