It’s in the mail
Elsewhere I weighed in on the widening conversation on Congressional modernization, noting some institutional barriers to 21st century legislating. But I omitted a first-hand anecdote that’s kind of an extreme encapsulation of the problem.
The problem, briefly: Congress is slow to update processes and technologies to maximize effectiveness.
An example: When the Executive Branch sends official communications to the Legislative Branch, the recipient has to attach an official received date. For standard recordkeeping hygiene but also because that date can meaningfully affect legislative options. Like the Congressional Review Act (CRA) which permits Congress a fast-track mechanism to repeal regulations. It’s cabined by a deadline attached to the reg’s “received by Congress”1 date which in practice turns out isn’t what you think.2
Typically agencies submit the required info separately to the House and Senate, electronically and by mail. As you can imagine the latter is much slower and involves multiple steps.3 Weeks could elapse between submission and delivery. Even physical delivery didn’t constitute “received by Congress.” It would be only when a staffer actually opened the envelope AND entered it into a database. Only the hardcopy counted.
It means the window for Congress to act under the CRA came down to the moment a staffer decided to go through the mail.
And it was a different person for each chamber. Yeesh.
We had to find a quicker way and a pandemic didn’t help. In 2020 what was a lethargic pace became catatonic. For the biggest items, very diligent agencies sent a courier, which still didn’t solve the open/log lag, but it helped. Congress’s early COVID protocols prevented non-staff entering the building. No way for offices to sign for delivery.
So I called the relevant offices and asked: “Since we already submit by email anyway, can we treat that timestamp as operative?” The discussion that ensued raised issues both legitimate and tedious. But the answer was “no” and trust me the reasons were not ones you’d find compelling.4 It boiled down to not having the adequate “systems” in place to do that.
In the end I advised - as there was no instructive adjudication on this - agencies to attempt delivery by courier for anything sufficiently important (entry restrictions were softening by this point) and we would use the signed handoff date as a record for any potential challenge.
The point is Congressional modernization, updating it to modern times, is absolutely needed but you probably have no idea how deep the necessity goes.
Human-in-the-loop (redux)
In March the always excellent ChinaTalk podcast (Jordan Schneider) touched on the at-the-time surprise DoW smackdown of Anthropic. I commend the discussants, and without intending to but at risk of being precious, there was one frustration.
One guest asserted in the absence of strong mutual alignment there are “laws” governing defense use of force that would be applicable and binding to AI. No one corrected that framing, including when the subject of direct experience developing DoW’s “rules” came up. This distinction is important.
I understand them as referring to the combination of then-DoD directive(s) and the JAIC-developed internal AI guidance issued by Sec. Esper in 2020. The point is the lack of actual “law” (strictly defined, statute) makes navigating novel policy questions that much harder for both parties.
There’s little to no expectation of stability in the rule regime. I fully expect the Esper memo to soon go conspicuously trashed.
Truly novel questions emerge, by the week, which have no guiding principles on which to build an answer.
Anthropic is not blameless here but disputes like this were inevitable in such an impossible policy environment. I’ll go strong buy on the “it’ll happen again” futures contract.
Couple years ago I was trying to tease out, at a broad level, what a then-House Committee Chairman thought would be an acceptable AI governance framework. He paused and said “Honestly? Y’know the Esper memo? I’d be just fine with codifying that into law. Let’s start there.”5
Now years later, I’ll still posit that impulse is correct. And continued failure to “start” somewhere isn’t doing anyone a favor, not AI labs or more importantly our defense capabilities.
Also note a compelling rejoinder by guest Henry Farrell (“AI is a bureaucratic technology. So is fighting war”).
Citing a different theater of war, an AI targeting system exhibited something like a 10 percent Type I error rate with little reconsideration by “human-in-the-loop.” Remarkably close to the rhetorical example Brad Carson gave arguing such an oversight heuristic is insufficient. I still believe it’s a necessary condition, but certainly not sufficient.
To the degree you do worry about scaled surveillance capabilities, the dilemma is real. Reminded me of the Census Bureau led study on re-identifying otherwise “de-identified” public data. Using what to mere mortals like us looks like totally impersonal Census elements, sophisticated data-jockeys could re-assemble private identities. Years before ChatGPT.
Don’t Believe Me, Just Watch
The Dispatch reported on the effort to put cameras in courtrooms. One bill requiring them in the Supreme Court and another to make it easier in lower courts unanimously (v/v) passed out of committees.6 Of course the high court maintains consistent institutional opposition. (In 1996 Justice Souter offered no quarter to such an idea, and also maybe didn’t know what a camera is)
The intuition is understandable, and not surprising the committee vote tracks popular opinion as the premise itself is democratic. The arguments in favor are also seductively compelling. On net, intuition says it’s been negative, by a a small-to-moderate amount [medium confidence].
The principle-based justification for cameras in Congress, say, is not conceptually wrong: accountability is monotonically increasing in transparency. Being watched leads to better behavior, ceteris paribus. But empirically that hasn’t been the outcome. Yuval Levin tells the story.
One advocate argues visual information is not adequately proxied by other means. But in my experience the structure of policy decision-making is predominantly written modes with some verbal communication.7
Worth asking: Is visual information extraneous? Biasing?
People had eyes at the country’s founding - what were the rules then, if only implied? Okay “camera” meant something else altogether, but how did they account for the role of visual information?
To be fair no one could have extrapolated the proliferation of social media and other tech, much less Congress’s endogenous behavioral changes (i.e., preening, confrontation, “clip-art,” etc.). It may have been nonetheless unavoidable as committees are entitled to their own rules and nothing would stop members/staff from self-broadcasting.
The OJ Simpsons trial broadcast…granted, a 1-of-1, but can we say it had no effect?
The Supreme Court has thus far managed to bar the door, inoculating it from the symptoms of this particular bug going ‘round. But immunity fades and it only takes a small viral load to incubate into illness.
Etc.
Law Firm of AI & AI: David Lat & Josh Barro on Serious Trouble got into attorney wages as a function of AI. Great discussion and insider-y understanding of some law firm idiosyncrasies, but seemed like they were unable to come to a mutually-agreeable model. Fairly simple comparative statics (conditional on your assumptions):
If you believe AI increases productivity of legal work (broadly), then reduces marginal cost. And almost all legal work is adjustable at the extensive margin.8
For “low value”/standard tasks: shadow price of marginal billed hour goes down, elasticity > 1 → aggregate demand increases while revenue per matter decreases.
Ex: legal support services increased after e-discovery.
For “superstar” segment (high value/risk matters, M&A): lower cost mostly captured as higher firm margin. Increases return to reputational capital + specialized expertise + management. Revenues flat to slightly up.
Three basic billing models: hourly, fixed retainer, contingency. If you think there will adjustment in payment models → clients open more matters.
Solve for x = the number of justices: A fascinating prompt on Advisory Opinions’ Supreme Court term roundup. Sarah Isgur asks: what is the correct methodological lens for assessing the court? Presented as implicitly unbounded. Each panelist in turn ably grabbed the ball and went ISO for a couple possessions.
Akhil Reed Amar went trick play. He invoked Condorcet’s Jury Theorem and the Wisdom of Crowds maxim to cheekily suggest expanding the court would improve it’s quality, and therefore he follows the median judge as a proxy for how the court is doing.9
But he forgot to mention the one super critical necessary condition: the population average from which the deciders are drawn have to be ≥ 50% correct, accurate, smart, whatever. Because when it’s not true the formula predicts as size increases accuracy collapses to 0.
5 USC 802(a): “…the date on which the report [the regulation] is received by Congress and ending 60 days thereafter…”
Yes I’m eliding technical details but the point will stand. But I’ll bet an official U.S. Senate Coaster Gift Set I know more about this than you.
After 9/11, all Congressional postal mail (even from federal offices) goes through an off-site sorting and security screening facility before delivery to the Capitol.
I don’t want to betray the other side of a private conversation, and in all fairness everyone was doing what they could within their constraints. Really it’s not an indictment of anyone in particular as it is of Congress as an institution.
This was an otherwise anti-regulation, GOP type in a private conversation.
For purely aesthetic reasons, I kinda favor the old school nature of the audio+photo the USSC has now. It actually makes it easier to focus on the presentation. It also gives it a local access-adjacent quality I dig.
I’m separating out the large task of negotiation occupying much of the policymaking production function. But this is non-public by design and cameras are immaterial.
Based on my time in an AmLaw100 firm, mix of hourly billing and uncapped retainer (allowances for overages bc all work is hourly tracked). Every client regardless of pay arrangement has one or more potential matters not yet undertaken—many small add-ons or discrete research projects etc.
Formally it applies to jury size (obviously) but any group decision-making over a binary choice—most general elections and legislatures, HOA associations, bank loan officers…some AI training methods.