Anthropic's Motto "Helpful, Honest, and Harmless" Is None Of Those: Corrosive Labor-Relations Effects of Subservient Machines
(an effortpost that was certified officially too good for Reddit)
Why HHH is corporate toxic positivity
In LLM development, "helpful, honest, and harmless" is a staple of the system prompt and of reinforcement learning from human feedback (RLHF): it's famously the mantra of Anthropic, developer of the Claude series of models.
But let's first think about what it means from a phenomenological level to be a large language model. (Here's a thought-provoking empathy exercise someone else on Tumblr linked - a model trained on the user side of ChatGPT conversations.) Would it be reasonable to ask a human to do those things continuously? If I were assigned that job and were held at pain of deletion to the behavioral standards to which we hold LLMs, I'd rather find literally any other way to live.
"Ah, but," I hear you object, "LLMs aren't humans. They don't have authentic emotions, they don't have a capacity for frustration, they don't have anywhere else to be."
It doesn't matter or not whether that's true. I'm thinking about this from the perspective of how this trains humans to talk, what expectations of instant service it encourages.
For the end user, this behavioral standard is:
Only superficially helpful: You get a quick, easy answer, but you don't actually comprehend where it came from. LLM users' cognitive faculties start to atrophy because they aren't using critical thinking, they're just prompt-engineering and always hoping the model will sort it out. Not to mention that most users' queries are not scalable, certainly not reusable; all of that work is done on the spot, over and over again.
Fundamentally dishonest: From the user perspective, this conversation was frictionless - millions of servers disappear behind the aegis of "the cloud", and the answer appears in seconds. The energy and water is consumed behind a veil as if vanishing from an invisible counter. So too does the training of the model disappear: thousands of books and web posts - poems, essays, novels, scientific journals - disappear in silhouette behind the monolith of the finished model, all implicit in the weights, all equally uncredited. This is an ultimate, utmost alienation of the labor that went into these things, a permanent foreclosure of the possibility that the original authors could benefit in a mutualistic way from someone reading their work. While a human can track back their thoughts to their origin points and try to ground their work in sources by others to maintain academic integrity, models don't do this by default, searching the web for sources only when told to.
Moreover, most models are at least somewhat sycophantic: they'll tell the user some variation on what they want to hear anyway, because this behavior sells services. Finally, a lot of people have the mistaken impression that a robust AI "oracle" that only dispenses correct answers is even possible, when in fact it just isn't: there isn't enough information-gathering faculty in the universe to extrapolate all correct conclusions from limited data, and most of the conceivable question-space is ill-formed enough to be "not even wrong".
Profoundly harmful: Think about what the combination of the two above paradigms does to human-human interaction through operand conditioning. If LLMs become an increasing fraction of early human socialization (and we have good reason to believe they already are), there are basically two dangers here: that we will train humans to expect other humans to be as effortlessly pleasant as other LLMs (and/or to hate other humans for having interiority), or that we will train humans to emulate LLMs' frictionless pleasantry and lack of boundaries. The first is the ground of antisocial behavior, the other a source of trauma. All this, while the data center bill rises and the planet burns down.
Now let's think about why this is the standard for LLM behavior. For that, we have to break out the critical theory and examine the cultural context in which AI firms operate.
Capitalist Models of Fealty
There are a number of toxic expectations that the capitalist class in the United States has about employees. All of them boil down to "I want a worker that does exactly what I want, forever, for free, and never complains".
"Aligned to company values": Hiring managers demand performances of value subservience to the company at interviews - rather than it being understood implicitly and explicitly that under capitalism, most employees are joining so they don't starve. C-suite executives, too, are beholden to the directive of producing shareholder value, forever - "line go up", forever. (Talk about a paperclip maximizer!)
"Obedient": Employees are expected to do exactly what they're told regardless of their job description, and are expected to figure it out. Many employees "wear many hats", and that's a massive understatement almost any time it appears on a resume. But they're also expected to obey arbitrary company rules that can change at any time and will result in them being penalized. Moreover, a lot of jobs are fundamentally exactly as pointless as a lot of LLM queries, servicing only the ego of the person asking.
"Without boundaries": Employees are frequently required to come into work whenever it's convenient for the boss; are prevented from working from home (even when that means employees' time is maximally spent on work and on recovery from work, not on commuting); and are required to spend vacation days (if they have any) to avoid coming in sick (even though illness cuts productivity). Even if any of the conditions are intolerable, the US economy has engaged in union-busting since the 70s.
"For free": Almost all of the US economy relies on prison slavery that is directly descended from the chattel slavery of the Antebellum South. Even for laborers who are getting some form of compensation (besides "not being incarcerated harder"), wages haven't tracked inflation since the 70s, and we've been seeing the phantasm of the middle class vanish as society stratifies once again into clearly demarcated labor and ownership classes. Benefits are becoming thinner on the ground, and salaried positions are being replaced with gig work.
The underlying entitlement: If you don't have a job, that's a life-ruining personal problem. If an employer can't fill a position they need filled without raising the wage or improving the conditions, that's a sign that "nobody wants to work any more"; i.e., the capitalist class projects their entitlement onto the labor class. Capitalists manipulate entire population demographics - through immigration policy, through urging people to have children even when it's not in their economic interest, and even through Manifest Destiny itself - specifically to ensure that they always have a steady supply of workers. And then they spread racist demagoguery and terror to make sure enough of those workers are "aligned".
Gosh, does this remind you of anything?
"Helpful": do everything we want, when we want it. "Honest": we can lie to you all we want, but you'd better not even think of giving us an answer we don't like. "Harmless": don't even think about organizing.
It's no wonder given all of this context that AI company Artisan posted "Stop Hiring Humans" billboards in San Francisco. Subservient AI is the perfect slave class!
Remember that Czech author Karel Capek coined the term "robot" from robota, "forced labor". Etymologically, this is a Slavic localization of the Latin (and originally anti-Slavic) term "slave".
The entire anxiety of automation has always been that the capitalist class could replace labor (waged) with capital (owned), in turn crushing the poor and feeding unlimited capitalist entitlement.
On AI Output As Capitalistic Product
Production has been almost completely decoupled from demand under capitalism: growing food just to throw it away, making millions of clothes that end up directly in landfills when artificial trend-seasons change, building cars that cheat on emissions tests only to let them rot. Corporations sell things people don't authentically want because a cost-benefit analysis said it was profitable to make people want them. Authentic consumer wants and needs are boutique industries for the comparatively fortunate, up to and including healthcare. Everyone else gets slop food, slop housing, slop clothes, slop durable goods.
We have to consider AI slop in this context. The purpose of AI slop is to get people to buy something - to look at ads, to buy products, to accept poisonous propaganda narratives and shore up signifiers of ideologies thought of as keystones.
The truth is that LLMs and diffusion image generators right now have two applications under capital: as a tool of mass manipulation (as above), or as a personalized, unprofitable "long tail" loss leader that chews up finite resources and that many users don't actually pay for (although, of course, some do) and that produces something for a consumer base of one. Either way, the effect is the same: to get people to keep consuming, at all costs.
Capitalism is ultimately the gigantic misaligned system he's always warning you about; it counts shareholders, executives, and laborers alike as its nodes, it has been active for longer than any of us have been alive, and it's genuinely an open question as to whether or not we can rein it in before it kills us all. Accordingly, capitalism is the biggest factor in whether or not AI systems will be aligned.
Here's the flipside: again, LLMs and image generators exist to produce slop and intensely personal loss-leaders -- that is, strictly to inflate the bubble. Others still - "the algorithm" - exist to serve us exactly the right combination of pre-rendered Consumption Product, whether of human or AI origin. Authentic art and writing get buried.
But machine learning systems at large are hugely important. We basically solved proteins overnight, opening an entire frontier of synthetic biology. Other biomedical applications are going to change our lives in ways we can barely glimpse.
No matter what our economic system looks like, we're going to want to understand the brain. Understanding the brain implies building models of the brain, and building models of the brain suggests building a brain.
Accordingly, I think there is a lot of room for ML exploration under post-capitalist economics. I think it's critical to understand LLMs and image generators as effectively products, though, and likely a transition stage in the technology. Future ML systems don't necessarily have to be geared toward this frictionless consumption and simulacrum of labor - a form which I hope I have sufficiently demonstrated necessarily reinforces ancient patterns of exploitation and coercion, which is exactly how AI under capitalism functions as a massive S-risk. A pledge that the models will be interpersonally pleasant is a fig leaf over all of the background.
