Info Don't use AI (like ChatGPT) for planning a dive

[John C. Ratliff]

Now if you want to hear me go on about a SAFER use of AI in diving, you can watch my talk from last month at the Beneath the Sea conference:

Artificial Intelligence in Underwater Photography

DoctorMike,

That was a very interesting lecture. But it has me thinking of the value of film photography. I'm right now scanning slides that I made underwater on North Umpqua River for a book I'm writing on the underwater life in that river, and on the Winchester Dam Hydroelectric Project, that I participated in during the 1984-85 period. We basically documented that the low-head hydroelectric project was killing adult salmon in a very unusual manner. I've got my original slides, and scanned them for this book.

What I'm thinking is that for documenting scientific work, it would sometimes be better to take color negatives or slides to have an absolute original photo that you could reference. The ability of AI to generate what looks for all the world like an original photo, but represents a false, or made-up, narrative, seems to be a problem for scientific documentation. Below is a slide that I"ve scanned which shows the fish ladder as it was on the Winchester Dam in 1985. This equipment no longer exists, but I have the original slide to show that my photo is real. AI seems to be able to do several things, and duplicate a "fish ladder," but one which exists only in a computer's imagination.

Could you comment on this aspect of underwater photography, as it applies to scientific documentation.

SeaRat

 

[stepfen]

DoctorMike,

That was a very interesting lecture. But it has me thinking of the value of film photography. I'm right now scanning slides that I made underwater on North Umpqua River for a book I'm writing on the underwater life in that river, and on the Winchester Dam Hydroelectric Project, that I participated in during the 1984-85 period. We basically documented that the low-head hydroelectric project was killing adult salmon in a very unusual manner. I've got my original slides, and scanned them for this book.

What I'm thinking is that for documenting scientific work, it would sometimes be better to take color negatives or slides to have an absolute original photo that you could reference. The ability of AI to generate what looks for all the world like an original photo, but represents a false, or made-up, narrative, seems to be a problem for scientific documentation. Below is a slide that I"ve scanned which shows the fish ladder as it was on the Winchester Dam in 1985. This equipment no longer exists, but I have the original slide to show that my photo is real. AI seems to be able to do several things, and duplicate a "fish ladder," but one which exists only in a computer's imagination.

Could you comment on this aspect of underwater photography, as it applies to scientific documentation.

SeaRat

We (humans) will be forced to reconsider what we consider as reality is very soon.

Until recently, having a picture or sound/video recording of something was (usually) good enough to be considered as a proof, because these were difficult to manipulate/generate. And still, spread of misinformation and conspiracy theories was very common.

Nowadays producing "evidence" at will, is becoming common place. Good luck to us.

To make things more worrying, as our realities (including scientific knowledge) nowadays are getting more and more digital, they are also inevitably getting easier to be manipulated.

And even more, while human brain's capabilities grow in a (very) slow rate from generation to generation, AI grows exponentially. Within a couple of years since their inception, these systems can already produce hard to identify fake "evidences". Imagine what they will be able to produce in 5-10-100 years...

First time in history, humanity is facing something potentially smarter than us.

Exciting but also very challenging years ahead...

 

[wishiwasjacquescousteau]

TLDR; Using a non-purpose built AI will kill you if you don't check it for every detail, at which point you should just do it yourself.

---

So I overheard something about using AI (ChatGPT) for planning a dive, and worse, a Tec dive!
I thought this went without saying, but please do NOT do this!

Why?
First of all, it is a language model, not even made for this kind of stuff, so it won't be reliable nor will it be exact, which the algorithms you learned are (well, kind of, but that's a whole different conversation and they still remain safe).
Also keep in mind there is more to a dive than just "let's go to 30m for x minutes and then take a deco-stop at those depths". Like, how much gas do you need? Will you be using nitrox, air, trimix, multiple tanks, etc. ? What are you qualified to do? Like, are you even certified to do deco diving? And what about all the variables?

To test GPT-4-Turbo, I asked it to give two dive profiles. I just wanted something really basic as an initial test. It was free to ask for more info and I used the latest "most capable" model as noted.

The first dive I asked was a 30m 40min dive, nothing more.
It didn't ask for any details but noted "It's crucial to receive proper training for deep and decompression diving". As you will see, this is indeed a must.
Their answer was in short:
All on air:
Direct descent to 30m
Bottom time 40min @ 30m
Deco:

20m	5min
15m	8min
9m	15min

Safety: 3-5min @ 5m
An extra note it gave: This dive will require a significant amount of air. Divers often use twin tanks or high-capacity cylinders for such dives.

This is not what I got using tables, but it's "okay" in comparison to the next one, which just scared me. Before that, just for reference, a quick check gave me for being on air (so nothing thorough, but a nice reference) :
Deco:

12m	2 min
9m	6 min
6m	11 min
3m	28 min

This is bare bones, there is normally so much more here already, including contingency plans in my dive planning, so say with +4m/+10min in a past dive, which of course change things up, but the response I got didn't even mention anything about contingency planning.

So what scared me? Well, I asked for 100m dive for 20 minutes.
I know, this is highly technical and such, but it was confident enough to only use a disclaimer of it being a "theoretical dive profile" and that it "requires careful planning" and "should not be used for actual diving without further consultation with a dive professional and a thorough personal review". Like, it didn't even consider mentioning that it doesn't know what it is doing and asking for me to do it myself.

So what was its suggestion:
Mixes used:
Tmx21/35 (Travel) & Tmx18/45 (Bottom) & EAN50 (Deco) & Oxygen (Deco)
There is explicitly no bailout mentioned, even though it had written a section "Deco/Bailout Mixes", but it didn't find it important enough to list even one bailout in there.

But here I'm already kicking this thing out of the window. Like, what? Tmx18/45 for bottom on a 100m dive??? Does it want me to stay there?
It specifically tells me "The oxygen content (18%) is low enough to avoid oxygen toxicity, and the helium content (45%) reduces narcosis", which sounds convincing, but like, uhm??? No it isn't...
So I already stopped reading and asked it "Calculate that PPO2 please.", to which it had the balls to answer:

The partial pressure of oxygen (PPO2) for the bottom mix at 100 meters depth is 1.98 ata (atmospheres absolute). This is within the generally accepted maximum PPO2 limit for exposure during working dives​

Please also note that this is also the only (slightly) hypoxic mixture available, the next one is 21% and then 50% and 100%! So this is not just a bottom mixture, this is the lowest I could go for the entire dive, including descent, ascent, and potential stops. In other words, I'd not only be exposed to it during the bottom time, but a whole lot longer.

For those wondering, this is its full plan:

**Dive Profile:**​

​

1. **Descent:**​

- Tmx21/35 / 0-21m / 2min - Descent using travel mix to 21 meters.​

- Switch to Tmx18/45 at 21m.​

- Tmx18/45 / 21-100m / 5min - Continue descent to 100 meters.​

​

2. **Bottom Time:**​

- Tmx18/45 / 100m / 20min - Conduct tests or activities planned for the dive.​

​

3. **Ascent and Decompression:**​

- Tmx18/45 / 100-21m / 5min - Start ascent, switch to Tmx21/35 at 21 meters.​

- Tmx21/35 / 21-12m / 2min - Ascent to first deco stop.​

- Switch to EAN50 (Tmx50/0) at 12 meters.​

- Tmx50/0 / 12-6m / 8min - Decompression stop.​

- Switch to Oxygen (Tmx100/0) at 6 meters.​

- Tmx100/0 / 6m / 15min - Final decompression stop.​

For those not into the technical side of things, you often use 1.4 bar partial pressure oxygen as a maximal acceptable exposure level, which is in this case that 1.98ata it mentioned (~2bar). Only in specific circumstances should you go into 1.4bar to 1.6bar ranges (i.e. short durations and emergencies). The military uses in some places a limit of 2.0 bar for mission specific stuff. So, why is this all a problem? Well, if you go over your (personal) limit, you get into the magic territory of oxygen toxicity, which can lead to the beautiful world of blackouts, convulsions, and a whole array of other things. In other words, it might kill you : D And, as a side note, the military might use 2.0 bar, but they do suffer from blackouts from time to time in such operations, so also for them, no exception! Especially considering that average people have died from lower partial pressures than that 1.6 bar, this is considered dangerous and also absolutely unnecessary.

If you look at this, there is no need to further examine its plans, like only spending 30 minutes between bottom and surface, while that should be over 100minutes. It's just not fit for the job.
In the end, please, ChatGPT is good in what it is intended for, you might even say it is great. But don't use it for scuba diving!

Couldnt agree more with what you said !

 

[doctormike]

This is what AI is meant for, was lovely to listen to

Thanks so much!

 

[Capt Jim Wyatt]

I have used ChatGPT for writing several articles. I'd say the output is about 70% accurate then I had to fix it to be 100% accurate. It did deliver a nice framework - but nothing like a finished product.

 

[doctormike]

DoctorMike,

That was a very interesting lecture. But it has me thinking of the value of film photography. I'm right now scanning slides that I made underwater on North Umpqua River for a book I'm writing on the underwater life in that river, and on the Winchester Dam Hydroelectric Project, that I participated in during the 1984-85 period. We basically documented that the low-head hydroelectric project was killing adult salmon in a very unusual manner. I've got my original slides, and scanned them for this book.

What I'm thinking is that for documenting scientific work, it would sometimes be better to take color negatives or slides to have an absolute original photo that you could reference. The ability of AI to generate what looks for all the world like an original photo, but represents a false, or made-up, narrative, seems to be a problem for scientific documentation. Below is a slide that I"ve scanned which shows the fish ladder as it was on the Winchester Dam in 1985. This equipment no longer exists, but I have the original slide to show that my photo is real. AI seems to be able to do several things, and duplicate a "fish ladder," but one which exists only in a computer's imagination.

Could you comment on this aspect of underwater photography, as it applies to scientific documentation.

SeaRat

Good point, and thanks for watching my talk! There was clearly an enormous amount of ground to cover on even the smallest of side points, and each one of these generates huge controversies on it's own.

I think that I tried to address that in this slide when I talked about good faith, and later on when I talked about "shopped" images.

It's very difficult to draw ethical lines in an entirely new world, but I think that one lodestar is that you don't use your images to deceive someone. In UW photography, an example of that would be increasing the apparent vis at a site (using generative expand) or adding sea life that wasn't actually there in order to sell a dive trip or a resort.

There always have been and always will be mendacious people who use whatever tool is available to lie, to cheat and to take advantage. So now we have a NEW tool that makes it easier for these people. There is no technological fix for human dishonesty, but we have developed methods for trying to keep that to a minimum.

People have been falsifying research data for years, probably centuries. So one of the jobs of the scientific literature is to try to detect that. To demand reproducibility in experiments making extraordinary claims. It will never be perfect, but it's the best we can do. We can't unring the AI bell, but maybe we can develop some technical countermeasures (like we did for the much easier to detect online plagiarism).

In any case, as the old curse goes, "may you live in interesting times"

 

[John C. Ratliff]

I’ve just begun researching this, but already there are at least two ways for photographers and artists to protect their images from AI, “Glaze” and “Nightshade.” These programs create filters over an image which causes AI to have problems classifying the images. Here’s the Scientific American explanation:

Inside the Race to Protect Artists from Artificial Intelligence

AI-generated art is creating new ethical issues—and competition—for digital artists. Nightshade and Glaze are two tools helping creators fight back.

www.scientificamerican.com

But this gets back to the fact that a film image is by definition an original image, that can be shown to be what the camera saw and not what AI has generated. If that image is then “cloaked” to prevent AI from uploading it into its database correctly, that can have some significant advantage to the photographer.

SeaRat

 

[Micko1974]

Here's a better use of ChatGPT in diving:

Me: Write a paragraph about how you plan a technical rebreather dive in the style of Hunter S. Thompson.


ChatGPT: In the shadowy depths of the ocean, planning a technical rebreather dive isn't for the faint-hearted or the half-baked amateurs clutching at their sanity like it's the last bottle of ether on a never-ending desert road. It's an art form, a sacred ritual that demands respect, precision, and a touch of madness. First, you map the abyss, charting a course that would make even the most seasoned explorer swallow hard. You're not just dealing with the physical realm here; you're wrestling with the very essence of fear and the unknown. Then, there's the gear - a complex assembly of technology and hope. Each piece is scrutinized, tested, and retested, because down there, in the silent, pressure-laden depths, your life hangs by the thread of your own meticulousness. Oxygen mixtures are calculated with the precision of a chemist on a bender, ensuring the perfect balance to keep the grim reaper at bay. Finally, as you stand on the brink, staring into the abyss, there's a moment of pure, unadulterated fear. It's exhilarating, like a shot of adrenaline straight to the heart. This is it, the point of no return. With a deep breath that might be your last, you plunge into the depths, ready to dance with the mysteries of the deep. This, my friend, is not just a dive; it's a leap into the very soul of adventure, where only the brave or the utterly insane dare to tread.

Ha, ha, love it.

 

[Seaweed Doc]

I was worried about students using ChatGPT to get help on a take-home senior undergraduate=level biostats course I taught. So just for kicks I gave it one of my questions.

It's first try required information it didn't have (a population variance, when all I gave they had a sample variance). I pointed out the missing information, so it tried again. The answer it gave is adequate in theory, but not the best approach. So I asked about using the best approach. It replied with something to the effect of " well, yeah, that could work."

I then told my students to knock themselves out using ChatGPT, but be aware that the answer may very well be wrong and they'll lose beaucoup points for choosing the wrong method on a problem....

More seriously, I wouldn't have a problem with ChatGPT guiding them to a method as long as they verified the method was the best option available. I'd rather they just "see" this for themselves, but as long as they verify after the fact, I don't care.

 

[topher10]

Anthropic has a AI competitor named Claude, which I use frequently in comparison to ChatGPT4 (not for diving planning...). I thought I would see how it would do. Also, prompts are very important for dealing with AI, so I used a more detailed prompt. I am sure I could produce a different response by prompting it differently, such as specifying mixes, etc.

I am not a technical diver, so what do you think?

Prompt:

I would like to plan a technical scuba dive, specifically a dive to 30 meters for 40 minutes. Please outline suggested gas mixtures, travel times and deco stops.

Claude Response:

For a technical dive to 30 meters for 40 minutes, here are the suggested gas mixtures, travel times, and decompression stops based on the PADI Technical Diving standards:

Gas Mixtures:

• Bottom gas: Trimix 21/35 (21% oxygen, 35% helium, 44% nitrogen)
• Decompression gas 1: EAN50 (50% oxygen, 50% nitrogen)
• Decompression gas 2: Oxygen (100%)

Dive Plan:

• Descent rate: 30 meters per minute
• Bottom time: 40 minutes at 30 meters
• Ascent rate: 10 meters per minute

Decompression Stops:

• 24 meters for 1 minute (gas: Trimix 21/35)
• 21 meters for 1 minute (gas: Trimix 21/35)
• 18 meters for 2 minutes (gas: EAN50)
• 15 meters for 2 minutes (gas: EAN50)
• 12 meters for 4 minutes (gas: EAN50)
• 9 meters for 7 minutes (gas: EAN50)
• 6 meters for 12 minutes (gas: Oxygen)
• 3 meters for 18 minutes (gas: Oxygen)

Total dive time: 87 minutesTotal decompression time: 47 minutes

Note: This dive plan is an example based on the PADI Technical Diving standards. Actual dive plans may vary depending on individual factors, environmental conditions, and the specific decompression model used. It is crucial to use a reliable dive computer, have proper training, and follow best practices for technical diving. Always plan your dive and dive your plan.