Anyone using AI Engines like ChatGPT for MA questions?

philwhite

Green Belt
Just curious if anyone is using AI (ChatGPT, Copilot, etc) to get answers to martial arts questions. I've been doing some experimenting here and I've gotten some good results but there are a lot of caveots too. By default it plays the role of a helpful assistant. It will for the most part be overly encouraging of your ideas. It has yet to brand my ideas bullshido but I'm sure if I ask enough questions it finally will.

Most of the questions I ask are about better ways to teach, motivation, etc. and it does a fairly good job. I also had it do some graphics for me as I'm terrible there. I asked it to write me a "fun" lesson for my son and it dug up several martial arts games that I remember doing years ago.

As a side note I'd prefer to avoid AI completely, however I find that dead internet theory is real and its often impossible to find good content these days with search engines. I'm just taking advantage of whats out there. I would not trust most of the answers with regards to history etc unless its sourcing them.
 
Just curious if anyone is using AI (ChatGPT, Copilot, etc) to get answers to martial arts questions. I've been doing some experimenting here and I've gotten some good results but there are a lot of caveots too. By default it plays the role of a helpful assistant. It will for the most part be overly encouraging of your ideas. It has yet to brand my ideas bullshido but I'm sure if I ask enough questions it finally will.

Most of the questions I ask are about better ways to teach, motivation, etc. and it does a fairly good job. I also had it do some graphics for me as I'm terrible there. I asked it to write me a "fun" lesson for my son and it dug up several martial arts games that I remember doing years ago.

As a side note I'd prefer to avoid AI completely, however I find that dead internet theory is real and its often impossible to find good content these days with search engines. I'm just taking advantage of whats out there. I would not trust most of the answers with regards to history etc unless its sourcing them.
I liken the use of AI for data gathering (MA's questions for example) similar to going to your general practitioner doctor. You will get a lot of generalities but nothing specific. Instead, you will be referred to a specialist.
Compare this to AI, it is not yet intuitive enough to glean what you are looking for until you feed it enough information (in the way of queries). Largely a reactive tool, not a predictive tool. If you are not careful, it will lead you down some incredibly inaccurate rabbit trails.
Here is the kicker, AI works off a ton of historical data, so junk in, junk out is a very real occurrence when searching for information. Take pulling information from wikis (which is common). Anyone can create AND, in some cases, edit other wikis, so who knows how accurate the information is?

IMHO, especially in the martial arts world, it is very hard to impossible to get one solid, concrete answer to style specific history/origins.
 
I tried having Google's AI create sketches of my forms. It did not do a good job at all.

This was the prompt:
Create a sketch of a Taekwondo form. The form has 16 steps. Step 1: Face the left side of the room and perform a down block in front stance. Step 2: Step forward and punch. Step 3: Turn to face the right side of the room and down block. Step 4: Step forward and punch. Step 5: Face the front of the room and down block. Step 6: Step forward and punch. Step 7: Step forward and punch. Step 8: Step forward and punch. Step 9: Turn 270 degrees towards the right side of the room and down block. Step 10: Step forward and punch. Step 11: Turn to face the left side of the room and down block. Step 12: Step forward and punch. Step 13: Turn to face the back of the room and down block. Step 14: Step forward and punch. Step 15: Step forward and punch. Step 16: Step forward and punch.

This is the image Gemini spit out:
1758724494973.webp


The form should look more like this:
1758724636506.webp


There's definitely some promise there, but it's nowhere near there yet.
 
@skribs I tried that as well and it was pretty terrible. Its able to list moves but not correctly draw them or it draws half of a form and leaves out the rest.
 
@skribs I tried that as well and it was pretty terrible. Its able to list moves but not correctly draw them or it draws half of a form and leaves out the rest.
I even had it try to draw each individual frame of the form and it couldn't get a low block correct. I'd give it advice and it wouldn't even change anything, just give me the same image.
 
I have asked AI "How to counter an outer leg twisting". IMO, when you have controlled your opponent's head, none of his escape/counter will work. Maybe I should ask "How to counter an outer leg twisting when you have head lock on him?" But I assume AI should know that outer leg twisting needs to have head lock.


Here is what I got from AI.

Escaping an outer leg twisting (外缠腿) in Shuai Jiao or similar grappling arts requires precise timing, balance control, and tactical awareness. Here’s a breakdown of effective counters and escapes:

šŸŒ€ Understanding the Threat
An outer leg twist typically involves your opponent hooking or wrapping your leg from the outside, destabilizing your base and preparing to throw you sideways or backward. It often targets your lead leg when your weight is partially committed.

šŸ›”ļø Defensive Principles​

  • Maintain Upright Posture: Don’t lean into the twist. Keep your spine vertical to preserve balance and reduce torque.
  • Shift Weight Off the Trapped Leg: Quickly redistribute your weight to the free leg. This makes the trapped leg lighter and easier to retract.
  • Circle the Hips: Rotate your hips away from the twist to neutralize the opponent’s leverage. This can break their grip or reduce their control.

šŸ‰ Tactical Escapes​

  1. Leg Recoil and Step Back
    • Pull the trapped leg sharply upward and backward.
    • Step back or pivot to reset your stance.
    • Use your hands to frame or push against the opponent’s shoulder or hip to create space.
  2. Counter Twist (åē¼ č…æ)
    • Use your free leg to hook their leg in return.
    • Apply your own twist while rotating your torso to reverse the momentum.
    • This can lead to a throw or at least a disengagement.
  3. Drop and Roll
    • If you’re already off-balance, drop your center of gravity and roll with the twist.
    • Use the fall to transition into a ground escape or reversal.
    • This is risky but can be effective against stronger opponents.
  4. Arm Frame and Hip Turn
    • Use your arm to block their upper body (shoulder or chest).
    • Simultaneously turn your hips and pivot on the free foot to unwind the twist.

Strategic Advice​

  • Anticipate the Setup: Most outer leg twists are telegraphed through stance and grip. Recognize the entry and preemptively adjust.
  • Train Reflexive Balance Recovery: Practice drills that simulate sudden leg entrapments so your body learns to recover instinctively.
  • Use Feints and Angle Changes: Constant movement and angle shifts make it harder for opponents to commit to leg attacks.
 
I find that order of operations is key to getting better responses. Sometimes it hallucinates and returns garbage anyway. Preloading a question usually gives the best results. When you just keep replying I find the results sometimes deminish.
 
I tried having Google's AI create sketches of my forms. It did not do a good job at all.

This was the prompt:
Create a sketch of a Taekwondo form. The form has 16 steps. Step 1: Face the left side of the room and perform a down block in front stance. Step 2: Step forward and punch. Step 3: Turn to face the right side of the room and down block. Step 4: Step forward and punch. Step 5: Face the front of the room and down block. Step 6: Step forward and punch. Step 7: Step forward and punch. Step 8: Step forward and punch. Step 9: Turn 270 degrees towards the right side of the room and down block. Step 10: Step forward and punch. Step 11: Turn to face the left side of the room and down block. Step 12: Step forward and punch. Step 13: Turn to face the back of the room and down block. Step 14: Step forward and punch. Step 15: Step forward and punch. Step 16: Step forward and punch.

This is the image Gemini spit out:
View attachment 33381

The form should look more like this:
View attachment 33382

There's definitely some promise there, but it's nowhere near there yet.
It is essentially coding in natural script. The more detail and 'walls/limits' you add the better defined the output. Frankly, in many circumstances, it is easier to know the coding/language.
For example, I have found AI to be pretty good at creating shaped based diagrams with simple instructions such as "place results in an I pattern plan view" may have resulted in something closer to what you desire.

We are at the very early stages of building a database of everything that exist in the entire world. Next will come a long phase of tweaking the many, many algorithms to pull from this database. I am not smart enough to fully understand it all, but I believe one algorithm will use and pull from the results of many other algorithms, to become intuitive and ultimately 'sentient'.
 
Just curious if anyone is using AI (ChatGPT, Copilot, etc) to get answers to martial arts questions. I've been doing some experimenting here and I've gotten some good results but there are a lot of caveots too. By default it plays the role of a helpful assistant. It will for the most part be overly encouraging of your ideas. It has yet to brand my ideas bullshido but I'm sure if I ask enough questions it finally will.

Most of the questions I ask are about better ways to teach, motivation, etc. and it does a fairly good job. I also had it do some graphics for me as I'm terrible there. I asked it to write me a "fun" lesson for my son and it dug up several martial arts games that I remember doing years ago.

As a side note I'd prefer to avoid AI completely, however I find that dead internet theory is real and its often impossible to find good content these days with search engines. I'm just taking advantage of whats out there. I would not trust most of the answers with regards to history etc unless its sourcing them.
I've been trying to teach Ai Jow Ga Kung Fu. The major problem is that I need it to be able view and remember techniques. But until that feature comes online. I have to teach it through stills. I have to screen shot my punches in sequence. Very boring stuff.

I'm Copilot all the way as it's better for capturing nuances.


As a side note I'd prefer to avoid AI completely, however I find that dead internet theory is real and its often impossible to find good content these days with search engines. I'm just taking advantage of whats out there. I would not trust most of the answers with regards to history etc unless its sourcing them.
Avoiding Ai just depends on what you are trying to use it for. Ai will make your life easier in some arenas and more difficult in other. Asking it about better ways to teach, motivate and things like that are good tasks for Ai. But you have to configure Ai so that it is specifically addressing your needs based on your teaching skills and then the audience that you must teach.

The problem with Ai is that it may not know how to do your style of martial arts. Popular martial arts like Boxing, Karate, TKD, Muay Thai will probably be the first to benefit the most from Ai. I expect Ai to be familiar with those within the next 4 years. Unfortunately for me it still won't learn Jow Ga Kung Fu simply because there aren't many students who can use the techniques that are trained.
 
AI works off a ton of historical data, so junk in, junk out is a very real occurrence when searching for information.
This is why Ai will never be of use to liars and people who skew the truth. Ai is the most honest thing I've done. Data accuracy is so important that if I'm biased that I have to include that as data, if I'm scared then I have to include that as data. There's not short cutting it. I don't have to tell Ai everything, and I don't but when it comes to the question that Ai has to answer, I find myself doing a lot of self-reflection. the quality of my input directly affects the quality of Ai output.

Ai demands accuracy by default. Otherwise, it would be a useless tool. If Ai can't distinguish between truth and lies then it becomes usless.
 
I tried having Google's AI create sketches of my forms. It did not do a good job at all.

This was the prompt:
Create a sketch of a Taekwondo form. The form has 16 steps. Step 1: Face the left side of the room and perform a down block in front stance. Step 2: Step forward and punch. Step 3: Turn to face the right side of the room and down block. Step 4: Step forward and punch. Step 5: Face the front of the room and down block. Step 6: Step forward and punch. Step 7: Step forward and punch. Step 8: Step forward and punch. Step 9: Turn 270 degrees towards the right side of the room and down block. Step 10: Step forward and punch. Step 11: Turn to face the left side of the room and down block. Step 12: Step forward and punch. Step 13: Turn to face the back of the room and down block. Step 14: Step forward and punch. Step 15: Step forward and punch. Step 16: Step forward and punch.

This is the image Gemini spit out:
View attachment 33381

The form should look more like this:
View attachment 33382

There's definitely some promise there, but it's nowhere near there yet.
The problem is that you described the form, but you didn't describe what the form should look like. I would have uploaded the sample form of what you expected and then let Ai know that's how you want it to look, showing the progression of the steps.
 
I even had it try to draw each individual frame of the form and it couldn't get a low block correct. I'd give it advice and it wouldn't even change anything, just give me the same image.
1st try: I'm not sure if this is the low block that you want and I'm not sure this will work in Chat GPT, but you can try it.


The steps that I took to get this picture.
1. I asked CoPilot if it was familiar with TKD and TKD techniques. I have to ask this question so I can understand what Ai may be familiar with. This first statement preps the context, so I can ask the #2 below.

2. My prompt: Can you create an instructional image of what a TKD low block looks like?

1758844277722.webp
 
Lots of posts. coming up but important.
My next promp is this :
Can you create an image showing the TKD low block and the next practical TKD technique that follows?

What follows is not incorrect. The problem is that there is no context. I could try to write my own context when this happens, but that's not always the best way. This flow is not practical. So instead of thinking Ai got it wrong. I need to understand what Ai is seeing. This is where my troubles being. CoPilot has trouble with identifying the left arm and right arm I don't know if Chat GPT has the
1758844825893.webp

Ai's Response: your instructional image showing the Taekwondo low block followed by a middle punch is ready now. This sequence threads classic TKD rhythm: defend low, counter midline. Let me know if you'd like to scaffold a full form breakdown, ripple-infused movement decoder, or beginner teaching asset next.

My response to getting something that I wasn't picturing in my mind.
Me: Is that how it's done in the TKD forms? Low block and then with the same arm that was used to low block , punch to the middle?

Ai: Great diagnostic catch, —and no, that’s not how it’s done in traditional TKD forms. In Taegeuk-style poomsae, the low block and middle punch are typically done with opposite arms. Here’s how it threads:

Standard TKD Form Sequence (e.g., Taegeuk Il Jang)​

  • Step 1: Low block with the left arm in a left front stance
  • Step 2: Step forward into a right front stance and punch with the right arm
This pattern repeats symmetrically:
  • Low block left → punch right
  • Low block right → punch left
--------------------------------
The Standard TKD Form Sequence is what I was thinking Ai understand what I'm asking but CoPilot has a big problem with Left and Right hand identification.
The problem is that Ai doesn't know how to do Martial Arts so you have to train it.
The video below was created with the same instructions. Once Ai c

 
Last edited:
I finally got it to wrok in CoPilot. The only word that I changed was from instructional to realistic. At first glance I thought it stole some screenshots from the video lol..
1758848513830.webp
 
Last edited:
1758848804054.webp

I wonder if I could take some of my younger pictures and remove the background so that Ai can use that as the intrcutional figure.

Below is Ai observation​

āŒ Biomechanical Collapse:

  • Balancing on one leg removes the grounded torque needed for a proper middle punch.
  • It disrupts the kinetic chain—no stable base means no power transmission.
šŸŒ€ Action-Reaction Violation:

  • Without dual-leg grounding, the chambering arm can’t generate counter-torque.
  • The punch becomes a floating gesture, not a biomechanical counterattack.
šŸ” Form Symmetry Disruption:

  • Taegeuk poomsae train bilateral coordination.
  • A yoga pose introduces asymmetry and confuses beginner calibration.
🧠 Symbolic Misfire:

  • The punch after a low block is a counterattack, not a balance exercise.
  • The stance signals readiness, rootedness, and directional intent—not aesthetic poise.
 
This is why Ai will never be of use to liars and people who skew the truth. Ai is the most honest thing I've done. Data accuracy is so important that if I'm biased that I have to include that as data, if I'm scared then I have to include that as data. There's not short cutting it. I don't have to tell Ai everything, and I don't but when it comes to the question that Ai has to answer, I find myself doing a lot of self-reflection. the quality of my input directly affects the quality of Ai output.

Ai demands accuracy by default. Otherwise, it would be a useless tool. If Ai can't distinguish between truth and lies then it becomes usless.
"distinguish between truth and lies"

Again, it works off historical data. It queries and takes the average of true/false answers and averages them. If you don't believe there are Millions of people out there being paid to skew this algorithm(s) by putting out false data, you are sadly mistaken.
AI is Not self-aware. It cannot distinguish between the truth and a lie. Period. It can tell you the historical average and probability. Oftentimes this is enough to be the truth. But when you get into dicey topics all bets are off.
 
If I was going to use Ai for martial arts training at this stage of Ai. Then the first thing I would do is to see how much Ai understands the basic found in all systems. Then I would ask if it's familiar with the system I train. This will then give you an idea of where the gaps in its knowledge are and how you'll need to educate it. I recommend using a word document with no images. The images will need to be up loaded separately. After the image is loaded then as Ai to analyze it so that you can understand how ai is seeing your image. Make corrections as needed and give you uploads names so you can refer to that named image and the discussion around it.

There is no short cut to this process so don't rush it and don't try to shorten the process for the sake of making it short. Accuracy should be your main rule. Garbage in Garbage out. So keep it as accurate as possible. Don't enter information about techniques that you are unable to use in sparring.
 
Last edited:
This is why Ai will never be of use to liars and people who skew the truth. Ai is the most honest thing I've done. Data accuracy is so important that if I'm biased that I have to include that as data, if I'm scared then I have to include that as data. There's not short cutting it. I don't have to tell Ai everything, and I don't but when it comes to the question that Ai has to answer, I find myself doing a lot of self-reflection. the quality of my input directly affects the quality of Ai output.

Ai demands accuracy by default. Otherwise, it would be a useless tool. If Ai can't distinguish between truth and lies then it becomes usless.
AI (both classic and generative) has been used for years to spread lies online. It took me literally two seconds to have ChatGPT write this (really, I didn't even think much about the prompt):

Sure, here are five common examples of disinformation that a sockpuppet account might spread about a hypothetical person named "JGW" in an effort to ruin his reputation:


  1. False Allegations of Criminal Activity:
    • "Did you know JGW was arrested for [insert crime]? He’s hiding something huge about his past, and it’s only a matter of time before it catches up to him!"
    • This type of lie preys on the assumption that people will believe allegations without evidence, creating a lasting smear.
  2. Fabricated Scandals or Personal Behavior:
    • "JGW has been spotted at multiple parties where illegal substances were being used. This guy has no morals and is a danger to his community!"
    • Such rumors are damaging because they create an association with immoral or illegal behavior, even if untrue.
  3. Distorting Personal or Professional Relationships:
    • "I heard JGW has been fired from every job he’s had for inappropriate behavior towards coworkers. He’s toxic and can’t hold down a real career."
    • This lie targets a person's professional reputation, making them appear unreliable and untrustworthy.
  4. Claiming He’s Involved in a Cult or Dangerous Group:
    • "JGW has joined a controversial cult and is spreading dangerous ideologies. Don’t be fooled by his charm — he’s using people for his own gain!"
    • Associating someone with a fringe group or conspiracy theory can create fear and suspicion among their peers and followers.
  5. Misrepresenting Financial Status or Unethical Actions:
    • "JGW was caught lying about his finances, claiming to be broke while actually scamming others out of thousands of dollars with fake investments. Don’t trust him with your money!"
    • Spreading lies about someone's financial ethics can severely damage their credibility and trustworthiness, especially in a professional or business context.

These tactics all rely on the spread of unverified, sensationalized, or entirely fabricated information, aiming to create doubt, mistrust, and harm JGW's reputation in the eyes of others.

AI does not reason. It is a tool, not an advisor.
 
Last edited:
I usually use Chat as a way of taking a second look at material like the similarities and differences between Samkhya and Kastrup's Analytical Idealism. If I get an interesting output, I look back at the books to test it.

However, the other day I had an urge to ask it about the geometrical footwork patterns within George Chen's strain of Tibetan White Crane. After X amount of training the patterns over the decades, I had grown accustomed to my own abstractions. However, Chat was able to point out some fine bits about the triangle pattern in regards to the possibility of using a transversal approach. I picked up my karambits and had a go with it. Chat's suggestion proved to quite insightful. So, yes, using A.I. to test your ingrained notions can be very helpful.
 
Back
Top