The case for letting people mess around with AI

The Ghibli effect

When OpenAI shipped its 4o image generator in March 2025, the same scene played out in thousands of companies: Slack channels filling up with Studio-Ghibli-style portraits of dogs, kids, weekend trips, and (inevitably) the CEO. Demand spiked hard enough that OpenAI publicly throttled image generation. It was easy to file under meme and assume it would burn out in two weeks.

It didn't — and adoption leads across the industry noticed the same pattern in their usage data: people who had never opened an AI tool showed up for the pictures, and a meaningful share of them stayed for the work. They tried chat. They tried the custom assistants their colleagues had built. The toy was the on-ramp.

Why play converts non-users

For dormant users, play removes activation energy. An image toy demands zero prior knowledge — no prompting technique, no mental model, no way to be wrong. Someone who just wants a cartoon of their dog crosses the threshold without noticing there was one. After that, the tool isn't foreign anymore.

For active users, play breaks frames. People who had filed AI under "text tool" watched it do something categorically different, and a broken frame makes people re-examine their other assumptions — that's when they wander into research modes, automations, and custom assistants they'd ignored for months.

And play recruits for free. A colleague laughing at a Ghibli dog is a more effective onboarding moment than any mandatory training deck, because it carries social proof instead of obligation — Duolingo learned the inverse the hard way, sparking a public backlash by tying AI use to performance reviews before quietly walking the mandate back. With 88% of organizations now using AI somewhere (McKinsey, 2025), the bottleneck is rarely access — it's the willingness to show up, and play manufactures willingness.

Making play pay

Play only compounds if it's visible and then bridged. The operating moves:

• Open a public play channel where silly experiments are explicitly welcome.
• Have leaders post their own messy attempts first — permission flows downhill.
• Spotlight frivolous-but-clever uses in all-hands, not just the serious wins.
• Run a monthly time-boxed show-and-tell; keep it to demos, not slides.
• Then bridge: hand each team one real workflow to apply what they found — our first-workflow rubric is built for exactly that handoff.

The organizations that struggle with adoption tend to share one decision: they skipped play and went straight for ROI. In doing so they skipped building the population of people capable of delivering the ROI. Play isn't a detour from the serious work — it's how you staff it. Moderna shows what that looks like at scale: once building your own AI assistant became normal, employees created around 750 of them in two months.