The Defence Production Act (DPA) has been reauthorised (and modified) by Congress since 1950, and in 1953 its powers very significantly reduced. I’m confident that it will continue to be passed—in a roughly similar form - for the foreseeable future. The current version was passed in 2019 under a Republican senate and is due for reauthorisation in 2025.
Since the Obama Presidency, there’s Republicans have begun to try to prevent bills proposed by Democrats from being passed by default. This is particularly easy for non-spending bills since for non-spending bills, 60 votes are needed to break the filibuster—a method used to prevent bills from being voted on—in the Senate. However, not only are defence bills consistently bipartisan, they have consistently high degrees of support from Republicans in particular. Therefore, I’m not concerned about the DPA not being passed by a Republican senate and a Democratic President when it’s next due for reauthorisation.
The DPA gives the President very wide-ranging powers since the goal of the act of the act is to ensure that the US industrial base is strong enough to fight and win any war the US might need to undertake. Concretely, this allows the President to instruct firms to accept contracts; incentive expansion of the industrial base; and a grab bag of other specific provisions aimed at making sure that the US production base is strong enough to win a war.
Until 1953 the act was much more powerful and essentially allowed the President to take control of the US economy. The act now doesn’t give the President authority to set wage or price ceilings, control consumer credit or requisition stuff.
Antitrust provisions
Various AI governance proposals rely on explicit, voluntary agreements between leading AI labs. For instance, this paper proposes a scheme in which AI firms agree to pause the rollout out and training of large models if one doesn’t pass an evaluation which indicates that it could act dangerously. I think it’s plausible that this agreement would be illegal under antitrust law. An agreement like this would be an explicit agreement amongst a small number of leading firms to limit supply.
It’s not at all clear that such agreements would be illegal—for instance, professional standards are not considered per se illegal but instead are judged under a rule of reason where their competitive effects need to be outweighed by their procompetitive effects. I won’t comment on this further but instead, refer the reader to this excellent by that looks specifically at anti-trust and AI industry self-regulation.
Section 708 of the DPA gives the President authority to allow firms to make agreements for that would normally be considered antitrust violations. Recently, this provision was used by the Trump administration during Covid-19.
Use in the Biden executive order
Some of the most AI safety-relevant elements of the recent Biden executive order on AI were authorised under the legal authority of the DPA. This includes:
Requiring AI firms to report training runs over some compute threshold and report red-teaming results
Instructing NIST to develop new AI safety standards
Instructing federal procurement to Notably though these elements don’t require firms to take any actions other than reporting their training runs and the results of red-teaming ect to the federal government.
The Defence production act and AI policy
Quick Summary
Gives the President wide-ranging powers to strengthen the US industrial base
Has been around without changing that much since 1953
Has provisions which allow firms to make voluntary agreements that would normally be illegal under antitrust law
Provided the legal authority for many of the provisions in Biden’s recent Executive Order on AI
The Defence Production Act
The Defence Production Act (DPA) has been reauthorised (and modified) by Congress since 1950, and in 1953 its powers very significantly reduced. I’m confident that it will continue to be passed—in a roughly similar form - for the foreseeable future. The current version was passed in 2019 under a Republican senate and is due for reauthorisation in 2025.
Since the Obama Presidency, there’s Republicans have begun to try to prevent bills proposed by Democrats from being passed by default. This is particularly easy for non-spending bills since for non-spending bills, 60 votes are needed to break the filibuster—a method used to prevent bills from being voted on—in the Senate. However, not only are defence bills consistently bipartisan, they have consistently high degrees of support from Republicans in particular. Therefore, I’m not concerned about the DPA not being passed by a Republican senate and a Democratic President when it’s next due for reauthorisation.
The DPA gives the President very wide-ranging powers since the goal of the act of the act is to ensure that the US industrial base is strong enough to fight and win any war the US might need to undertake. Concretely, this allows the President to instruct firms to accept contracts; incentive expansion of the industrial base; and a grab bag of other specific provisions aimed at making sure that the US production base is strong enough to win a war.
Until 1953 the act was much more powerful and essentially allowed the President to take control of the US economy. The act now doesn’t give the President authority to set wage or price ceilings, control consumer credit or requisition stuff.
Antitrust provisions
Various AI governance proposals rely on explicit, voluntary agreements between leading AI labs. For instance, this paper proposes a scheme in which AI firms agree to pause the rollout out and training of large models if one doesn’t pass an evaluation which indicates that it could act dangerously. I think it’s plausible that this agreement would be illegal under antitrust law. An agreement like this would be an explicit agreement amongst a small number of leading firms to limit supply.
This skirts pretty close to being a criminal violation of US antitrust law. Under this law, various forms of agreements between firms to fix prices are considered illigal no matter what firms say is the justification for them (this is known as per se illegal.) Agreements to limit production are considered just as illegal.
It’s not at all clear that such agreements would be illegal—for instance, professional standards are not considered per se illegal but instead are judged under a rule of reason where their competitive effects need to be outweighed by their procompetitive effects. I won’t comment on this further but instead, refer the reader to this excellent by that looks specifically at anti-trust and AI industry self-regulation.
Section 708 of the DPA gives the President authority to allow firms to make agreements for that would normally be considered antitrust violations. Recently, this provision was used by the Trump administration during Covid-19.
Use in the Biden executive order
Some of the most AI safety-relevant elements of the recent Biden executive order on AI were authorised under the legal authority of the DPA. This includes:
Requiring AI firms to report training runs over some compute threshold and report red-teaming results
Instructing NIST to develop new AI safety standards
Instructing federal procurement to
Notably though these elements don’t require firms to take any actions other than reporting their training runs and the results of red-teaming ect to the federal government.