Imagine you wake up sometime in the early 1940s and the first thing you do is check Twitter. (Yes, there’s Twitter in my 1940’s, just go with it).

You’re in the 2019 1940’s and you read a Tweet from a prestigious, private organization that’s funded by a notorious billionaire.

“Hey guys! We are so excited to release the latest paper in our series on democratizing nuclear fission. Check us out at https://arxiv.org/abs/@(#%($; title of the paper is “Initiating Chain Reactions of Fission in U-238 Atoms” by Einstein, Oppenheimer, Furman et al.”

If you’re appalled at my Extreme Hyperbole™, you should be. If you’re also mad at me for not getting the history of nuclear fission correct, I’m mad at me too and I’ll correct myself to say that probably the first paper that started mankind down the road to nuclear power was by Otto Hahn and Fritz Strassmann, and it happened in 1938. Anyway sorry not sorry for my very bold stylistic choice, and technophobia is certainly not the answer to this dilemma, but the issues of AI Export and publishing research are present and real.

I digress.

Back here in regular 2019, research on machine learning is moving forward at an astounding pace (in case you haven’t noticed). Just yesterday, OpenAI, a non-profit funded by Elon Musk had a big day in the media when they released a new model that is a step forward on the road to robots being able to simulate human language text effectively, and to generate (not just understand) intuitive language.

I’m no researcher, but I do have a metered (and pragmatic) fear of certain ML models that mimic human behavior, and a pretty sizeable problem with research coming out of the field we all know as Deep Fake. If used for ill, these things can streamline, amplify, and multiply the effects of foreign trolls and fake news election interference.

OpenAI says that their release of this text generation research was done in an “ethical” way, with respect for the power that the technology imbues. They claim to have trained much larger models than the ones they put out there, and have refused to release those at all. Instead, they held back 3 of the 4 models they trained, and declined to publish any of the code they wrote to make it.

The problem is, they did release the paper. And if you’re into machine learning, there’s absolutely nothing stopping you from building the big models yourself.

Their “respectful and ethical” approach to releasing this information has been lauded all around AI Academic Fairness Twitter since it came out, and has sparked a bit of a debate.

When is research too dangerous to release?

This isn’t the first time that humanity has asked itself this question. In the life sciences, there must be research to analyze and assess various nasty pathogens with the aim of curing them, but of course those can always be used for genocide as well. In that field, there’s this thing called Dual-Use Research of Concern, or DURC for short.

Surprise! DURC of potential biological weaponry is heavily regulated.

Nobody likes regulation, especially when it has the potential to restrict your life’s work, or prevents you from making money doing what you do. In fact, studies show that life sciences researchers working on protected pathogens don’t even consider their work to be potentially problematic. Talk about denial.

That link I threw in the paragraph above is really interesting because it asks researchers what frameworks of regulation they’d like to see. Few of them said that regulation shouldn’t be imposed, but they heavily disagreed on how to do it. Everyone was sad that it might slow down the pace of their work (and many of them worried about foreign competition pressing forward regardless; a common theme in AI these days).

Here are the options, according to the researchers studied in the paper:

1. Case by case regulation (every researcher has to get approval)

2. General rules regulation (like a code of conduct)

And the regulation would have to be enforced (likely through funding) either by:

1. The Government

2. Universities

3. Journals

4. Independent Advisory Body

5. National Funding Agency

I’m not going to propose an answer in my tiny little blog. I will say that the debate has to move forward. It has to be an interdisciplinary discussion, involving a diverse group of minds. And it has to happen soon.

It’s important to note that regulation slows progress and restricts scientific freedom, which really sucks. In the life sciences, we’re talking about research that could potentially save lives were one of these particularly horrifying pathogens to make it out into the public.

I don’t want to stop progress, but an interesting point in this discussion is that the degree of regulation should be proportional to the likelihood that it will be used for evil.

Right now we have no reason to believe that Russia will be releasing weaponized anthrax on us, but we do know with complete certainty that they will attempt to use any and all available technology to interfere with our democratic elections.

So, is OpenAI doing the right thing here by policing themselves? Perhaps. Perhaps instead it simply makes for a really nice boost to their brand and reputation. Perhaps it is an attempt to prevent regulation that we desperately need. Either way, we have to do something about the fact that Deep Fake research regulation is completely missing from the political conversation right now.

Write to congress! Do it now!