Is the AI capability gap really an interface problem?

Mollick argues that AI has been "running ahead of AI accessibility" for a while — that the capability overhang isn't primarily a limit of the models but of "how people interact with it." His target is the default access point: a free-tier chatbot window. "A chatbot is fine for a quick question, but it is a bad way to get real work done," he writes, and he cites a cognitive-load study he doesn't name in full detail: a small group of financial professionals doing a complex valuation task with GPT-4o1, with their cognitive load measured turn by turn from the transcripts. The study found a real productivity gain from AI use, offset by the chatbot format itself — "giant walls of text, offers to pursue new topics, and sprawling discussions" that overwhelmed users, who then failed to reorganize messy conversations, while the AI "just mirrored back whatever disorganized structure the user provided." Both sides compounded the mess, and the professionals hurt most were the least experienced — "exactly the people who could benefit the most from AI."

Mollick's reasoning works through three emerging fixes, each closing the gap a different way. Specialized interfaces built for one profession: only coding has a "really complete" one (Claude Code, Codex, Antigravity), because the labs are "staffed by programmers," while Google's Stitch, Pomelli, and NotebookLM are rougher attempts at the same for design, marketing, and research. Personal agents that drop the interface altogether: OpenClaw let people "talk to an AI in the way that you would a person" over WhatsApp or Telegram, and Anthropic's Cowork plus Dispatch is the safer, sandboxed answer to the same insight — "people don't want a chatbot. They want an agent that works on their actual files, with their actual tools, accessible the way they talk to people." And generative, on-the-fly interfaces, where "the AI generates the right interface for the moment" instead of a company pre-building one — Claude's in-conversation, adjustable visualizations are his example. His forecast: "every new interface that closes even part of that gap will feel like a leap in AI capability, even when the models haven't changed."

This sits alongside What makes an AI system feel like a colleague rather than a chatbot?, which makes the same not-bigger-models argument from the colleague-interface side; Dispatch is Mollick's concrete instance of that shift. It also extends Do generated interfaces outperform text-based chat for most tasks? by supplying the forecast that measured finding implies: generated interfaces will keep substituting for pre-built ones as the default way AI meets people. And it complicates Does AI assistance help less experienced workers most?: in that customer-support study, less experienced workers gained the most from AI assistance, while in the valuation study Mollick cites, less experienced workers were hurt most by the chatbot interface's cognitive load — same population, opposite outcome, depending on whether the interface organizes the work or dumps it on the user.

The valuation study is reported at second hand, with no name, author, sample size beyond "a small group," or publication given, so its cognitive-load finding can't be checked against its own method here. Mollick's Dispatch examples are his own single-user anecdotes (a morning briefing, one PowerPoint update), not a study of Cowork's reliability, and he says as much — "error-prone in practice." The three-category taxonomy of fixes is his own synthesis, not a measured comparison of which approach closes the gap fastest. What the excerpt does support, at the strength of an informed practitioner's argument rather than a finding, is that interface design is doing real work in how "AI disappointment" gets explained, independent of whether model capability itself has moved.