A Warning from a Computer Scientist: Develop a Research Framework for Ethical AI Development

A computer scientist's plea to stop the unethical use of LLM technologies.

I strongly believe we are now at war. But the war isn’t a physical conflict, it’s how to develop, or not develop AI.

I strongly believe we need an ethical AI research framework. As a computer scientist, I have seen firsthand the development of AI, and I have met researchers along the way that truly believe what they’re doing is right.

I remember being shown GPT-2 as a Large Language Model (LLM). It couldn’t do anything. It could just about string sentences together. And now, it feels like the technology is used by millions of people every day for who knows what reason. They’re not all academics. Why are they using it?

Computer Science is a profession that I think has lost its grace within the general public, and rightly so. People are sceptical of new technologies, and I think this is actually quite a good thing. I sit here, as a computer scientist, and I look at what’s been created, and I can’t help but to scream. I think the public is right and we too, as computer scientists, are fed up of ‘AI experts’ and ‘techbros’ spreading misinformation, like Elon Musk.

We could have used this technology to help children in classrooms learn with local models tailored to what the teachers need. We could have used this technology as a very small scale local model personalised spell and grammar checker that helps you to not focus on how your message is written, only how it’s delivered. We could have used this technology to automate boring, pointless bureaucracy in regard to forms, which aren’t usually designed for a human in mind anyway and nobody should have to stuck doing data entry.

Instead, what we got, was complete control of this by a few companies. OpenAI. xAI. And I do genuinely believe that these researchers think that they are good in the world. But I can’t tell you, as a computer scientist, if they are.

I cannot go in and audit the source code of anything they do. I can obviously see the little bits of research, like local models, but I have no idea how these commercial models work. And when I ask it, it tells me all these AI techniques and algorithms that it uses.

But does it? Does it actually those? Because the LLM will just tell me what other people think about it. It does not know anything about itself. It doesn’t know its own emotions. It is merely a robot. And people forget that.

But who programs these robots? Who is responsible for feeding it data? Which workers? How are they compensated? What framework do they work under? How are they feeling about the work they do? Are they becoming AI engineers because they want to, or because they need to do it to survive as their jobs slowly get taken away.

So I would like to propose a research framework. I know that this may be buried under some more authoritative voices, for I am merely a Batcheor of Science in the topic, but I am still a student of Computer Science for life, and that’s something nobody can take away from me.

Here's my personal thoughts about using AI such as LLMs, and these may change over time:

• AI must be a tool for academic use and policy making use only. AI cannot be forced on people who are just trying to get by, i.e. people in entry-level jobs. AI use there must be freely consensual, and I do not believe that this is possible under a capitalist framework.

• AI use must be sparing, and every query must be carefully considered. It must be the initial point of research, and not something that is a finished product in any way.
  I believe that there’s only two more ethical uses as a form of harm reduction (I do not think that using AI as it currently is, is ethical):

  • Initial research in a form of a search engine to see what other people are saying. You’re basically asking an librarian by doing so, as you are essentially using it as a form of a Google. And that, is how you should treat the output it gives you if you are not using it directly for AI research. It’s a better Google. And you need to check our sources more than Google because it can absolutely tell you shit that isn’t real.
    
    But I’d rather you asked an actual librarian - they are heavily underpaid and underappreciated and librarians will typically know better than a search engine.

  • In order to keep yourself informed about AI research. You may experiment and test the current prototype. This is an accepted use of AI research as it is important to understand the current capabilities of the machine so that we know what the problems are with it.

• AI research must be public research. Every part of AI development should be auditable and computer scientists like me, should be able to view the source code, no matter where they may be. This research MUST be Open Access – no matter how complex this web gets; there are some exceptionally smart people who could even use LLMs to understand LLMs.

• Where possible, local models should be used. This allows for accountability of the resources used, since this is a measurable outcome. We must assume that the public models, until proven otherwise, are grossly inefficient and optimised so that they survive in a capitalist framework. We must assume they will feed us propaganda about the wonders of capitalism.

• Anyone who breaks the rules, must be held accountable. If someone is caught blatantly stealing AI output, they must be stopped and they must be educated on the importance of using AI ethically. And while it isn’t possible to have consequences for bad search queries, much in the same way that it’s not possible to have consequences for bad search queries you type into DuckDuckGo, I do believe in calling people out for this kind of behaviour.

• You should not pay for commercial AI. A major chunk of funding for AI is venture capital. By paying into it, you are perpetuating the cycle of resource abuse by being complicit in its use.

• Consider if you need to use AI. Is there another option? Could you use some of the work that has been already done by others, and therefore saving resources. Do you need to train yet another model? Use what has been already done, and optimise your research as much as possible.

• Consider how large your LLM model you’re using is. A larger model has more problems. But a small scale LLM model can be used to model behaviours in language just as effectively. You don’t need to use a big model most of the time, and that is where most of the resource concerns come from.

Those are my rules I’ve come up for AI research thus far. I think it’s an incredibly difficult problem. This problem is whether we, as computer scientists, develop the Manhattan project, the pending doom that will cause numerous amounts of suffering.

I do not think that AI inherently solves societal problems by allowing everyone to use it. As a computer scientist, I do believe in limited uses of ethical AI, such as in true industrial applications where automation will allow others to become creatives, and we can support more creatives in this world as a result.

But creativity must not become automated, for it is then not creative. To be creative is to be human. It cannot inherently help someone to be more creative, because all it will do is say the same things that everyone else has said, because it hasn’t got the ability to create new ideas.

Create nuclear power stations. Not the atomic bomb. That’s my message to my colleagues. Please consider the impact of the work you’re doing and steer this bloody technology away from profit-making. It is not going to solve humanity. Please at least advocate for a Universal Basic Income and for the working classes if you keep developing AI technology. And consider whether you, personally, need to be the one working on it.

Thank you for reading my post. I hope you have come away more informed.