The Tablet That Copied the Paper: Where Trust in Health AI Comes From

In 2007 I built a tablet application that added almost nothing new. Birth-to-three programs in Milwaukee kept each child's Individualized Family Service Plan, the IFSP, on paper. It is the plan every child in early intervention has: what the family is worried about, what they hope for, which services will come, and who is responsible. The forms had been designed by health care providers and used for many years. The people who filled them out were program associates, and once a form was filed, the information in it was hard to find again. I was a health informatics faculty member with a small team and a Tablet-PC, and our design was simple on purpose: we copied the paper form onto the screen. The program associates found it easy to use. There was no learning curve, because they already knew the form.

Key Takeaways

• The 2007 tablet was a rule-based system: people wrote every rule, so trust in the paper form carried over to the screen.
• In the metadata architecture of an AI-powered health record, time and source give AI its power, and each opens a new way to be wrong.
• Learned rules are not visible by default, so trust must be earned through transparency and governance. That work is real and urgent.

Most of us know the feeling: a new screen at the clinic or the bank that looks just like the old paper form, and we trust it without a second thought. That was my tablet. In 2026 the technology can do far more than copy a form. I keep returning to the 2007 project because it shows me the one thing the new technology does not get for free.

Where the Tablet's Trust Came From

Trust was not a problem we had to solve in 2007. The tablet was trusted because the form was trusted, and the form had earned that long before we arrived.

Three things stood behind it. Health care providers had designed the form, so it asked for what they needed to know. It had been in use for many years, so everyone knew its habits and its gaps. And nothing in it was hidden: every word on the page had been written by a person.

In the vocabulary of AI, the tablet was a rule-based system. People wrote every rule it followed: which questions to ask, in what order, what counted as a complete plan. The form was the rulebook, and anyone could read it.

When we moved the form to a screen, all of that came with it. The tablet changed where the information lived. It did not change what the information was, or who stood behind it. That is why the program associates could pick it up and keep working.

What AI Adds, and What It Does Not Inherit

Today the successor to a form like that is an AI-powered personal health record, and it is a different kind of thing. AI belongs to the world of machine learning, where people do not write the rules. The machine learns them from data. That is the source of its power: it can adapt to each family, keep adjusting as their situation changes, and do this across thousands of records at once.

It also has far more to learn from. The paper form held a snapshot, updated at each review. A record built for AI can hold a life in motion. In my textbook I describe eight kinds of metadata such a record needs. Two of them show both the power and the problem.

The first is time: when things happened, in what order, and how long after. Take migraine, a condition far from early intervention but easy to picture, and the case I use in a 2026 paper on AI agents in personal health records. No single reading explains a migraine. The pattern lives in sequence: an attack that follows a short night of sleep, after a delay. A person rarely sees that for themselves. A system that tracks time can. But people often log an attack after it is over. If the system reads the time of the entry instead of the time of the attack, it ties the migraine to whatever was logged nearby, such as a morning coffee, and reports a trigger that does not exist.

The second is source: where each piece of information came from. This lets a system weigh a sensor reading differently from a memory typed in days later. But the system writes into the record too. Its own earlier summary, "chocolate triggers your migraines," can come back later looking like new evidence, and the system ends up confirming itself.

Both mistakes arrive in fluent, confident sentences. "Avoid coffee" looks like advice. It is a rule the machine learned, and nothing on the screen says it rests on a logging habit.

This is why the trust does not transfer. On the tablet, every statement on the screen was something a person had written, following rules a person had written. In an AI-powered record, the screen shows statements nobody wrote, produced by rules nobody wrote. A program associate can no longer check the page against the rulebook, because there is no rulebook to read. This is the new challenge, and making the tool easier to use will not solve it.

How Trust Gets Earned

I do not think there is a shortcut, but the old form shows where to look. It was trusted for three reasons. Each one becomes a question you can ask of any AI tool put in front of you, in a clinic, an agency, or a classroom.

Who stands behind it? The form was designed by the providers who relied on it, and they answered for it. When people write the rules, that governance comes built in. When the machine learns the rules, the job has to be assigned on purpose: someone must decide what the system may do, check what it has learned, and answer when it is wrong. This work is real and urgent. The tools are already arriving.

What has it already done reliably? The form had years behind it. An AI tool has to start small: a few narrow jobs, done well, before anyone is asked to depend on it for more.

Can it show where each statement comes from? The form had nothing hidden. AI has to show its work on purpose: where each statement comes from, when the events behind it actually happened, and when the record is too thin to say anything at all. As I argued in Fluent is not the same as right, a polished answer tells you nothing about whether it is right. People need something they can check.

If the answer to any of the three is silence, the trust has not been earned yet.

What I do not know is how long this takes. The form needed years.

Observer's Insight

In 2007 we got trust for free. Providers had written the rules, years of use had made them trustworthy, and the tablet inherited all of it. AI learns its own rules, so it has no such inheritance. It can do far more than the form ever did, and for that reason it has to build its own track record: open about its sources, honest about its limits, governed by people who answer for it, one small job at a time.

Track Record Makes Trust

As a member of the Springer Nature Author Affiliate Program, I may earn a small commission from purchases made through this affiliated link to my book. Support the author by checking out my textbook, Artificial Intelligence in Public Health: https://tidd.ly/4mH9389