ChatGPT is three years old and most users are still barely touching what it can do

OpenAI has given a name to a problem that has been sitting in plain sight. Ian Silber, the company’s Head of Product Design, describes what the team calls “capability overhang”: the gap between what ChatGPT can actually do and what the vast majority of users manage to extract from it. The label matters because it places the burden on design rather than on the underlying model. The hard part, in OpenAI’s framing, is not making the system more capable. It is getting people to use the capability that already exists.

Silber points to the interaction model as where that design work is happening. ChatGPT is moving away from text-in and text-out toward something more tactile. He describes a new pattern in which the product returns content in a block that a user can directly manipulate, typing with it, brainstorming, and editing specific parts. That is a different kind of relationship with the tool than pasting a reply into a separate document, and it suggests the product is trying to make depth of use feel lower-friction rather than asking users to develop new habits on their own.

The behavioral data that Nathaniel Whittemore, Founder and Chief Executive of Superintelligent, surfaces makes the design challenge legible in a different way. He cites a figure showing that 43.5 percent of occupation-specific ChatGPT use at work involves tasks belonging to a different occupation than the user’s own. In other words, people are not primarily using the tool to do more of what their job title describes. They are reaching across occupational lines, drafting things they would previously have handed to someone else, running analyses outside their formal domain. The tool is already reshaping how people define the scope of their own work, even if most of them are not using it at full depth.

We believe in this idea of this sort of capability overhang where the product you know vast majority of the people are getting a sliver of the true value that they could out of the bottle Ian Silber

What makes Whittemore’s data more pointed is what sits alongside it. He notes there is zero evidence of clear AI fingerprints in aggregate United States occupation data, this despite approaching three years since ChatGPT’s release, and despite artificial intelligence being cited as the number one stated reason for job cuts for five consecutive months. Those two facts do not cancel each other out, but they do create a genuine puzzle. Workers are crossing occupational lines with the tool. Employers are invoking AI as justification for cuts. And the macro employment data shows nothing. The gap between those three readings is not resolved by any of the evidence on hand, and anyone who claims otherwise is getting ahead of what the numbers actually say.

Harry Stebbings, Founder of 20VC, adds a related corrective aimed at the startup and venture conversation specifically. His argument is that the do-it-yourself threat to small and mid-size software businesses is a short myth. The more credible competitive pressure on products like HubSpot and Monday comes not from users building their own replacements with ChatGPT, but from purpose-built, AI-native competitors at the low end that are already very strong. That is a narrower claim than it might first appear: Stebbings is not saying incumbents are safe, only that the vector most commonly cited in the investor conversation is probably the wrong one to watch.

Taken together, these threads point to a product and a market that are both further along and less legible than the standard framing suggests. ChatGPT is already changing how people scope their own work. OpenAI knows the depth problem is a design problem and is actively rearchitecting the interface to address it. And the macro data, despite years of use and visible employer behavior, has not yet moved in ways that are easy to interpret. The honest position is that something significant is happening at the level of individual behavior, the mechanism by which it will show up in aggregate numbers is not yet established, and the design choices being made right now about how people interact with these tools will shape which way that resolves.