Comment by SamoyedFurFluff
7 days ago
The thing that makes me nervous is the statement that they plan to use AI. AI? The thing that is mathematically incapable of perfection, on finance information, for which perfection is table stakes? Not to mention all the privacy issues (although that boat has sailed).
The people in charge have a pathological hatred for the IRS. AI is just an excuse to continue destroying the capabilities of the IRS. In the meantime, they’ll keep borrowing to fund the government while telling everyone it’s ok because they slashed programs that make up a tiny portion of the budget. This can go on until there is a major economic shock related to US debt, but honestly, most of them will be dead by the time that happens.
I thought I would give the Treasury the benefit of the doubt for a moment and check whether they meant LLMs like we're all assuming, or possibly a more specific finance-focused type of AI. Like how we have specialist neural net AI helping with radiology.
Looking at their official info document[1]... "a secure AI-based chat solution"... "AI-assisted code development"...
Okay they mean LLMs, carry on.
[1] https://home.treasury.gov/system/files/136/Treasury-AI-Strat...
I am not following, a lot of things get turned into python calculations, so the LLM is not doing the precise math.
There are two forces at work:
1. Rich cheats for whom complexity is the goal. Reduced enforcement benefits them without the guilt. They can construct nonsensical schemes but if no one ever audits them they get to feel like they are paying what they owe despite being freeloaders.
2. Strangle the baby types: they hate the federal government. They deliberately want to reduce its income to force cuts to government spending (programs and staff). If they can they will cut other parts of the government then use that to justify reducing taxes. Nothing else matters except shrinking the federal government as much as possible via any means possible. These types also enjoy taking any government service that works and people like and making it as terrible as possible to kill popular support thus making it easier to cut the program entirely.
If you think the state of the average tax return is "perfection as table stakes"..... you might be disappointed.
Most tax returns these days are prepared and submitted electronically so the basic work of the arithmetic involved should be as close as possible to perfect already. Evaluating that is going to be pretty mathematically intensive though and LLMs have been pretty bad at that. Tool usage has gotten it better so maybe they'll just hand off the validation to the existing traditional computing and mostly be vibes based, 'does this return look legit?' evaluation.