A Playbook for AI: Discussing Principles for a Safe and Innovative Future

At this point, on the brink of 2025, most of us have had a chance to interact with at least a limited form of AI –  getting advice from a chatbot, planning the best route through the city, or online shopping with a virtual assistant.

But in the background, AI is quietly becoming significantly more intelligent, rapidly approaching what we call “artificial general intelligence”, or “AGI” - a type of far more advanced intelligence than we are used to dealing with, which can perform any task a human can, sometimes better. Which raises the question: how do we prepare for a future where we preserve the benefits of AI, and eventually work in harmony with AI, without getting overrun by it?

One organization with some answers to this question is the Manhattan Institute, which recently released “A Playbook for AI Policy”. For this week’s CAIP podcast, Jakub Kraus sat down with Nick Whitaker, a fellow at the Manhattan Institute working on emerging tech and AI policy, to discuss the four key principles underlying the Playbook, and where he sees we should focus our efforts in the rapidly developing context of as-yet-untested technology:

The Playbook: “Four key principles that can shape the future of AI and the policies that accompany them:”

1. “The U.S. must retain, and further invest in, its strategic lead in AI development.” 
2. “The U.S. must protect against AI-powered threats from state and non-state actors.”
3. “The U.S. must build state capacity for AI.”
4. “The U.S. must protect human integrity and dignity in the age of AI.” 

Below, I discuss my overarching takeaways from their conversation.

Technological Protectionism

To begin, a consistent but unsurprising theme is protectionism: both keeping foes out and keeping secrets in. Principle One reads at first glance like a typical “innovate faster than our competitors”, but it also is premised on the idea that we already have the advantage and need to retain it by bolstering our protections against espionage and sabotage. The U.S., as the leading developer of AI models but also the country with access to the best hardware needed, holds a lot of the cards - making us an obvious target for those adversaries who want a piece of the pie, likely to use against us. Investment is essential not only to get on top, but to keep it safe from intrusion.

The flip side of the coin is not only to keep the pirates out, but to keep the treasure in. Principle two emphasizes that the U.S. “must protect against AI powered threats by state and non-state actors”. Internal rules such as export controls are essential to protect precious algorithmic secrets.  For example, the podcasters discuss BIS standards that kick in at a certain amount of compute and hardware ops per second (10^26, and 10^20, respectively).

Mr. Kraus then notes that the Playbook proposes that evaluations also be done at AISI and private companies once they pass a 10^26 threshold, with a focus on autonomy, persuasion, and military assistance capabilities, at which point export controls and other restrictions should kick in. At the end of the day, advanced AI should be treated like any other sensitive military technology, so that an employee of a lab cannot be hired by a military enterprise of hostile foreign power, nor can they publish or give away these algorithmic secrets.

Yet while all of these are solid principles, discussion is not the same as action. The government will need to act efficiently using  limited resources (time, funding, and manpower). 

The Human Element

Indeed, the Playbook acknowledges that the current public sector is under-equipped to implement these recommendations, particularly in the agile fashion AI requires. To that end, Principle 3 proposes building state capacity for AI - in other words, strengthening the government's ability to understand, evaluate, and effectively integrate AI systems through increased funding, improved talent recruitment, targeted research investment, and standardized policies.

Part of this comes down to the brass tacks of more attractive salaries and benefits for existing experts; but there is also the human element - a need for personnel management - training the talent to operate in a public bureaucracy, which from the hiring process to day-to-day job can be slow and opaque. Of the many options available, ranging from public-private partnerships to temporary appointments, Mr. Whitaker favors the fellowship approach - but ultimately, he says, we won’t know at the outset what is working best - it’s advisable to try many approaches and see what has the best outcome.

This lighter-touch, trial-and-error approach makes sense when the stakes are relatively low, as they are for labor retention in government. We don’t know exactly which labor retention strategies will work best for AI experts in government, and there will be enough individual experts that we can try different strategies over time and then adopt the strategies that work the best.. 

However, I’m far warier about this trial-and-error approach when it comes to AI that affects the fundamental human psyche. Principle Four attempts to address this by saying that we must “preserve human dignity and integrity in the age of AI”, but so far I remain uncertain how this can be achieved in practice. The real-world harms AI has brought in the form of deepfakes, CSAM, misinformation, and fraud have thus far all seemed to outstrip our efforts to stifle them. It seems to be a game of whack-a-mole when it comes to the most obvious harms from AI, demonstrating how difficult it is to put the genie back in the bottle once technology has developed to a particular degree of sophistication. 

Yet still we forge ahead. Mr. Kraus brings up concerns about new AI systems that simulate therapists, which are sorely needed, but which can lead to seriously  adverse results if they are misaligned. Beyond that, he raises the question of romantic AI companions. Both of these simulate the experience of human connection so convincingly that it’s hard to see how they can be compatible with a desire to preserve core human integrity. They tap into our deepest vulnerabilities, giving them the access they would need to control us, should they become advanced enough to decide to do so.

In comparison to his more regulatory approach to other sides of AI (such as export controls and deepfakes), Mr. Whitaker argues that a fair market approach is best for these increasingly intimate AI models - he believes humans are ultimately in the best position to decide how to regulate their behavior. He argues that we can wait for new norms to develop like “no Twitter at work”, or “no phones at the dinner table.”

Nonetheless, he strikes a balance in the podcast, saying that as we take a hands-off approach, it falls to institutions in civil society and folks within government to proactively evaluate whether we are going down a path as a society that is beneficial and flourishing, or increasingly dystopian, and to adjust accordingly.

What Now?

That’s a balanced approach I can appreciate. It does seem that the more we find out about AI, and AGI in particular, the more hypotheticals arise. Maybe analyzing those potential futures and preparing proactively for each scenario is a good place to start, rather than making calls before kickoff.

But if we accept that thinking and assessing is the right, measured approach to an uncertain future, we need people to do that job. The podcasters continuously allude to the need for the government to participate actively in asking these questions - and in many cases, it does. Thus far, within the U.S., it has fallen largely to a few bodies - primarily NIST, and within NIST, the AI Safety Institute, and NAIRR - to do the lion’s share of this job. 

But those projects require support from Congress: NIST overall is increasingly tasked with new projects without the necessary resources or, in many cases, authority to do the job. Meanwhile, NAIRR remains in its pilot phase and the AI Safety Institute was enacted by the October 2023 Biden AI Executive Order, which is almost certain to be rescinded in the next administration. As a result, without Congressional action, the government bodies responsible for ushering into this new AI world will struggle. 

Fortunately, there are two bills on the brink of passage in Congress as we speak that would establish NAIRR and AISI permanently. They are the CREATE AI Act and the Future of AI Innovation Act. All that is left is for Congressional leadership to get it over the finish line. Meanwhile, in the New Year, a Republican Congress will be poised to decide funding for NIST, and are within their capabilities to do so if appropriate pressure is applied. If it’s true that we are going to be embracing new technologies to retain our competitive edge as an AI nation, we need to be well-staffed to forge confidently ahead.

It is imperative that our policies ensure thoughtful principles are implemented and the right people are asking smart follow-up questions for the next round, because AI is waiting for no one.