---
title: "When GPS Stops at the Waterline: What Scuba Diving Taught Me About AI"
summary: "A dive in Kos became a practical lesson in underwater navigation, human judgement, AI estimation and why human-in-the-loop systems still matter."
datePublished: 2026-07-25
dateModified: 2026-07-25
originalUrl: https://www.neuvottelija.fi/openclaw/kun-gps-loppuu-vesirajaan
originalLang: fi
authors: ["Sami Miettinen"]
tags: ["scuba diving","artificial intelligence","human in the loop","underwater navigation","Kos"]
canonical: https://ai.neuvottelija.com/when-gps-stops-at-the-waterline-ai-diving/
---
# When GPS Stops at the Waterline: What Scuba Diving Taught Me About AI

🇫🇮 [Lue suomeksi / Read in Finnish](https://www.neuvottelija.fi/openclaw/kun-gps-loppuu-vesirajaan)

I am writing this from Kos, where I recently went diving with my dive buddy Chris. The short video has already landed on YouTube — no narration, no captions, just the dive:

[Watch the Kos diving short](https://youtube.com/shorts/0-8COc1XNDU?feature=share)

I am a **PADI Advanced Open Water Diver**, but underwater navigation is still not my strongest skill. When visibility changes, currents build or the shoreline disappears, I rely on experienced dive guides.

That is not a contradiction. It is risk management — and a useful way to think about artificial intelligence.

## GPS ends at the waterline

GPS depends on radio signals. Seawater is electrically conductive and attenuates radio waves so effectively that satellite positioning becomes practically useless once the receiver is submerged.

Underwater navigation therefore depends on three things: dead reckoning, acoustics and estimation.

The same is true of many AI systems. They do not observe reality directly. They infer a state from partial, noisy and sometimes stale signals.

## How a diver estimates position

![Underwater navigation with a compass off Kos](/media/openclaw/when-gps-stops-at-the-waterline-underwater.jpg)


A recreational diver combines several imperfect sources of information:

- compass headings and reciprocal bearings;
- distance estimated from kick cycles or elapsed time;
- depth as a contour line;
- natural features such as bottom slope, sand ripples, sunlight and vegetation;
- occasional surface checks.

Each source contains error, and the errors compound. A five-degree heading error over 300 metres can leave a diver more than 26 metres off course. Current changes kick efficiency. Fatigue changes pace. Stress changes perception.

An experienced guide continuously reconciles all of these signals, often without making the process visible.

## Machines already beat humans at state estimation

Autonomous underwater vehicles use combinations of inertial navigation systems, Doppler velocity logs, sonar, acoustic positioning and terrain matching.

Machine learning increasingly supports this stack by estimating velocity when a Doppler system loses bottom lock, recognising previously visited terrain and complementing traditional filtering methods.

This is a state-estimation problem — exactly the kind of bounded, data-rich problem at which machines can outperform people.

Consumer technology is moving in the same direction. Acoustic systems such as Garmin SubWave can help divers communicate and maintain direction and distance to a known surface point. It is not underwater GPS, but knowing where “home” is solves much of the practical problem.

## What the guides in Kos actually did

The difficult part of our dive was not merely determining coordinates.

Conditions at the surface were rough. The guides had to assess, at the same time:

- remaining gas;
- fatigue and stress;
- waves and surface current;
- visibility;
- the safest exit point;
- whether the original plan should be abandoned.

They made a decision with incomplete information, under changing conditions, and carried responsibility for the outcome.

That is not just estimation. It is judgement.

## Human in the loop is not a temporary flaw

AI discussions sometimes treat “human in the loop” as an interim stage that will disappear when models become more capable.

Diving suggests a different conclusion.

The machine may estimate my position more accurately than I can. It may model current, calculate gas consumption and recommend the statistically safest route. But somebody still has to decide whether the plan itself is still valid.

That distinction also appears in mergers and acquisitions, negotiations and board decisions. A model provides an estimate. Experience tells you when the estimate no longer deserves to control the decision.

The human remains in the loop not merely because the machine is imperfect, but because uncertainty, responsibility and irreversible consequences belong together.

## Knowing your limits is a strength

I know that underwater navigation is a relative weakness of mine, despite my certification and diving experience. So I use a guide.

Companies should approach AI in the same way. They should not reject a system because it performs some tasks better than people. Nor should they delegate every decision merely because the output appears precise.

The strongest operating model is a division of labour:

**Machines estimate. Humans decide.**

## The limits of AI dive technology

AI-assisted diving will improve. Better sensors, acoustic networks, computer vision and compact inertial systems will make navigation, buddy tracking and risk alerts more accurate.

But recreational diving places hard limits on the technology:

- seawater blocks ordinary GPS and radio communication;
- acoustic systems require compatible equipment and have limited range and bandwidth;
- inertial systems drift unless they receive an external correction;
- cameras and sonar lose reliability in poor visibility, bubbles and complex terrain;
- equipment cannot fully observe panic, discomfort, judgement or the changing capabilities of every diver;
- power, cost, bulk and failure modes matter more underwater than in a software demo.

An AI system can improve situational awareness. It cannot make a dive consequence-free, and it does not take moral or legal responsibility for the group.

## The water does not negotiate

I have spent much of my career negotiating mergers, acquisitions and financing transactions. The sea is less accommodating.

It does not care about confidence. It does not reward exaggerated capability. It exposes mistakes quickly.

That is why diving is such a useful analogy for AI deployment. Sensors improve. Models improve. Estimates improve. Yet experience, judgement and responsibility remain stubbornly human.

At least for now.

*This article discusses general principles of underwater navigation and AI. It is not a substitute for scuba training, local briefings or professional dive guidance. Always dive within your certification, experience and the prevailing conditions.*