Regulators can’t shut Pandora’s box.

There’s a lot coming out in the news right now about regulation of generative AI models (AI for short). Legislators and company CEOs are banding together to call for increased regulation, with OpenAI’s Sam Altman testifying to Congress that he fears AI could cause ‘significant harm to the world’ and calling for a licensing system to be implemented. However it’s clear that the growth in open-source AI means regulation is already futile. All regulation will achieve is raising legal barriers for smaller enterprises and stifling innovation by pulling up the ladder.

Open-Source AI is freely available

Right now if you head across to Hugging Face (an open-source AI model repository) you can choose from over 12,000 AI models that provide ChatGPT-like text generation, download any of them, and train/run a significant number of them locally. In many cases the repositories contain the original training data. The best of these open-source models produce comparable results to the OpenAI GPT series. To get started an engineer only needs to type a couple of shell commands like these:

git lfs install
git clone https://huggingface.co/distilgpt2

Regulating AI or issuing licences won’t prevent bad actors from accessing them. They already have access to them.

Open-source models are developing exponentially

It’s helpful to understand how fast AI models are developing in the wild. A recently leaked internal Google document provides a helpful timeline of how quickly the wild AI models developed from the first freely available leaked model (LLaMA) to a GPT-3 comparable model (Koala). It took around 39 days, or 1.3 months. That’s a curve that makes Covid’s spread look sluggish.

Demand for AI products is huge

Users want AI-powered services. Since ChatGPT was launched around 6 months ago the service has accumulated around 100 million users. By comparison, it took Facebook four and half years, Instagram two and half years, and TikTok nine months. AI products are popular because they solve big user problems (writing is hard, creating compelling images is hard) and they are extremely easy to use.

The history of peer to peer file sharing shows that where a new technology is freely available, developing exponentially and solves a big user problem regulation does nothing to limit availability. It wasn’t until Spotify came along dealt the final blow that illegal music distribution pretty much ended. The situation with AI is comparable, but user demand is actually far greater for AI.

Don’t needlessly stifle innovation

Attempting to regulate AI won’t work. The underlying technology is too available and developing too quickly for a conventional regulatory cycle to keep pace with. A licensing regime will do nothing to stop bad actors from developing products that leverage open-source AI models for harm.

Instead regulation will halt innovation by small companies (like Eject) who are trying to harness AI to make the world a better and more productive place. We will have to comply with and bear the costs of a licensing regime. Bad actors and individuals will be able and willing to ignore any licensing requirements.

You might point out that we have an obvious interest in AI regulation, since at Eject we’re currently building a product that leverages AI. Fair enough. But we hope that legislators and regulators are prepared to listen to our point of view because we’re actively creating AI-derived and open-source technology. Businesses like ours are the future of the economy. Right now the world is at a pivot point and these decisions are big ones. It’s important regulators base them on a realistic appraisal of the developing situation on the ground.