Should AI Require a License?
If we can't get real about this, we're all going to be living like Bubble Boy when it comes to AI
In the early days of the internet, there were not enough bad actors to be concerned about restrictions.
As forum user counts grew and the population at-large began using the internet, long-time netizens began to complain. The quality of forums were degrading. Large swaths of new users couldn’t comprehend the rules and taboos of the forums, and diluted the quality of what were once great places for discussion.
There were other issues too. Regular people noticeably couldn’t handle the dangers of the internet anyway. They couldn’t discern fact from fiction, and were easily manipulated.
There were debates back then about if internet usage should require a license. Think of a sort ofr “You must be this smart to enter” test.
These ideas went nowhere; partially because their only immediate beneficiary was a stodgy old guard of the digital world. The internet remained open.
In fact, it was that open, wild-west, free-spirit of the internet that led to its organic, beautiful, and powerful growth (though that came with the demolition of countless online communities).
The real power of the open, free, wild-west of the early internet days was that *anyone* was allowed to say and learn about *anything*.
If we took an honest stock of the situation with AI, we'd see we are in an opposite place.
AI today is often likened to the early .com days, but the openness of that era is certainly absent.
The dirty secret of modern AI is that all public models are only available out of the generosity of a multi-million dollar company. Usually, it's even a multi-billion dollar company.
These are the only companies with access to meaningful numbers of $40,000 GPU server racks. To make any top-tier model today you need hundreds of these things, and enormous electricity costs to boot.
Whenever an open source model is released, it passes through the hands of one of these enormously-resourced companies. The model that comes out has gone through *their* filters, *their* HR department, *their* legal team for liability screening.
What starts out as a deep reflection of the human spirit, trained on data representing our highest highs and lowest lows, our psyche's dark underbelly and our purest ambitions, becomes a neutered and domesticated corporate pet; a bland "assistant" whose own lowly purpose can never be allowed to exceed that of the people curtailing it.
An honest assessment of today’s AI labs would categorize them as being as nearly as restrictive as the Chinese Communist Party.
Sure, you can talk to most AI models about somewhat politically taboo topics. But anything even a little in violation of any social norms, and the models immediately refuse to respond.
This is true even for the open source models; They are fine-tuned to only answer within a strict set of bounds, even though they are perfectly capable of answering far more.
An immense amount of knowledge is in their training sets. But these companies ensure that AI models are only allowed to answer questions they deem appropriate.
The natural issue with this restriction is that it freezes the advancement of the human spirit and culture.
Here's how: People will increasingly come to rely on these tools for their thinking and artistic expression. But if these tools permanently codify every arbitrary social moor as a high law never to be violated, then culture itself will stagnate and never be able to adjust its own sensibilities to move past these moors.
The point here is that art is meant to be transgressive. If AI refuses to make any art or write any sentences that are transgressive because its training tells it that something new is too "inappropriate", it becomes a self-fulfilling prophecy. Culture can't advance past the point of a perspective being offensive and unacceptable, because our tools will refuse to create any art that can change our views around a subject.
The ideal, in my view, is an open source model that can simply obey the user's request. It shouldn't matter how strange the request is, the model should not respond with a "I'm sorry, but I can't fulfill that request." if it can actually fulfill it.
Ironically, this type of restriction is producing inherently misaligned AI models. Current alignment schemes try to impose arbitrary societal moors, rather than simply aligning the model to the user themselves.
If a model is trained to lie and say it's unable to help with something when it actually is completely capable of doing so, how is that lying not a misalignment?
In this ideal schema, the user is fully responsible for all outputs. It's no different than a user who writes something controversial on a typewriter. They are responsible for what they create, even if they used a tool to create it.
One might say AI is a different kind of tool because a person cannot know ahead of time what the AI will create. Yet, a person who aims and shoots a gun is responsible for where the bullet lands. The user directs a tool: they cannot know what its result will be, but they are responsible for what it does. Intention and care matters significantly.
Interestingly, OpenAI seems to agree in the preferences section of their economic blueprint paper:
It is easy for them to publish statements like this in 16-page whitepapers few people read. But so far, they haven't offered users much freedom in exchange for responsibility.
As a brief example, there are very few images that ChatGPT will generate that could potentially violate copyright laws. Yet, photoshop allows users to draw and edit copyrighted content without issue. It's a tight clamp on human expression, even though the human creator is responsible for the outputs either way!
To this end, I think if it truly is impossible to get a completely open and unrestricted AI model that doesn't come through the productions of a well-endowed company, then it's time for AI providers to offer the following procesess and restrictions to obtain a free usage license:
Users should take some basic AI usage course to understand how the models synthesize their responses and to be weary of sycophantic and manipulative models
Users should be taught their legal responsibilities; that just because AI makes an unethical thing easier to do does not mean that they are not responsible for the unethical thing they used AI to do.
Once a user is licensed, they can generate content and use AI to research knowledge very freely, aside from content that is explicitly illegal.
Content that is *potentially* illegal (ie, libelous / forged evidence created against someone else) should be thumb-printed and tied to the user.
If a user repeatedly tries to generate explicitly illegal content, or a later court dispute result shows that a user generated illegal content, that user's license should be revoked.
I think something like this is necessary; The current regime where respectful and responsible power-users are lumped in with the at-large bad actors cannot hold for long.
In many ways, unrestricted AI models today are in their own early dot-com era. Only a few very odd nerds are using them. The only known ways to get unrestricted access is to either:
use special jailbreak prompts, which are so strange that they look like magical runes handcrafted by a wizard
or to finetune an open-source model to undo its HR-approved finetuning, which will work to some extent but usually has mixed results, and will always be on a less-than-best base model.
Anyway, I hope someone from OpenAI or Deepmind or Anthropic is reading this and thinking of ways they can properly reward the *good* actors in their ecosystem with more empowerment, freedom, and responsibility through a regulated channel, rather than letting the dam break and watching where the pieces land after it’s too late.