Sensing Our Way In The World

On war robots, newborn care, and a holistic view of perceiving machines.

Edited by Althea May Atherton.

---

I have spent the last two weeks acutely aware of sensors and sensor error. Newborns can communicate in finite ways, useful signals all, but a scream is an alert, not a diagnostic. For those who require immediate attention in a Neonatal Intensive Care Unit, or NICU, being born means a rapid transit from nude in the womb to swaddles and an array of sensors, providing doctors and nurses readings on heart rate, respiratory rate, and blood oxygen levels, among others. These are useful as both tracked and diagnostic measures, and in the NICU or elsewhere, the sensor data is routed and charted through a computer and display, which can automatically trigger alerts if, say, blood oxygen drops too low.

What makes this a Wars of Future Past post, and not just a tired new father scraping content from a hospital stay, is the way those sensor readings diverge from reality. No sensor system survives contact with a baby.

"Pulse oximetry," explains Johns Hopkins , "is a test used to measure the oxygen level (oxygen saturation) of the blood. It is an easy, painless measure of how well oxygen is being sent to parts of your body furthest from your heart, such as the arms and legs."

The clip-on probe sends light through the extremity and reads that information, which is then translated into a percentage of blood oxygen level, and displayed. Althea and I first picked up a home pulse oximeter in the earliest days of the ongoing Covid-19 pandemic as another home diagnostic tool. These past couple weeks, pulse ox readings have defined our child's hospital stay. For infants, the pulse oximeter is a light-emitting tube taped to either the child's foot or hand.

Most of the time, when a baby sleeps, the sensor can stay in place just fine, though even a sleeping baby can displace the reader with toes or fingers. Awake, and especially awake and unswaddled, a pulse ox is trivial to disrupt. When that happens, through action as routine as a stretch, burp, or cry, the machine reading the pulse oximeter assumes its data remains correct and the patient is in danger. Then, in the hospital machines monitoring our kiddo, that assumption and data can after a short interval turn into a beeping alert, and then a blaring crisis.

This happens, especially, when changing a diaper.

Kelsey Atherton in the NICU, with his finger gripped by the hand of his newborn. It was a sensor-rich environment, this selfie camera was about the least intrusive angle. (Photo by author)

It would, from a hospital standpoint, likely be malpractice to massively increase the warning threshold, or to toggle back the care-summoning automated alerts. A machine with false alarms is far more lifesaving than a machine with insufficient alarms.

Still, on the fifth bleating blare of the bottle field, triggered by a baby's toes curling in contentment, it is hard not to ask why we rely on a tool so prone to error in this way.

My beat, generally, covers tools concerned with rapidly and involuntarily expediting the end of life. "First, do some harm," is a not-inaccurate assessment of military purpose. The big questions of military technology are how to facilitate, direct, and control that harm, and these are questions increasingly mediated by sensors and sensor reading analysis.

When I write about robots, what I am really writing about are machines that have the shortest possible distance between sensor input and machine reaction.

Sensor error, then, is an essential and inevitable part of the story of automation. On July 3, 1988, while operating in the Persian Gulf, the cruiser USS Vincennes' automated AEGIS anti-air system detected a plane approaching that potentially matched a fighter jet known to be in Iran's military. Errors in understanding and checking the sensor reading, as well as the decision by a nearby vessel to not pass along contradictory information, all contributed to the Vincennes' commanders decision to shoot down the detected plane. Two anti-air missiles destroyed Iran Air flight  655, killing all 290 people on board.

Relatives of victims of Flight 655 visit location of shootdown, image by Majid Jamshidi of Fars News Agency. Via Wikimedia Commons, licensed under a Creative Commons Attribution 4.0 International License.

Vincennes is among the highest profile and cleanest cut examples of sensor error, along with human error, resulting in such an entirely avoidable tragedy. Among the problems identified by the Navy afterward was an eagerness among the crew to fight, an information display system that was confusing to read, and human trust of machine identification of a threat.

Sensor error becomes part of a broader problem, where humans assume machines to be less fallible than they actually are. In war, that error can come at every part of the process, and accurate sensor readings can give false confidence to wrong assessments made by the humans directing the machines.

There are ways to avoid and reduce these errors, which are good and important, but the nature of the work of war makes some such error inevitable for a profession built to direct harm at speed.

It is also a case for everyone, in every role in life, to consider sensing machines as not describing reality perfectly, but as potentially flawed observers. After a week of deep familiarity, I now see the pulse oximeter as a machine that is sometimes right when it is blaring an alert. As my child slept softly beneath the pediatric klaxons of an ICU bed, I checked to see if the actual culprit was not bad blood but curled toes.

---

Thank you all for reading. Our child was born early this month and as you can tell, we’ve had the opportunity to get familiar with the sensors that oversaw his first days. I am, in a fashion, still on “freelancer paternity leave,” which is trusting by grace that what I saved up beforehand is enough to carry my family through some vital but unpaid time off. Supporting this newsletter is a way to support me and Althea and Baby Atherton through all of that.

I love writing for you. I drafted this cribside in the hospital a few days ago (everyone is home and fine now). It means so much to have the support I get from readers like you and to write directly, without having to convince an editor that the sensor on my baby’s foot is a good angle for talking about war robot sensor error first.

Thank you for your continued support and patience, and I look forward to writing for you more in the future.