I am extremely disappointed in the recent senate hearing on AI regulation. The level of authoritarianism and control required to truly regulate AI will bring about a draconian world in which the very values we wish to protect will be destroyed.
The 3 witnesses failed to address the positives AI will bring and do not focus on solving the root of the problem at hand. Instead they are targeting the very freedoms of our world in an attempt to regulate AI. They are asking congress to control online speech to prevent interference of elections, mandate ID verification on social platforms, and even requiring DRM on all chips to make sure users do not try to jailbreak models.
The witnesses are Dario Amodei of Anthropic (Google’s AI), Yoshu Bengio (Quebec AI) and Stuart Russell (Professor of University of California Berkeley).
I will list the most significant and concerning statements higher up.
- 1:25:00 Stuart Russell tells senate members that the only means to control users from jailbreaking models is to have chips themselves reject certain prompts and models (DRM).
- 38:24: Yoshu Bengio discusses how fine tuning of open source models are used to circumvent the censorship and that he wants the government to regulate this.
- 41:30: Stuart Russell wants all chat from any LLM to be saved so that it can be compared with a database to determine the origin of the conversation.
- 39:50: Yoshu Bengio wants the government to prevent open source AI from being released going forward.
- 37:59: Dario Amodei wants to require watermarks of all AIs including open source. This is a privacy violation unjustified given the good LLMs are doing.
- 56:24: Senate member Klobuchar seems to be pushing for an answer to regulate open models to prevent them for being used for scams even though the benefits open source models are providing significantly outweighs the downside.
- 2:07:18: Yoshua Bengio and Yoshu Bengio discusses regulation of open source, and especially open source going into the future.
It is important to also mention that not all points given in the hearing were unjustified. Some suggestions such as the requirement for transparency by larger corporations when it comes to red flag capabilities from an AI are reasonable. However some requests given to congress show a disregard for the basic freedoms and values of democracy that will be destroyed if regulation is not considered extremely carefully.
Full Hearing: https://piped.video/watch?v=hm1zexCjELo
To be honest, regulating AI models seems very important to me, because one day humanity (or companies) will trust AI with a dangerous amount of power. The training data use will probably originate from the internet, making the model corruptible and susceptible to errors. However, spying and logging every user prompt is clearly not the answer.