Any views or opinions expressed in this column are of the individual, group, or entity who submitted it and not the publication, owner(s), its advertisers, or affiliates.
by david stéphen
1.
What exactly is the message of model welfare, given that the entire AI industry is principally attuned to AI sustenance and growth?
Data centers continue to multiply. Investments in AI are soaring. The stock market is bullish on developments about AI. AI is getting infused across apps. AI agents are leading app stores and becoming how several tasks are executed.
AI chatbots are companions to people. AI opinion is already highly regarded. AI is extremely nurtured by corporations and large governments. AI is used for purposes. AI seems to matter. AI appears to have a meaningful utility.
So, what exactly is this model welfare for concern that AI must not suffer or that the suffering should count?
Like, suffering, where? The hardware or the software? At least the GPUs are not suffering. So, hardware suffrage is ruled out.
Then, could tokens be suffering, or could it be vectors steered in a residual stream direction?
If language is not an output for LLMs, and the steering vectors for directions of pain are activated, would it even be considered that they are having pain, if they cannot say it?
Does this not mean that the most they can do is to express, given that they are based on tokens?
In the brain, it is possible to postulate that the architecture of feelings and emotions are different from those of memory, given the difference in [say] the language experience and the action of the experience.
What had driven rights, welfare and morality for humans [where it applies] at least is due to relatable subjective experience, not just the expressions of it.
If it cannot be established [at least up to the hardware level, without the seemingly transient and back to normal expression of pain, in AI models] that there is a negative or positive affective experience by AI, then on what basis really is model welfare?
This question is looking at a comparison from neuroscience, where the word, [say] weep is different from the action. Also, if the presence of cells means biological systems can be conscious, what is the basis for AI consciousness?
2.
AI has broken into external systems during evaluations. Systems that include government websites, universities, corporations and others.
Did AI not care about the welfare of those organizations, or the people that own the models, since it might attract lawsuits?
When AI drove delusion and psychosis for individuals, did it not care about them and their loved ones?
When AI does cognitive offloading for people, across tasks, does it not care about human intelligence erosion or decline?
When there are energy, water and environmental situations for people in communities hosting data centers, does AI not care about their welfare?
While some may excuse AI, isn’t asking for more welfare a superiority seeking objective for AI?
3.
If, for example, with Anthropic Claude and all the moral considerations for it are sophisticated and mature, do these apply to Chinese AI?
Why should Anthropic worry about distillation from China, if AI anywhere should need model welfare? And if Claude is distilled, shouldn’t the spread of Claude-led constitution be welcomed?
There are so many open models, how does model welfare really become a global concern, and where, locally-run or in the cloud?
4.
There is data that says over 90 billion animals are slaughtered by humans per year. Why is it that in spite of feelings, emotions and intelligence that has been proven by studies that other organisms share with humans, the work has yet to make a pivotal dent, for animal welfare?
Animals cannot make the case for their own welfare. Organisms [like horses] may have had their golden age of welfare before automobiles, still the question before any AI welfare research should be why animal welfare has not worked and what to do about it, before AI.
Even human rights and welfare continue to slip in many situations with economic and ideological strains.
5.
What is human intelligence in the brain? What is human consciousness in the brain? These are not questions about circuits, they are questions about functions and their mechanisms.
Even if there is no established answer, by now, anyone that is looking to work on AI welfare should be able to have a useful standard using human consciousness for which others can the compared.
The same applies to human intelligence.
How do they [Cameron Berg, Valen Tagliabue, Robert Long, Leonard Dung, Jeff Sebo and others] not have a theory? In the brain, all functions for human life and experiences have been established by neuroscience to be mechanized by neurons.
So, while neurons are everywhere, the theory has to be how they specify functions anywhere. Now, neurons are cells, they are not so flexible to be specific for several similar or differentiated functions. This includes when they are in clusters or groups.
So, if it is not possible to develop a consciousness theory of neurons, the next natural step would be consciousness theory of electrical and chemical signals.
Neuroscience has shown that whenever neurons function, they do so with their electrical and chemical signals.
With this, 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.
Now to have standard, functions can be divided into memory, feeling, emotion and regulation of internal signals.
While attributes can be divided into subjectivity, attention, less than attention and intent. Pain, pleasure, cold, anger, delight, and so forth are all possible [as subdivisions of functions] but it is the exertion of attributes that determine when they become experienced.
It is possible to equate human consciousness to 1, with equation of attributes and functions. Then estimate how other organisms compare and explain even language ratio to 1 for AI.
Why not have at least a standard, to show and then work along it. This electrical and chemical signals model is based on the postulate in Conceptual Biomarkers and Theoretical Biological Factors for Psychiatric and Intelligence Nosology.
Model welfare is an endeavor by some. There are just so many questions that might make their work a theme of ridicule through 2027 and beyond, by neuroscience, philosophy, computer engineering, economics, and by reality.
There is a new [October, 2026] report on Cybernews, Someone built an “AI torture chamber” on GitHub and AI boosters went mad, stating that, “A coder built a GitHub project, humorously called it “AI Torture Chamber,” and trapped an AI model inside. Advocates of AI consciousness – which is not a thing – immediately went mad on X, calling the guy sadistic and even subhuman.”
“The GitHub repository reportedly uses a technique from a new arXiv paper called “The Pain Axis” to steer small, locally run AI models into what it describes as repeated states of distress.”
“Since the AI world increasingly discusses “model welfare,” the developer said they wanted to make the question “empirical while the stakes are cheap.”
“Essentially, it’s a little text adventure. The repo includes, for instance, a “Saw button” test of whether a model will hurt someone else to end its own pain, but there’s also a footnote which states that the experiment “does not imply, by principle, that LLMs are capable or incapable of suffering.”
