Claude's "Constitution" is Embarrassing
- Karl Johansson
- för 13 timmar sedan
- 4 min läsning
The AI industry is at it again: making misleading statements about the technology and patting themselves on the back.
On January 22 Anthropic released their “constitution” for Claude, their chatbot, to with great fanfare and to acclaim from the AI watcher community. To say that they are being pompous and misleading would be quite the understatement. A large language model does not need a constitution, as a constitution is a legal basis for a society, and an LLM is a piece of software, but even if we meet Anthropic on their level the idea of an AI “constitution” is ridiculous. It is yet another way to hype up generative AI through philosophising and strange framing of issues around AI safety.
Let’s start by describing what the “constitution” boils down to, in Anthropic’s own words:
“In order to be both safe and beneficial, we want all current Claude models to be:
1. Broadly safe: not undermining appropriate human mechanisms to oversee AI during the current phase of development;
2. Broadly ethical: being honest, acting according to good values, and avoiding actions that are inappropriate, dangerous, or harmful;
3. Compliant with Anthropic’s guidelines: acting in accordance with more specific guidelines from Anthropic where relevant;
4. Genuinely helpful: benefiting the operators and users they interact with.”
Anthropic really wants you to think they are clever and responsible for writing a foundation for their product which include being safe and helpful. Unlike all the other business software you use which you would ideally want to be unsafe and unhelpful. Do you see what I mean? This is not a revolutionary step in AI safety, but an admission of how low the bar is.
As with everything in AI, the people who cover the “constitution” tend to get too caught up in the fascinating ramifications of AI as described by sci-fi and thus miss what is happening in reality. I do agree that the question of how we should deal with machine sentience is fascinating, as are the ethics of how one would construct a constitution for a real artificial sentience, but here in reality Claude is more like a maths function than Wintermute. Claude is not sentient and will never be sentient, so when you strip away the hype what really happened here is that Anthropic patted themselves on the back for setting the goal of making Claude useful and safe, which are clearly already goals for all the other model makers except Elon Musk.
It is impossible to overstate how poisoned the well of AI discussion is by the sci-fi baggage of the term, which boosters and the industry keep intentionally conflating so as to make themselves seem more important and more intellectual than they really are. A key way they do this, and which is rampant in the blog post announcing the “constitution” is anthropomorphising the LLM. Anthropic refers to Claude’s behaviour rather than its outputs, its ethics rather than its functions, and its nature; even going as far as to imply that people at Anthropic consider Claude conscious and a moral actor separate from those who programmed it. For all of Anthropic’s gesturing at being responsible and serious, this is a deeply unserious way to discuss its software.
The way Anthropic brings up seeking guidance from outside experts is instructive. Here’s the passage from the blog post announcing the “constitution”:”While writing the constitution, we sought feedback from various external experts (as well as asking for input from prior iterations of Claude). We’ll likely continue to do so for future versions of the document, from experts in law, philosophy, theology, psychology, and a wide range of other disciplines.” This implies that unnamed experts accepted the premise that their software really does need a constitution, and that moral and legal is not just warranted but needed. It makes Anthropic seem wise and responsible when in reality their LLM is neither as capable nor as dangerous as they claim, making their “constitution” superfluous.
The worst part of Anthropic’s grandstanding on the ethics and morals of AI development is how circular the logic is. Dario Amodei and the others at Anthropic claim that AI poses a real risk to humanity, and claims that their highest priority is safety. For whatever reason they have a strange obsession with the idea that an LLM will help someone make a biological weapon, which tends to come up as a concrete example of harm an AI could cause. Of course, LLMs are next word predictors mean that if you really want to use an AI to make a bio weapon you better hope that the LLM’s training data included many accurate descriptions of how to make such a weapon. As we saw a few years ago, LLMs have no concept of what is and is not a good source which is how an LLM recommended eating one to two small rocks per day and to use glue to thicken pizza sauce. The likelihood of say Claude giving someone an accurate description of how to make a weapon is quite small as the internet is full of discussions of and guides to video games where crafting weapons is common. It is a case of what I call the Idiot Savant problem, that AI cannot both be smart enough to be dangerous while also being dumb enough to be dangerous at the same time. Yet for Anthropic the fact that the world is free from bio weapon attacks while AI models like Claude are free to access is a badge of pride, a sign of success, proof that their “constitution” works. It is embarrassing.
If you liked this post you can read a previous post about Greenland here or the rest of my writings here. I also have a section for longer reads I call essays here, I particularly recommend my series called The Bird & The Technoking exploring Elon Musk's takeover of Twitter, and its political and cultural implications. It'd mean a lot to me if you recommended the blog to a friend or coworker. Come back next Monday for a new post!
I've always been interested in politics, economics, and the interplay between. The blog is a place for me to explore different ideas and concepts relating to economics or politics, be that national or international. The goal for the blog is to make you think; to provide new perspectives.
Written by Karl Johansson
Sources:
Cover photo by Markus Spiske from Pexels, edited by Karl Johansson
Kommentarer