News
Where the learning runs: the first question for an AI feature
November 17, 2026 · PenDraco
Of 5,101 U.S. adults in a Pew survey, 73% feel they have very little or no control over the data companies collect. 56% frequently click agree without reading. Where an AI feature does its learning is a stronger control than consent.
In May 2023, Pew Research Center asked 5,101 U.S. adults about their personal data. Of those adults, 73% feel they have very little or no control over the data that companies collect about them. And 56% frequently click agree on a privacy policy without reading it.
Consent puts the protection of data on that click. A phone keyboard shows a stronger control. The control is a design choice: where the feature does its learning.
A keyboard is the hard case
A keyboard reads what a person types, one key at a time. Apple writes the rules for keyboards on its phones. "Keyboards handle some of the most sensitive user data", its documentation says.
The documentation adds that a keyboard can read the selected text and the text around the cursor. There is one limit on that reach. In a secure text field, "the system temporarily replaces your custom keyboard with the system keyboard". So a third-party keyboard does not see a password that a person types there.
A keyboard also learns. It improves its suggestions from the words a person uses. That makes it a small AI feature with a sensitive input. An AI feature that learns from a company's data has the same shape.
The platform's control rests on one switch
Apple gives a keyboard 2 modes. The first is the default. "By default, a keyboard has no network access and cannot share a container with its containing app", Apple's guide says. For that keyboard, "the system ensures that keystrokes can’t be sent back to your server or anywhere else".
The second mode is open access, and it removes that block. The user grants it with a switch in Settings. Apple describes what the user knows at that moment. "When they enable this, they know their keystrokes are available to the keyboard developer", the documentation says.
Read that sentence beside the Pew figures. The control assumes a person who reads one switch and understands it. More than half of the adults in the survey frequently click agree without reading. This is PenDraco's reading of the 2 sources together: consent is a weak control for a product to rest on.
The learning can run where the typing happens
The stronger control is to keep the learning on the device. Then the typed text does not need to leave the phone.
Public research shows that this works for a keyboard. Hard and 8 co-authors published the study in 2018. It shows "the feasibility and benefit of training language models on client devices without exporting sensitive user data to servers". The team trained a next-word model on phones in place of a server.
The accuracy result matters most to a buyer. The team tested both models on data held on the phones. The model trained on the phones beat the model trained on a server. Its top-1 recall was 0.8 points higher, a relative gain of 5%.
The study also shows the cost. A phone took part only while it charged and sat idle on an unmetered network. So the learning used hours when the phone had spare capacity. A business system can do the same: train at night on the hardware that already holds the data.
The idea has a name in privacy practice. FTC staff describe data minimization as the concept that "companies should limit the data they collect and retain". The NIST Privacy Framework lists a related outcome: process data "to limit observability and linkability". One of its examples is that "data actions take place on local devices". When the learning stays on the device, the company needs no copy of the typed text.
iSureType keeps its learning on the phone
PenDraco built iSureType, a keyboard for iOS, on that rule. The keyboard learns the words a person uses, and it does that work on the phone. The source code of the keyboard has no network calls.
The product does ask for open access. Its App Store listing gives the reason: "Allow Full Access on the keyboard so it can read the shared clipboard". With that switch on, Apple's system block does not apply. The protection for typed text is the design of the keyboard. A buyer can inspect a design. A policy asks for trust.
The same rule guides PenDraco's client work: put the learning where the data already is. PenDraco builds AI systems that run inside the client's environment.
What the sources do not show
Pew asked adults about companies in general. It did not ask buyers of business software, and it did not ask about keyboards.
The keyboard study covers one product and one task. On data that a server logged, the recall of the 2 models was comparable. Its method still uploads model weights from the phones to a server. The authors say that users must "trust that the aggregation server will not scrutinize individual weight uploads". The study trains one shared model from many phones. It does not describe iSureType, whose learning stays on one phone.
We did not measure iSureType against that study's benchmark. The 2 statements about its code describe the keyboard, which is the part that reads the keys.
The NIST framework is voluntary. Its cover page says its contents "do not have the force and effect of law". This article reports what the documents say. It is not legal advice.
Ask where the learning runs before you buy
The decision comes before any comparison of features. A feature that learns from sensitive data learns somewhere, and the buyer chooses where.
- List the data that the feature reads, and mark the sensitive fields.
- Ask where the model learns from your data: on your own systems or on the vendor's servers.
- Ask what leaves your systems while the model learns. Ask for the design, not the policy.
- Test accuracy on your own data before you accept that the learning needs the vendor's servers.
If an AI feature in your company reads sensitive data, start with step 2. The embedded sprint puts one workflow into production in 30 to 60 days. Get in touch.
Sources
- Pew Research Center, How Americans View Data Privacy, October 18, 2023: the 5,101 adults, the 73% and the 56%. - Apple, Configuring open access for a custom keyboard, read November 12, 2026: the system block, the switch and the sentence on sensitive data. - Apple, App Extension Programming Guide: Custom Keyboard, updated October 19, 2017, archived: the default mode and the secure text field. - Apple, Creating a custom keyboard, read November 12, 2026: what a keyboard can read. - Hard and 8 co-authors, Federated Learning for Mobile Keyboard Prediction, arXiv, version 2 of February 28, 2019: the training on phones, the recall result, the idle phones and the trust in the server. - Federal Trade Commission staff, Internet of Things: Privacy & Security in a Connected World, January 2015: data minimization. - National Institute of Standards and Technology, NIST Privacy Framework, Version 1.0, January 16, 2020: subcategory CT.DP-P1 and the cover page. - Apple App Store, iSureType, read November 17, 2026: the request for Full Access. - The statements about the keyboard's learning and its source code are PenDraco's reading of its own code for the App Store build.