Of course, there’s a lot of speculation about deploying AI to replace federal employees for routine administrative tasks, and to make the remaining ones more efficient on those same tasks. Pretty obvious, I would think.
But what about the potential impact of AI on federal policy making? Not at all routine, often very complex, a lot of written output, almost always done under deadlines. Sounds a lot like university term papers, something which AI LLM models are apparently pretty good at. Will AI have an impact on federal policymaking?
It’s hard not to believe that AI LLM models are already in extensive (though perhaps unofficial) use by government policy staffers, not to mention think tanks, etc. For sure at DOGE — a house specialty, as it were.
So, as part of my ongoing exploration of AI using stuff I know about, I had a long-ish ‘discussion’ with the free version of ChatGPT about the WIFIA Loan Program. The unedited output is above.
Here’s my overall impression: Simple questions get ‘mainstream narrative’ answers because presumably that forms the bulk of the model’s database inputs. But if you push with harder, devil’s advocate-type questions, there’s much more subtlety and substance than I would have expected — there appears to be some synthesis of disparate sources and active logic going on, and all expressed in eerily smooth and cogent language.
(Btw, as the model ‘contemplates’ and slowly scrolls out an answer, you get a visceral sense of physical effort: billions of chips firing away, heat rising from the frames, electricity pouring in and heated cooling water pouring out. For one question on a free model. Now…how about that essential physical infrastructure again?)
It takes some knowledge to push the model into a deeper and logically sequential ‘exposition’ — you need to know what questions to ask. But in a policy-saturated environment, the staffer using the AI will probably have more than enough, just from their everyday general conversation. The purpose, cost and benefits of federal programs must be discussed all the time, one assumes (hopes?), even if little of those discussion ever becomes public. The federal programs are their ‘product’, and the day job is to put much thought into defending them — and attacking competing ‘products’. In that context, sophisticated users pushing AI to deliver deeper and more nuanced answers on policy issues, for defense, offense, or ‘war games’ will (or more likely, has) become commonplace.
The net result might be this: AI effectively lowers the cost and increases the efficiency of deep dives into policy matters. If there are equally sophisticated users on both sides of the issue (likely?), the whole policy discussion will likely have more than depth and nuance than before. Whether it’s ‘improved’ with respect to the public good might be another question (in one sense, AI ‘improves’ armed conflict — cui bono there?) and of course, everything will continue to be processed and packaged for public consumption, as always. Maybe it’s only safe to say for now that policy debate could become more interesting for dedicated and knowledgeable fans — a game played at a higher intellectual level, in effect.
The impact in this way on federal financing programs might be particularly significant. That’s because the combination of government policy and finance is intrinsically complex and multi-faceted — there’s a lot going on — and AI tools could be especially effective in efficiently digging out the nuances and synergies that otherwise wouldn’t be considered, e.g., by a hard-pressed staffer toiling away at 2 a.m. for a hearing the next day.
Well, we’ll see. In the meantime, below are the questions (with page numbers) I asked ChatGPT in the course of our ‘discussion’, starting from basic purpose and going to economic rationale, some political context, relationship to state & local lending funds and actual cost. Regarding the last one, you can see that I drilled the model pretty hard on a very specialized area, FCRA interest rate re-estimates. Its analysis and answers on that were surprisingly cogent. I’m not sure what to think about that yet.
Program Additionality
1. The WIFIA Loan Program lends to investment-grade public water agencies financing low-risk projects. But these agencies have many financial options, including access to the tax-exempt municipal bond market where interest rates are often as low as those offered by WIFIA. What purpose does WIFIA serve? Is the Program necessary? (page 1)
2. Does WIFIA compete with the municipal bond market? (page 3)
3. Is WIFIA necessary in terms of strict additionality? That is, given typical borrowers’ other finance alternatives, does it make any difference in terms of water infrastructure? (page 5)
4. Imagine you are trying to downsize and decentralize the federal government. Would WIFIA be a good candidate for elimination? Can state and other local governments perform the same function? (page 7)
Program Loans Based on Federal Financing Strengths
5. Does the federal government have any unique capabilities, relative to the debt markets or local governments, in providing infrastructure finance? Exclude transfer payment or loss absorption capability based on US scale. (page 9)
6. Are interjurisdictional coordination, mission-driven and regulatory leverage really financing strengths? Or just policy add-ons or riders that the federal government is enforcing by offering attractive financing terms? That is, offering a ‘carrot’ (the benefit) to get borrowers to accept the ‘stick’ (policy riders that they would not have otherwise done)? (page 12)
Program in Political Context
7. Is [WIFIA] ‘politically attractive’ to the [Biden] Administration? (page 14)
8. Try again. To the Trump Administration? (page 15)
9. How might DOGE people look at WIFIA Program? Are they likely to have a negative view? (page 17)
10. Could WIFIA have an essential role in water infrastructure finance based on federal functional financing strengths? For example, loan features like a very long term that can facilitate local funding for local projects. (page 18)
Program as a ‘Wholesale’ Lender
11. Would it also make sense for WIFIA to utilize unique functional strengths in lending to SRFs and other state & local infrastructure financing agencies? The rough analogy would be a ‘wholesale lender’ (WIFIA) to other local ‘retail lenders’ (state & local agencies). Could this a unique role? (page 20)
12. Would WIFIA make a good ‘retail lender’ (e.g. direct loans to many small projects throughout the country)? Why not? (page 23)
Program’s Actual Taxpayer Cost
13. A WIFIA loan has a ‘rate lock’ — the interest rate is locked at loan commitment, but loan drawdown may occur much later. Since cancelling the commitment if rates fall does not incur a penalty, doesn’t the rate lock work as an interest rate option? (page 24)
14. Per FCRA, the cost of a WIFIA loan to the taxpayers is estimated when the loan is drawn. If US Treasury rates have risen since loan commitment, the cost will exceed the discretionary appropriation allocated to the loan. The additional cost of such interest rates estimates will become an off-budget mandatory appropriation. Can you estimate the current scale of WIFIA’s off-budget mandatory appropriations? (page 26)
15. WIFIA borrowers can pre-pay their loans anytime without penalty. If interest rates fall, borrowers will refinance their loans. If rates rise, they will not. Over time, does this mean that the WIFIA loan portfolio will increasingly be concentrated in loans with low interest rates? (page 27)
16. Doesn’t this imply that the WIFIA Program is far more expensive than its discretionary appropriations would indicate? (page 29)
17. If DOGE conducts this kind of financial analysis on WIFIA, and estimates the true scale of Program cost, are they likely to recommend cutting it? (page 31)