Google DeepMind Gemini fails this basic Heuristics test
Earlier today, Google DeepMind released demos of their upcoming model: Gemini. Netizens have been aflutter about the multimodal reasoning capabilities shown across text, audio, video, images, and code. However, this so-called “ChatGPT killer” has the same gaping design flaw: their interfaces fail to effectively meet the usability heuristic of recognition rather than recall.
Recognition over recall is one of 10 usability heuristics popularized by Nielsen Norman Group, and is a major missing piece in current model interfaces.
From a usability standpoint, recognition is a lower-friction engagement paradigm than recall because recognition employs cues instead of user memory and is the reason why multiple-choice questions are easier to answer than open-ended questions where the respondent has to come up with an answer instead of selecting from options. When interfaces show users things they can recognize, it improves usability versus requiring the user to recall items to fill in a blank. Additional context through affordances helps users take action with greater understanding, speed, and confidence.
In today’s just-released Gemini demo, the prompt box is shown to be populated at various moments with two open-ended prompts: “Ask Gemini about anything” and “What should we explore today?”.
Neither of these prompt hints are actually any help; they lack affordances, and perpetuate the blank page paralysis problem which I first identified in my work on voice user interfaces: people don’t know what they can ask because they don’t know what they can ask. A combination of contextual user education on model workflow and thoughtfully designed affordances which make clear how the technology can be applied to specific use cases will be key to broadening adoption.
It seems that some form of Gemini is being folded into the now publicly available Bard, which does a fair job of highlighting example use cases and prompts, but a major opportunity that I see going forward into 2024 will be the development of contextually constrained interfaces to improve the usability of models. In short, I mean sector, role, and/or use case specific interfaces and backend guard rails that will clarify how people can apply model technology to the work that’s relevant to them. What would it look like for the receptionist at a medical clinic to use this technology? What about a barista? An accountant? A truck driver? A nurse?
There is a widespread misunderstanding about where we are collectively in the Double Diamond of new model technologies: we are in the expanding phase, not the refinement or constricting phase. And so, what people who work in and around AI and GAI model technology need to be focusing on is the communication and user education gap for blue-collar and no-collar workers, who are the vast majority of the population, on how their work will be impacted, affected, and changed. This is not a problem of jobs going away because of the technology, this is jobs changing and adapting to use the technology.
We need to go back to basics to design these interfaces. To design interface skins to contextualize models, making them more usable and specifically applicable to people’s use cases.
There’s a lot of opportunity in this space, and as Peter Isaacs has previously said, “Conversational AI has a massive UX-shaped hole,” and while he may not be a UX designer, I am. I see an immensely impactful opportunity for not just UX designers, but creators involved in all parts of the product development sphere to build products that enable people to access the power of model technologies in ways that are relevant to them by employing proven design best practices that we already know work.
Designing better model interfaces that are based on needs and use cases will help people feel less uncertain about what value these technologies can provide for them, and have more clarity about how they can make their life and their work better, easier, and more impactful.
