A professor asked me that question recently, sincerely, and with real concern: are we forcing students to use AI? My answer was immediate. No. But their employers will. Their fieldwork sites already are. Capstone sites are expecting it. And patients are using it, whether anyone gave them permission or not.

That's the entire case for AI literacy in one exchange, and it's why the new AI in Practice: Tech section exists.

Here's what I keep seeing that most of our profession isn't tracking closely enough. AI is being built directly into the technology of our field right now. Rehab tech. Agetech. Caretech. Documentation systems. Assessment platforms. The tools your clients will use at home, the systems your future employer will require, the devices being designed to do things occupational therapists have always done. This is happening with or without us in the room.

And that's the problem. Because right now, most of these tools are being designed for OT without enough input from OT. When a caretech company builds an AI that decides how to prompt an older adult through their morning routine, do they understand occupation the way we do? When a documentation platform bakes AI into clinical reasoning, is anyone at that table who was trained in clinical reasoning?

If we don't understand these tools, we can't shape them, evaluate them, push back on them, or advocate for our patients around them. We just get handed them and told to comply. Learning basic AI literacy isn't about becoming pro-AI. It's about earning a voice in the technology being developed in our own field, technology we, or our clients, will almost certainly have to use.

You don't have to like it. I've been plenty critical of AI in these pages, and I'll keep being critical. But "I don't like it" and "I don't understand it" are two very different positions, and only one of them lets you do anything about it. Refusing to learn doesn't stop the technology. It just guarantees you'll have no say in how it shows up in your practice.

This week's headline is a perfect example of exactly what I mean.

Researchers presented a new AI tool at the European Society of Cardiology congress in Munich that can detect signs of heart failure and heart valve disease from a standard ECG in less than two seconds. Here's why that's remarkable: the traditional ECG has been used for a century to read the heart's rhythm and rate, but on its own it cannot detect structural heart disease. That normally requires an echocardiogram, an ultrasound scan that patients often wait months to receive.

This AI extracts patterns from a routine ECG that the human eye simply cannot see. Trained on more than 1.6 million ECGs from Brazil plus several million recordings from the US, it identified up to 81% of heart failure cases and up to 90% of heart valve disease cases in a US study of roughly 67,000 patients. The team, led by Imperial College London, called it "superhuman AI," and said the next step is designing handheld AI-led ECG readers clinicians could use anywhere.

This is the same pattern we keep seeing in the good AI stories: the technology doesn't replace the clinician. It flags the patients who need a human's attention faster. The goal is to fast-track high-risk patients for the scans and treatment they'd otherwise wait months for, and even to catch heart disease opportunistically in people who had an ECG for some unrelated reason. Earlier detection means earlier intervention, which changes outcomes.

Watch the gap between "presented at a conference" and "in your hospital." These results are early and not yet peer-reviewed in final form, and "superhuman" is a headline word, not a clinical guarantee. Watch also for the handheld version, because that's when this stops being a research story and becomes something you might actually encounter in a care setting.

Follow the timeline, because OT lives all along it. Earlier detection of heart failure and valve disease means patients enter the care continuum sooner, which shifts when cardiac rehab begins, when energy conservation and activity tolerance work starts, and how we plan for the occupational impact of a diagnosis. A patient flagged early is a patient we may see earlier, at a different stage, with different goals.

But the bigger point is the one from this week's hot take. This is AI aimed at the right problem, by the right people, augmenting a clinician rather than replacing one. That is the model of AI integration we should be advocating for across healthcare. And notice: this tool was built by cardiologists and imaging scientists who understand the heart. The tools being built for our field need that same thing: people who understand occupation in the room. That only happens if we're literate enough to be there.

Back in issue #10, I shared an "AI for good" story: researchers at the Rochester Institute of Technology using AI-powered drones to detect landmines faster and more safely than ground-based methods. The "what to watch for" I flagged then was specific. The team had promised to release their dataset as open-access, and I wrote: watch how open science accelerates humanitarian AI.

They followed through. The RIT team, working with the nonprofit Demining Research Community, collected a large-scale, multi-sensor dataset at a controlled test field in Oklahoma, deploying over 140 inert landmine and unexploded ordnance targets across surface, partially buried, and fully buried configurations. A portion of that data, a hyperspectral collection, has now been released through a peer-reviewed publication and will appear in the IEEE InGARSS 2025 proceedings, with the full dataset pending journal review. In other words, the open-access promise moved from intention to published reality.

And it went international. The team partnered with the Royal Military Academy of Belgium, scattering over 110 replica mines across varied terrain and vegetation to test how the system performs outside a single controlled field, which is exactly the kind of real-world validation that separates a promising demo from a deployable tool.

The humanitarian-AI story I told in issue #10 is quietly doing what I hoped it would. The benchmark datasets are becoming public, which means researchers worldwide can now build on this work instead of starting from scratch. This is slow, unglamorous, deeply important progress. It won't trend on social media. But somewhere down the line, a farmer walks back onto safe land, or a child walks to school on a cleared path, because open science let this work move faster. That's still what "AI for good" looks like, and six months in, it's still delivering.

(I'll keep following this and report on it periodically.)

The first of our new rotating spotlights. This one connects directly to this week's headline.

This week's heart-disease AI is one example of a much bigger shift already reaching rehabilitation: predictive analytics. These are machine learning models that forecast a patient's likely functional outcomes from data collected at admission, before the recovery even unfolds.

I presented on this earlier this year during a neurorehab symposium.

The model is trained on huge sets of past patient data, learning which admission-day factors tend to predict which outcomes. Then, for a new patient, it takes their clinical and imaging data and produces a forecast: how much function they're likely to recover, and how quickly.

In stroke care specifically, this is already striking. Machine learning models are predicting three-month functional outcomes in acute ischemic stroke with AUC scores above 0.90, which is a high level of accuracy. The strongest predictors the models lean on will sound familiar to any neuro OT: stroke severity, age, the timing of rehabilitation initiation, and dysphagia screening.

Look at that list of predictors again. "Rehabilitation initiation timing" is on it. That's us. The timing of when OT begins is one of the factors these models weigh most heavily, which is a data-backed argument for early referral that we can actually point to.

More broadly, better prediction enables better-targeted intervention. If a model can forecast a patient's likely trajectory, we can calibrate the intensity and focus of therapy to match, maximizing the neuroplasticity window when it matters most. This is the neuroplasticity principle we've talked about all along: challenge at the right level, at the right time, drives recovery.

The promise is precision. The risk is misplaced trust. A prediction is a probability, not a destiny, and a model that forecasts a poor outcome must never become a reason to under-treat a patient or write off their potential. That would be the exact opposite of what OT stands for. As these tools enter rehab settings, the OT in the room has to understand what the model is, and isn't, saying, so that a forecast informs care without capping it. Which is, once again, why the literacy matters.

Staying current isn’t optional. It’s beneficence.

You can fundamentally disagree with AI. You can be critical of how it's built, who profits, and what it costs. I am, often. And at the same time, you carry a professional duty that doesn't bend to your feelings about a technology: the duty to stay current enough to give your patients the benefit of the best available options.

Continuing education about emerging tools, including AI, is part of beneficence. It's how we ensure our patients benefit from what's actually possible for their recovery and quality of life.

Our AOTA Code of Ethics frames beneficence as acting for the good of our clients, and that includes a clear, often-overlooked obligation: maintaining competence and staying current with the knowledge and tools that affect care. Beneficence isn't just being kind in the moment. It's making sure the care you provide reflects what's genuinely available to help someone, not just what you happened to learn in school.

Think about this week's stories through that lens. If an AI tool can flag heart disease earlier, or a predictive model can help time a stroke patient's rehab to maximize recovery, then those tools are becoming part of what "good care" means for those patients. A clinician who refuses to understand them isn't staying neutral. Over time, they risk falling behind on options that could genuinely benefit the people in their care.

Ask yourself:

You can hold both. You can advocate loudly against AI's harms and still learn it well enough to serve your patients, because those aren't opposites. They're two parts of the same professional responsibility. Beneficence asks us to know what's possible for the people in our care. In a field where AI is arriving whether we like it or not, staying literate is how we keep that promise.

Here's why I keep coming back to the practical stuff, and why the paid issues this month go hands-on with it.

Understanding where AI is showing up in our field, like the predictive analytics spotlight above, only matters if you also know how to work with these tools, evaluate them, and use them without handing over your clinical judgment. That's what the rest of September is for. The paid issues go into real scenarios: clinical situations, faculty and research workflows, and the business-side tasks that quietly eat your week. Each includes a scenario, a "which tool would you pick" poll, and a real look at how I'd actually do it.

If the Tech spotlight is the why, the paid issues are the how. And if you've found this useful, the most valuable thing you can do is forward it to one colleague who still thinks they can sit this out.

If this is your first article, thank you so much for being here! As I build this, I want to make sure it’s helpful to you. Please comment or email with any feedback or suggestions!

What would you like to read more of? Please let me know via email!

Until next time,

Pooja A. Patel, DrOT, OTR/L, BCG

learn. innovate. empower.