Stanford published the 2026 AI Index this week. Over 400 pages of data. But the thing that stayed with me wasn't a headline number.
It was two curves going in opposite directions at the same time.
Everyone's reporting the upward curve. Enterprise AI adoption hit 88%. Generative AI reached 53% of the population in just three years — faster than the PC, faster than the internet. The usual breathless coverage followed.
The downward curve barely made the news.
The Foundation Model Transparency Index — which tracks how openly major AI companies document their models — dropped from 58 to 40 in a single year. While adoption was peaking.
More people using these models. Less information about how they actually work.
This might sound like a policy issue. A researcher's problem. It's not. It's a product problem.
Here's why.
You're building a product. There's a model at its core. You don't know the training cutoff. You don't know what edge cases make it inconsistent. You don't know how it was fine-tuned or what data shaped its behavior in the specific domain you're deploying it in. And this not-knowing is getting deeper, not shallower.
Meanwhile, your users trust you more every quarter. They're entering their data. Following recommendations. In healthcare, they're prescribing, planning treatments, making clinical decisions based on what your system surfaces.
So: increasingly opaque infrastructure. Increasing trust load sitting on top of it.
These two things can't both be true indefinitely. Either trust needs to come down, or transparency needs to go up. Right now, neither is happening.
I see this directly in the work I do on Medibulut. Clinicians using the system expect answers to "what did the AI say?" They're not asking "why did the model generate that output." The gap between those two questions is enormous.
The first one is a PM responsibility. I can answer it, or build a UI that answers it.
The second is, increasingly, unanswerable — because the model companies are telling us less, not more.
And that gap widens every month.
There are two wrong reactions to this.
One is to ignore it: "The model performs well, users are happy, no problem here." The other is to freeze: "If I can't explain it fully, I won't build it."
Both miss the point.
The right response is layer management. The model can be opaque. But how its output connects to your product, how it's presented to users, when it can be overridden, where decisions get logged — those are things you control. The cleaner those layers are, the less that opaque core becomes a liability.
Transparency responsibility is shifting from the model to the product.
That's the shift I keep coming back to.
"AI transparency" used to be a concern for researchers and regulators. PMs watched from a distance. That's no longer viable.
Model companies are becoming less transparent while user trust is increasing. Your product sits between those two realities. You manage the middle layer.
Don't wait for the model to be transparent. Build transparency yourself.
Explain why the output is what it is — even when the model doesn't. Show that decisions can be overridden. Document when AI is involved. Log what happened. These aren't technical details. They're product decisions.
What Stanford can't measure is this: in enterprise settings, when an AI output goes wrong, who gets to say "we didn't know"?
Regulators are starting to ask that question. And when they do, the answer can't be "the model."
The model won't be in the room. You will.
That preparation isn't technical. It's not a model upgrade. It's a product decision — made now, before someone forces you to make it later.