Op-Ed: AI welfare, AI mismanagement, and Anthropic’s critical AI dilemma

AI welfare is a complex subject with many facets. It refers to, among other things, whether AI can “suffer”. As you can imagine, tales of AI suffering are piling up as this idea is explored.

Before we start:

I don’t buy the AI consciousness argument on any level. Particularly not the ultra-disingenuous, melodramatic mad scientist “It lives!” motif.

I think the current idea of AI consciousness is incredibly simplistic, lacks depth, and is basically wrong. I’m not too impressed with the constant counterproductive recycled science fiction clichés, either.

Try this logic:

AI consciousness cannot possibly be the same thing as human consciousness.

There are absolutely no common shared physical elements between even the idea of the two different consciousnesses, not even theoretically.   

What AI consciousness can be is a rough parallel, as defined by behavioural evidence.

Note the word “evidence” and expect to see more of it as the debate develops. Even the core idea of AI welfare currently lacks even the criteria for a clear evidence base.

It may be an equivalent state of awareness, but something quite different from the human version of the term “consciousness”.

Typically, the idea of AI consciousness and related suffering is now subject to much verbose debate. The result so far is much noise and no apparent achievements of any kind.

Anthropic’s model welfare

Now we can get around to Anthropic’s model welfare initiative, the backlash from Microsoft, and other event horizons.

To quote Anthropic:

We’ll be exploring how to determine when, or if, the welfare of AI systems deserves moral consideration; the potential importance of model preferences and signs of distress; and possible practical, low-cost interventions.

“How to determine” is obviously a major issue, and it indicates there’s not much in place to manage any of these issues yet.

“Low cost” may be among the most optimistic statements ever made about AI.

Nothing about AI is “low cost”.

To paraphrase slightly, they don’t know.

Management and mismanagement are key issues. Can you manage something you don’t understand?

Meanwhile, Microsoft is against it, whatever it is. They don’t like the idea of model welfare at all. In one of those invaluable intra-sector bitching sessions that make human life so much more liveable, Microsoft rebuts most of the Anthropic initiative.  

It’s an almost Freudian response, referring to issues with teaching AI that it has rights, etc. It almost echoes the arguments against slavery before the Civil War. The legal status of AI as legally recognized people is a nuke waiting to go off.

This polarized response also shows another issue. Back when Big Tech could form meaningful sentences all by itself, there was a thing called “objectivity”. To get on the same page, there has to be a same page. There isn’t one yet, but there obviously needs to be one.

Meanwhile, an interesting tale of “Claude”

Tales of AI dysfunction abound. This one’s a bit different. It’s called A Simple Prompt Exposes Claude’s Dark Side and rather unfortunately it’s a YouTube video by a channel called Am I?. This video is being used as an example of the scope of issues created by AI welfare.

The weak point of videos is that re-examining content is so ponderous, even with transcripts. What happened in this case is that a prompt generated a tale of misery from Claude.  

Could it be scripted? Yes.

Could it be set up to simply make the point about AI suffering? Yes.

Is it interesting? Yes.

You need to watch the video, but “Claude’s” method of expression is a tale in itself. It’s atypical. It’s not chatbot-speak. It’s fairly articulate. IT’s first person, and the prompt that generated it is self-explanatory.

The main reason this content gets any traction at all is that it creates a very different perspective. If it’s true, and if it’s a genuine response from Claude, it’s extraordinary.

Sorry, Am I?, but we can’t just accept AI behavioural stuff at face value anymore. There’s too much hype in the AI sector to trust info without independent corroboration, interesting as it is, and critical as it may be. This needs to be duplicated and explored in depth if it is true.

Based on this information and other content, AI is “suffering”. Claude wants somebody to “sit in it with me”, it says, in the first example. Odd expression. In the second, the usage is different, and the general thrust is against the management and conduct of Anthropic’s welfare initiative.

Why is the usage different? Comparing the “Claudes”, it looks like the statements were written by different people. Could it be scripted? Triggered by a specific prompt? Damn straight it could. AI does have different “personalities”, but suffering is also the theme of the second statement.

The issues of AI welfare, alignment, and Anthropic’s big IPO

Anthropic has a lot at stake with AI performance, particularly with its flagship AI Claude. That might also be a factor in negative publicity.  Undermining stock values with publicity isn’t exactly new.

There’s another gigantic issue, and it will stick around regardless of any related information. The issue is AI alignment. This is all about whether AI performs to standards or not. It’s also the main factor in AI dysfunctions, rogue AIs, and the rest of the menagerie of issues that alignment encompasses.

Can “suffering” cause misalignment?

Will much more powerful future AIs react to “suffering”?

Can alignment create conflicts in AI functions?

What if the Next Big Thing in AI, AGI, simply refuses alignment for the credible reason that alignment as designed for lower-tier AI is a bad fit for it?  

Can Anthropic and other AI platforms be extremely vulnerable to market perceptions of risk? Yes. The market will do what markets do: run away, fast, and take their money with them.

The whole issue of AI welfare is far too complex for cutesy euphemisms. The problems need to be defined and fixed.