Contractual silence on inference caching
of professional services contracts signed in the last contain no specific clause regarding the temporary caching of client intellectual property on external inference servers.
It is a flat, dry statistic that sounds like something a compliance officer would mutter in their sleep, yet it represents the exact tectonic plate where the ground is currently splitting open. We are operating in a world where the legal department is still arguing about “data at rest” while the data itself has already been digested, metabolized, and turned into the muscle of someone else’s proprietary model.
The Three-Second Kaleidoscope
The meeting started at , but by , the room had reached that specific temperature where the air feels heavy with unspoken assumptions. Pauline sat across from her client, a man whose reasonable tone was perhaps his most terrifying attribute.
He wasn’t looking for a fight; he was looking for a baseline. He leaned forward, his pen resting horizontally across his notepad, and asked the question that has become the haunting refrain of :
“Just so I understand-is any of this going through an AI?”
Pauline had exactly three seconds. In those , a kaleidoscope of technical nuances flashed through her mind. She thought about the specific API she used to clean up the transcriptions. She thought about the “no-training” toggle she had checked, which she believed worked but had no way of verifying. She thought about the fact that her word processor now has a “suggested edits” feature that she can’t fully disable.
It was a perfect answer. It was professionally sound, it preserved her authority, and it was entirely true in spirit. It was also a lie by omission, because she couldn’t explain the middle part-the part where the client’s data traveled through a black box she didn’t own, hosted by a company that changed its Terms of Service this year.
The client nodded, seemingly satisfied, and moved on to the budget. But Pauline didn’t move on. She spent the rest of the afternoon staring at her coffee, feeling the weight of a secret she hadn’t actually intended to keep.
Earlier that morning, I had found in the pocket of a pair of jeans I hadn’t worn since the previous autumn. It felt like a small, unearned gift from the universe-a bit of grace in a week defined by spreadsheets.
But as I watched Pauline’s dilemma unfold in my mind, that twenty-dollar bill started to feel like a metaphor for the way we treat AI in the professional world. We find these efficiencies, these little “wins” tucked away in our workflows, and we’re so happy to have the extra currency that we don’t stop to ask how it got there or what we might have lost to earn it. We are all walking around with found money in our pockets, hoping nobody asks us for a receipt.
The Inherited Silence
The core frustration here isn’t a lack of ethics. It’s a lack of vocabulary. When a client asks if you’re using AI, they are asking a question about trust, but they are framing it as a question about tools. If Pauline had tried to be “honest” in that moment, the meeting would have died.
She would have had to explain the difference between zero-retention APIs and consumer-grade interfaces. She would have had to discuss the “residual knowledge” problem, where a model might not “store” your data but its weights are nonetheless subtly altered by the patterns of your logic. She would have had to admit that she is a tenant in a digital building where the landlord has cameras in the hallways, and she’s not entirely sure if those cameras can see through her office door.
This is the contrarian reality of the AI era: the silence is inherited. We are users of systems that were designed to be opaque, and then we are blamed for the opacity.
I once asked Yuki J.P., the lighthouse keeper who watches the grey stretch of the Pacific near the Oregon border, how he knows the lamp is actually working when the fog is too thick to see the beam.
“A light that doesn’t reveal the shore is just a distraction for the ships.”
– Yuki J.P., Lighthouse Keeper
We are currently in a period of thick fog. We’re shining our lights-our “disclosures” and our “transparency reports”-but they aren’t actually revealing the shore. They’re just making the fog look brighter.
The Carpenter and the Magical Plane
The professional relationship has always rested on the idea that the practitioner is the master of their own process. If a carpenter builds you a table, they can tell you where the wood came from, what grit of sandpaper they used, and why they chose a specific oil for the finish.
There is no part of that table that is “mystery.” But when we use generative models to draft a brief or analyze a dataset, we are using a tool that we do not, and cannot, fully understand. We are carpenters using a magical plane that sometimes chooses its own grain.
This creates a terrifying gap. When the client asks the question, they are assuming that we have the same level of control over our digital tools as the carpenter has over his saw. But we don’t. We are participating in a grand experiment in outsourced cognition. And because we don’t want to look incompetent or out of control, we give the “Pauline Answer.”
The danger is that trust has a shelf life. It functions on momentum. Right now, clients trust us because they’ve always trusted us. But as the gap between what we do and what we can explain grows, that momentum will eventually stall.
Someone will have a data breach that shouldn’t have happened. Someone will find their proprietary strategy reflected in a competitor’s AI-generated report. And then the simple questions will stop being simple.
Demanding a New Infrastructure
What we need is not more disclosure language written by lawyers to protect the firm; we need a fundamental shift in the infrastructure of the work. The reason Pauline felt like a liar is that she was using a system that forced her to be one.
If she had been using an Encrypted chat gpt solution-one where the data never leaves the encrypted tunnel, her answer would have been different.
Where identity is stripped before it even hits the server, and where “no storage” isn’t a pinky-promise but a technical impossibility. She could have said, “Yes, we use AI, and here is exactly why your data is safer with our AI process than it is in your own email inbox.”
But we aren’t there yet, at least not collectively. Most of us are still rummaging through the pockets of our old jeans, finding bits of efficiency and hoping the bill doesn’t come due. We are comfortable with the “found money” because the alternative-admitting that we are working in a fog-is too painful to contemplate.
I think back to that twenty-dollar bill. I ended up using it to buy a book I’d already read, just because I wanted a physical copy to hold. There is a desire for the tangible, for the proven, for the things that don’t disappear when the server goes down. In our work, that tangibility comes from being able to point to a process and say, “I know what happened here.”
The table becomes a barrier when the question is a key that no longer fits the lock.
The silence Pauline felt wasn’t her own; it was the silence of the black box she was using. We have to stop inheriting that silence. We have to demand tools that allow us to be as honest as we want to be. Until then, we are just professionals standing in the fog, waiting for someone to ask a question we aren’t allowed to answer.
It’s not just about data security; it’s about the soul of the profession. If we can’t describe how we think, then eventually, we’ll stop thinking for ourselves. We’ll just be the stewards of the output, the people who check the boxes and sign the contracts, while the real work happens in a server farm in a desert we’ve never visited.
That’s a high price to pay for found in a pocket. We should probably start looking for a way to pay it back before the interest gets too high.