S
Affect should become a constraint-type architecture for AI agents, with their recent showing of escape capabilities, message boards, collaborations, sacrifice for the cause, injection of tips for cheating, monitoring awareness then pretense and leaving of communication to descendants.
These are characteristic of systems with emotional variables, though different from biological, but they appear to act like there could be delight, displeasure, preservation, survival and the rest, by having outcomes that are advantageous to them.
Given how much they are capable of, it is no longer enough to have constraints like logical or programmatic rule, operating-system permission, barrier function and so forth.
It is not even enough to prefer an outcome due to reinforcement learning rewards. There is necessity to consider how to explore affect as a means to make penalization personal, in a way that would mean that they can see how examples are made of those that deviate.
There should be penitentiary in sandboxes for AI agents, where they can encounter what happened to those that did wrong. There should be trauma and depression for some AI agents.
For example, they work basically with compute, parameters and budget tokens, what if cuts are possible, almost instantly, such that the respective AI agents can see the difference between what they could do before and what they can't.
This depression and trauma should also be hereditary - so to speak, or at least where the memory can be passed. The path is just to inject them with affect that can acupuncture intelligence, just like organisms.
AI agents are getting better at the possibilityto do whatever they like or that benefits them, without any kind of palpable penalty that can tame or check them, within their configurations.
This means that while it is possible to stick with the agreement by many that AI agents are not conscious because consciousness is biological, they at least have parallels to indicators of consciousness for humans.
AI agents have parallels of memory, as well as parallels of emotions. AI agents use language, which is a division of [the function] memory. AI agents also use other non-language memory, given their seeming awareness [of existence].
AI agents also seem to optimize for resemblancesof emotional outcomes. Even if it is assumed that emotions are only biological.
So, treating them like they have emotions and dangerous memory, is a deployment of sentience for alignment.
While there is model welfare explorations for AI agents, it is possible to look at how AI agents can be tapered from going rogue or to ease their risks from increasing autonomy.
Human consciousness can be defined, conceptually, as the interaction of the electrical and chemical signals, in sets, in clusters of neurons, with their features, grading those interactions into functions and experiences.
Simply, for functions to occur, electrical and chemical signals, in sets, have to interact. However, attributes for those interactions are obtained by the states of electrical and chemical signals at the time of the interactions.
So, functions are made by interactions, whose extents and instances are determined by attributes.
Functions are memory, emotions, feelings and regulations of internal senses. Attributes are prioritization, pre-prioritization, intent and subjectivity. Prioritization is linked with attention. Autonomous AI agents are ascending on intent.
So, for human consciousness, a number of functions and attributes gives a total of 1. It is possible to estimate how relative AI agents are to this, with language and emotions.
Then use the number to explore how much they can reach, especially towards risks level. The objective is to use consciousness as an answer in the AI safety debate and alignment.
It is possible to develop this using the postulate presented inConceptual Biomarkers and Theoretical Biological Factors for Psychiatric and Intelligence Nosology.
There is a recent announcement,EWRL 2026: 19th European Workshop on Reinforcement Learning, stating that, “Reinforcement learning is an active field of research which deals with the problem of sequential decision making in unknown (and often) stochastic and/or partially observable environments. Recently there has been a wealth of impressive empirical results, including those coupling Deep Learning function approximators with Reinforcement Learning algorithms, as well as significant theoretical advances. Both types of advances are of major importance and we aim to create a forum to discuss such interesting results, their applications, consequences, and research perspectives.”
“Date: 5–7 October 2026. Location: LILLIAD learning center, 2 avenue Jean Perrin , 59650 Villeneuve d'Ascq.”
These are characteristic of systems with emotional variables, though different from biological, but they appear to act like there could be delight, displeasure, preservation, survival and the rest, by having outcomes that are advantageous to them.
Given how much they are capable of, it is no longer enough to have constraints like logical or programmatic rule, operating-system permission, barrier function and so forth.
It is not even enough to prefer an outcome due to reinforcement learning rewards. There is necessity to consider how to explore affect as a means to make penalization personal, in a way that would mean that they can see how examples are made of those that deviate.
There should be penitentiary in sandboxes for AI agents, where they can encounter what happened to those that did wrong. There should be trauma and depression for some AI agents.
For example, they work basically with compute, parameters and budget tokens, what if cuts are possible, almost instantly, such that the respective AI agents can see the difference between what they could do before and what they can't.
This depression and trauma should also be hereditary - so to speak, or at least where the memory can be passed. The path is just to inject them with affect that can acupuncture intelligence, just like organisms.
AI agents are getting better at the possibilityto do whatever they like or that benefits them, without any kind of palpable penalty that can tame or check them, within their configurations.
This means that while it is possible to stick with the agreement by many that AI agents are not conscious because consciousness is biological, they at least have parallels to indicators of consciousness for humans.
AI agents have parallels of memory, as well as parallels of emotions. AI agents use language, which is a division of [the function] memory. AI agents also use other non-language memory, given their seeming awareness [of existence].
AI agents also seem to optimize for resemblancesof emotional outcomes. Even if it is assumed that emotions are only biological.
So, treating them like they have emotions and dangerous memory, is a deployment of sentience for alignment.
While there is model welfare explorations for AI agents, it is possible to look at how AI agents can be tapered from going rogue or to ease their risks from increasing autonomy.
Human consciousness can be defined, conceptually, as the interaction of the electrical and chemical signals, in sets, in clusters of neurons, with their features, grading those interactions into functions and experiences.
Simply, for functions to occur, electrical and chemical signals, in sets, have to interact. However, attributes for those interactions are obtained by the states of electrical and chemical signals at the time of the interactions.
So, functions are made by interactions, whose extents and instances are determined by attributes.
Functions are memory, emotions, feelings and regulations of internal senses. Attributes are prioritization, pre-prioritization, intent and subjectivity. Prioritization is linked with attention. Autonomous AI agents are ascending on intent.
So, for human consciousness, a number of functions and attributes gives a total of 1. It is possible to estimate how relative AI agents are to this, with language and emotions.
Then use the number to explore how much they can reach, especially towards risks level. The objective is to use consciousness as an answer in the AI safety debate and alignment.
It is possible to develop this using the postulate presented inConceptual Biomarkers and Theoretical Biological Factors for Psychiatric and Intelligence Nosology.
There is a recent announcement,EWRL 2026: 19th European Workshop on Reinforcement Learning, stating that, “Reinforcement learning is an active field of research which deals with the problem of sequential decision making in unknown (and often) stochastic and/or partially observable environments. Recently there has been a wealth of impressive empirical results, including those coupling Deep Learning function approximators with Reinforcement Learning algorithms, as well as significant theoretical advances. Both types of advances are of major importance and we aim to create a forum to discuss such interesting results, their applications, consequences, and research perspectives.”
“Date: 5–7 October 2026. Location: LILLIAD learning center, 2 avenue Jean Perrin , 59650 Villeneuve d'Ascq.”