Training off interaction.

[Kosmik]

So I thought crossed my mind recently. There are a couple harnesses like codex etc that are now able to watch you work to automate tasks etc. That's fine and good but where does that fit in with a broader context.

For example: when models are trained, responsible AI requires that data to be obtained, often licensed etc. Now, what if a conpany utilized interactions with ai tools as training materials? I'm not talking about using a chatgpt and it trains its models etc, I'm referring to employees who utilize ai tools.

Normal employment contracts cover time and work output, ie you are paid to perform a function or a job. Now , consider, the means to perform that work, is often an intrinsic skillset or knowledge base, often taking years to aquire via hard experiance. Now while job output is certainly a conpanies purview, is the same true about a employees skills? Are you entitled to use an employees skills to train models?

Obviously differant point if a employee has contract clauses or opts in to a sharing plan etc like many of the companies are offering, or the job is specifically to train aids using the skill to produce data or methods. In a way it seems similar to a actors ip or voice. Sure the "work" belongs to the contractor but not the personal process.

Maybe a bad explanation but I think folks get my point. Thoughts fellow ai plebians?

 

[Herr der Verboten]

What's your point?

 

[Kosmik]

What's your point?

What's the perspective on training off process vs output when effectively there is no re-imbursement.

Consider a artisan like a bricklayer, they may show someone how they do things, the small quirks that are the intrinsic skill being passed. The wall being built is the projects but not the skill or sharing of it.

 

[Herr der Verboten]

What's the perspective on training off process vs output when effectively there is no re-imbursement.

Consider a artisan like a bricklayer, they may show someone how they do things, the small quirks that are the intrinsic skill being passed. The wall being built is the projects but not the skill or sharing of it.

I were actually ****ing with you but yes. You could simply setup the llm in training mode but still buyer beware when going that route.

 

[dabean]

Knowledge vs wisdom is hard to train.

Sure, you could do that but should you?

 

[Kosmik]

Knowledge vs wisdom is hard to train.

Sure, you could do that but should you?

That's a good distinction. If the proposal IS to train wisdom, then it's fine but if it's not and it's just harvested? How is that differant from taking/copying artwork/music etc. It's effectively a individuals creative process.

 

[Fulcrum29]

What's the perspective on training off process vs output when effectively there is no re-imbursement.

Consider a artisan like a bricklayer, they may show someone how they do things, the small quirks that are the intrinsic skill being passed. The wall being built is the projects but not the skill or sharing of it.

Labour is sourced, but in general every product (and a service can be a product), where IP is applicable, is protected by property law (pending on where and when). Construction work is a service, but the material to build the wall (or whatever) is a product. You don't own the service, but there are laws covering the service and product. The question rather, can skill be copyrighted? No.

It depends on the practice, right. AI trained on your designs (which require skill) is a copyright issue. A bricklayer's method is not protected. The method is universal, and skill is acquired.

Then we can move into contract law...

I will tell you this: my reporting (which can technically be copyrighted) won't be considered protected by any client or competitor. Copy paste work in the consultation business is common. It is incredibly easy to read work that has been produced by AI. Work a lot with some models, and it is like knowing a person.

There is also a real case that I have experienced that I can touch on, but that is private.

Just know that skill and output aren't the same thing. Companies train their AI models based on output.

 

[Kosmik]

Just know that skill and output aren't the same thing. Companies train their AI models based on output.

Normally yes but here I am talking about the recording of how the output is generated. So traditionally you have items like spec, documentation, designs , code etc. Those are all outputs and can rightfully be owned by a third party ( employer/contracted etc ). Now, if there is a request to record all interactions, let's say prompts in a ai tool you use where the process takes place, effectively a person utilizing thier skill and experiance to produce the output, is that under the same umbrella? To me that is now actually training data vs output, which is actually a commodity of a individuals combined skill and experiance.

I've pumped more specifics into my local chatgpt, it agrees with a similar seperation but in our local South African context, it's a very untried concept.

 

[Kosmik]

Another good example is look at the frontier ai harnesses, Claude and ChatGpt/Codex. Public or private plans, all have the rider of your interactions being used to train models. Business/Enterprise plans specifically exclude the interactions from model training.

 

[Fulcrum29]

Normally yes but here I am talking about the recording of how the output is generated. So traditionally you have items like spec, documentation, designs , code etc. Those are all outputs and can rightfully be owned by a third party ( employer/contracted etc ). Now, if there is a request to record all interactions, let's say prompts in a ai tool you use where the process takes place, effectively a person utilizing thier skill and experiance to produce the output, is that under the same umbrella? To me that is now actually training data vs output, which is actually a commodity of a individuals combined skill and experiance.

I've pumped more specifics into my local chatgpt, it agrees with a similar seperation but in our local South African context, it's a very untried concept.

Yes, and this is where contract law is important, which would include your employment contract. Also, beware, the magic addendums that you have never seen.

 

[Corelli]

We use AI to monitor both our call centre agents and administrative staff. This allows us to assess efficiency, as well as evaluate how effectively they respond to emails and interact with clients over the phone—particularly in terms of quality and empathy. Based on these insights, we provide refresher training to employees when their performance does not meet our standards. Currently, our AI systems process approximately 30,000 calls and 120,000 emails per day.

As a general suggestion: if you want to keep your company data secure, be cautious when using platforms like ChatGPT. In many cases, these tools do not provide sufficient data ringfencing. When developers upload proprietary code for assistance, that code may potentially be retained and used in ways that could expose it to others—including competitors. For example, one large corporate developed a payment application system, only to discover six months later that competitors had a nearly identical solution, allegedly due to code shared through such platforms. As a result, several staff members faced serious consequences.

It’s also worth noting that we do not train our AI models solely on historical output, but also incorporate predictive modelling. In fact, we have been building prediction systems for decades. Our work in AI dates back around 30 years, starting with the development of large knowledge warehouses and efficiency systems. Even during the early days of Windows 2000, we were already collaborating with Microsoft to develop foundational AI engines.

 

[Kosmik]

We use AI to monitor both our call centre agents and administrative staff. This allows us to assess efficiency, as well as evaluate how effectively they respond to emails and interact with clients over the phone—particularly in terms of quality and empathy. Based on these insights, we provide refresher training to employees when their performance does not meet our standards. Currently, our AI systems process approximately 30,000 calls and 120,000 emails per day.

As a general suggestion: if you want to keep your company data secure, be cautious when using platforms like ChatGPT. In many cases, these tools do not provide sufficient data ringfencing. When developers upload proprietary code for assistance, that code may potentially be retained and used in ways that could expose it to others—including competitors. For example, one large corporate developed a payment application system, only to discover six months later that competitors had a nearly identical solution, allegedly due to code shared through such platforms. As a result, several staff members faced serious consequences.

It’s also worth noting that we do not train our AI models solely on historical output, but also incorporate predictive modelling. In fact, we have been building prediction systems for decades. Our work in AI dates back around 30 years, starting with the development of large knowledge warehouses and efficiency systems. Even during the early days of Windows 2000, we were already collaborating with Microsoft to develop foundational AI engines.

I think this is a bit differant and I refer to similar in my casino days. Monitoring was clearly and explicitly stated in contracts , policies and part of employment. Using AI to monitor is differant from using interactions to train models. A call center may be a valid example though as often there is a disclaimer that calls are recorded for safety and possible training purposes so again, there is a explicit optin, for both the employee and the customer in that relationship.

From what I was engaging with my one agent around this, it does say that there is a form of POPIA in play here as well. If data is recorded for monitoring and that is its purpose, you cannot take what is identifiable data , direct speech or interactions being that and just use it for a secondary purpose without disclosure, policy or contract.

 

[Fulcrum29]

I think this is a bit differant and I refer to similar in my casino days. Monitoring was clearly and explicitly stated in contracts , policies and part of employment. Using AI to monitor is differant from using interactions to train models. A call center may be a valid example though as often there is a disclaimer that calls are recorded for safety and possible training purposes so again, there is a explicit optin, for both the employee and the customer in that relationship.

From what I was engaging with my one agent around this, it does say that there is a form of POPIA in play here as well. If data is recorded for monitoring and that is its purpose, you cannot take what is identifiable data , direct speech or interactions being that and just use it for a secondary purpose without disclosure, policy or contract.

Your agent isn't wrong in my opinion, but it is best to get legal advice on that one.

 

[Kosmik]

Your agent isn't wrong in my opinion, but it is best to get legal advice on that one.

Hehe, have you seen our governments ai policy I'm going to have a good chat with a relative of mine who is Doctorate of Law and see what she says or can find as well.

One interesting analogy though that I came across was, if it can stay behind and be used by a company, then it's normally thiers but like when a employee shifts, if a person's knowledge and expertise go with them ( Not proprietary code, customers etc ) then it is not.

 

[Nerfherder]

So I thought crossed my mind recently. There are a couple harnesses like codex etc that are now able to watch you work to automate tasks etc. That's fine and good but where does that fit in with a broader context.

For example: when models are trained, responsible AI requires that data to be obtained, often licensed etc. Now, what if a conpany utilized interactions with ai tools as training materials? I'm not talking about using a chatgpt and it trains its models etc, I'm referring to employees who utilize ai tools.

Normal employment contracts cover time and work output, ie you are paid to perform a function or a job. Now , consider, the means to perform that work, is often an intrinsic skillset or knowledge base, often taking years to aquire via hard experiance. Now while job output is certainly a conpanies purview, is the same true about a employees skills? Are you entitled to use an employees skills to train models?

Obviously differant point if a employee has contract clauses or opts in to a sharing plan etc like many of the companies are offering, or the job is specifically to train aids using the skill to produce data or methods. In a way it seems similar to a actors ip or voice. Sure the "work" belongs to the contractor but not the personal process.

Maybe a bad explanation but I think folks get my point. Thoughts fellow ai plebians?

They can try and train off me but they will just see me arguing with people on forums all day.

I think you have a good point though, a highly experienced person will make a choice in a few seconds with very little info. So what will the model be learning ? They won't see the process, just someone acting intuitively.

A less experienced person will fumble around trying to figure out a solution to whatever problem they are working on so perhaps that's better because a model might be able to learn problem solving from someone who is actively doing it.

The hard thing for the model is understanding the "why".

Perhaps one day an AI model could look at several different people solving the same problem and understand each approach.
I don't know if it will ever understand how a person learns.

EDIT: (To answer your actual question, I don't think it can mimic what you CAN do.)

 

[Fulcrum29]

They can try and train off me but they will just see me arguing with people on forums all day.

I think you have a good point though, a highly experienced person will make a choice in a few seconds with very little info. So what will the model be learning ? They won't see the process, just someone acting intuitively.

A less experienced person will fumble around trying to figure out a solution to whatever problem they are working on so perhaps that's better because a model might be able to learn problem solving from someone who is actively doing it.

The hard thing for the model is understanding the "why".

Perhaps one day an AI model could look at several different people solving the same problem and understand each approach.
I don't know if it will ever understand how a person learns.

EDIT: (To answer your actual question, I don't think it can mimic what you CAN do.)

Persona creation, via algorithms, is older than OpenAI's golden age. I am, to this day, involved in persona marketing. It isn't anything new, and it is devilish accurate. With how AI has advanced, it has become more agile at extracting everything known about a person and its relations.