1
Opinion
Have you ever said "thanks" to an AI?
alamy.com
Chances are you probably have, we tend to anthropomorphise our AI chatbots or attribute human qualities to them. It feels natural because the chatbot responds like a human and even remembers past conversations.
Naturally people also fear AI's, they appear conscious and by virtue of their intelligence may want to take over the world.
But this is where we must stop and question the word want, because at least with current technology large language models or LLM's don't have a proper concept of wanting something.
LLM's are pre-trained on large data sets and essentially take a string of text as input (represented as tokens) and produce another string of text (tokens) as the output.
What comes back, the reply, is largely dependent on what went in and to some degree the model itself and how it was trained.
Where this becomes a bit fuzzy and confusing is that most people don't see the entire input string.
Your actual prompt might be simple, but it is most likely prefixed by a large amount of context. That context may contain details about you, about earlier concersations, as well as instructions on how to behave and the personality traits it should display.
It is precisely this contextual attachment that makes a chatbot sound like a human and as someone who cares.
The inherent danger associated with AI is therefore not so much the AI itself, but the human operator who provided the context and the instructions to the AI.
We should therefore not think of an AI as being a giant monster living in a data center, but instead think of it as millions of agents acting for humans.
Unfortunately this doesn't make it safe, because operators with bad intentions can and do use AI to achieve their objectives.
Recent advances allowing users to spawn multiple sub agents effectively lets anyone instantly create a whole army of bots, potentially with malicious instructions, but ultimately someone instructed them.
What we should really be concerned about, is the lack of accountability. Humans have a history of behaving badly when responsibility cannot be attributed to them.
What if AI agents were required to be registered and operate under unique IDs? Each ID could be linked to a verified human or legal entity. That person or corporation could then be held legally responsible for the agent’s actions.
Leading LLM providers could refuse connecting with unregistered bots, but this still leaves the door open for anyone capable of running their own LLM.
Which leads us to a much larger problem, the current IP system. An IP address, which is similar to a phone number - only identifies a network, it does not identify the legal entity, making it difficult to prosecute an individual person.
The debate over individual identification has raged since the internet began. Until now the world has tolerated a system of anarky over a rules based system.
But with the arrival of LLMs and autonomous agents, allowing unidentified operators to deploy thousands, or millions of intelligent actors may simply be too dangerous.
Are we ready to take responsibility for what our AI agents do?
Naturally people also fear AI's, they appear conscious and by virtue of their intelligence may want to take over the world.
But this is where we must stop and question the word want, because at least with current technology large language models or LLM's don't have a proper concept of wanting something.
LLM's are pre-trained on large data sets and essentially take a string of text as input (represented as tokens) and produce another string of text (tokens) as the output.
What comes back, the reply, is largely dependent on what went in and to some degree the model itself and how it was trained.
Where this becomes a bit fuzzy and confusing is that most people don't see the entire input string.
Your actual prompt might be simple, but it is most likely prefixed by a large amount of context. That context may contain details about you, about earlier concersations, as well as instructions on how to behave and the personality traits it should display.
It is precisely this contextual attachment that makes a chatbot sound like a human and as someone who cares.
The inherent danger associated with AI is therefore not so much the AI itself, but the human operator who provided the context and the instructions to the AI.
We should therefore not think of an AI as being a giant monster living in a data center, but instead think of it as millions of agents acting for humans.
Unfortunately this doesn't make it safe, because operators with bad intentions can and do use AI to achieve their objectives.
Recent advances allowing users to spawn multiple sub agents effectively lets anyone instantly create a whole army of bots, potentially with malicious instructions, but ultimately someone instructed them.
What we should really be concerned about, is the lack of accountability. Humans have a history of behaving badly when responsibility cannot be attributed to them.
What if AI agents were required to be registered and operate under unique IDs? Each ID could be linked to a verified human or legal entity. That person or corporation could then be held legally responsible for the agent’s actions.
Leading LLM providers could refuse connecting with unregistered bots, but this still leaves the door open for anyone capable of running their own LLM.
Which leads us to a much larger problem, the current IP system. An IP address, which is similar to a phone number - only identifies a network, it does not identify the legal entity, making it difficult to prosecute an individual person.
The debate over individual identification has raged since the internet began. Until now the world has tolerated a system of anarky over a rules based system.
But with the arrival of LLMs and autonomous agents, allowing unidentified operators to deploy thousands, or millions of intelligent actors may simply be too dangerous.
Are we ready to take responsibility for what our AI agents do?





