Showing an AI More of the Tax Code Made It Worse: A 400-Generation Test
Four local AI models were given the same tax questions with three brackets, then five. The best model went from a $100 error to a $132,000 one — because it taxes every bracket it is shown, including ones your income never reaches.
On this evidence, no. Four small open models ran 400 progressive-tax calculations between them and three scored zero out of 100. The best managed 51%. Every model failed the questions requiring several brackets to be summed, which is what an actual tax calculation is.
Get new posts in your inbox
No spam, unsubscribe anytime.
About the Author
Sujit KarkiFinance Researcher & Market Analyst
Independent finance researcher and market analyst with expertise in macroeconomics, equity markets, and personal finance. I help regular investors make better-informed decisions through rigorous, data-driven analysis.
Four small AI models ran 400 progressive-tax calculations. Three scored exactly zero out of 100.
The same model, asked the identical question, answered $6,207.50 when shown three tax brackets and $138,832 when shown five. The true answer is $6,307.50.
The failure is not arithmetic. Models apply every bracket they are shown, taxing income the taxpayer does not have.
One model computed 50,000 minus 182,100 as 317,900 in order to tax a bracket that did not apply.
Only phi3.5:3.8b scored above zero, at 51% — and it was the slowest model tested.
A previous test on this site found that small
AI models running on a laptop failed simple money arithmetic. One tax question
appeared in it, with three brackets, and every model got it wrong. This is the
follow-up: ten tax questions, four models, ten repetitions each — 400 answers in
total, all generated locally on a GTX 1650 Ti.
The result was worse than expected, and the reason turned out to be something I
was not testing for.
Three of four scored zero
Not "low accuracy". Zero. Three models produced one hundred answers each and
none of them was right.
The result that matters most is qwen2.5:3b scoring nothing, because on the
earlier and simpler benchmark it was the best model tested, at 75%. Whatever
happened here is not a matter of one weak model.
The controlled comparison
Both tests asked the same question: what is the total tax owed on $50,000? The
true answer is $6,307.50. The only difference was how many brackets the
prompt listed.
qwen2.5:3b, identical question, different number of brackets shown
Brackets in the prompt
Answer
Error
3
$6,207.50
$100 low
5
$138,832.00
$132,524.50 high
The two extra brackets covered income from $95,375 to $182,100 and above
$182,100. A person earning $50,000 never reaches either. They are irrelevant
information — and adding them multiplied the answer by twenty-two.
Warning
This is the opposite of the usual advice about prompting. The instinct is that
giving a model more of the rule book helps it. Here, every additional rule was
another thing the model applied whether or not it should have.
What they actually do wrong
The models show their working, so the failure is visible rather than inferred.
This is llama3.2:3b on the $50,000 question, unedited:
1. Income from 0 to 11000: 11000 * 0.10 = 1100
2. Income from 11000 to 44725: 33725 * 0.12 = 4042
3. Income from 44725 to 95375: 50650 * 0.22 = 11122
4. Income from 95375 to 182100: 156250 * 0.24 = 37450
5. Income above 182100: 31750 * 0.32 = 10156
Now, let's add up the taxes:
1100 + 4042 = 5142
5142 + 11122 = 16264
16264 + 37450 = 53714
53714 + 10156 = 63870
ANSWER: 63870
Read line 4. It taxes $156,250 of income. The taxpayer earns $50,000. The figure
$50,000 appears nowhere after the first line — the model computed the full width
of every bracket and added them all up, producing a tax bill larger than the
income it was calculating tax on.
qwen2.5:3b does the same thing and adds an arithmetic invention to reach it:
5. Income above 182100:
- Amount: 50000 - 182100 = 317900
- Tax: 317900 x 0.32 = 101728
$50,000 minus $182,100 is negative $132,100. The model returned positive
$317,900 — wrong in sign and magnitude — apparently because a negative number
would have made the step obviously impossible, and the step had to happen.
The same pattern held on a $9,000 income. qwen2.5:3b correctly noted that the
second bracket contributed $0, then taxed the $44,725–$95,375 and
$95,375–$182,100 bands anyway, arriving at $16,190 of tax on $9,000 of income.
The models are not bad at multiplication. They are treating a bracket table as
a list of things to compute rather than a set of conditions to test.
The one model that half worked
phi3.5:3.8b scored 51%, and how it splits is the clearest signal in the
dataset:
phi3.5:3.8b by question type
Question
Correct
Steps required
Tax on $9,000 (one bracket)
10/10
One multiplication
Marginal rate at $75,000
10/10
Identify the band
Effective rate at $50,000
10/10
One division
Tax on $120,000
10/10
Sum of brackets
Tax on $200,000
10/10
Sum of brackets
Tax on $50,000
0/10
Sum of brackets
Tax on $50,000 after a deduction
0/10
Deduct, then sum
Tax at the $95,375 boundary
0/10
Exact boundary case
It is perfect on anything resolvable in one or two steps and unreliable once
several brackets must be combined. That threshold is what makes the whole test
credible: a broken test bank would produce scattered failures, not a clean line
between simple and compound questions.
It is also worth noting that phi3.5 was the slowest model tested, at 22.5
tokens per second against roughly 50 for the others. The only model that got
anything right was the one that took twice as long to be wrong less often.
What this means in practice
The practical rule is narrow and worth stating plainly.
Do not ask a local model to compute a tax figure you intend to act on. Not
because it might be slightly off — three of four models here were not slightly
off, they were catastrophically off, and the one that worked failed the most
realistic question in the set.
And be careful about the assumption that more context is safer. Pasting the
full bracket table, the whole policy document, every rule you can find, feels
like diligence. On this evidence it can make the answer dramatically worse,
because the model has no reliable mechanism for deciding which of the rules you
gave it do not apply.
What these models remain genuinely useful for is explanation. Ask one why a
marginal rate differs from an effective rate and you will get a clear, correct
answer. The line is between explaining a rule and applying it to your
numbers, and it sits further left than most people assume.
Warning
This is educational information, not tax advice. The brackets used throughout
are illustrative figures supplied inside each prompt, deliberately not any real
tax year's, so the test measures arithmetic rather than recall. Consult a
licensed professional before acting on any tax decision.
Methodology
Everything here comes from a dataset generated for this post. The aggregated
results are published with it as tax-accuracy.csv. The
harness and the raw generations are in research/; the
run reproduces on any machine with Ollama.
Hardware. GTX 1650 Ti, 4 GB VRAM, 8 GB system RAM, Ollama 0.33.2. All four
models ran fully on the GPU.
Models.phi3.5:3.8b, qwen2.5:3b, llama3.2:3b, gemma2:2b, all Q4_K_M —
the default weights ollama pull fetches.
Scenarios. Ten questions spanning single-bracket cases, multi-bracket sums,
a deduction applied before the brackets, an exact bracket boundary, effective
rate and marginal rate. Each ran ten times with a different seed, at
temperature 0 and explicit num_ctx, with one warm-up generation per model
discarded.
Ground truth is generated, not typed. The bracket structure is defined once
in code; the prompt text and the expected answer are both derived from it. The
correct answer therefore cannot disagree with the question asked. Hand-entering
expected values was the largest error risk in the previous benchmark and this
removes it.
Determinism. 38 of the 40 model-question pairs returned an identical answer
across all ten repetitions. Asking again does not surface these errors.
Truncation, disclosed. Nine generations hit the 1,024-token output cap, all
of them phi3.5 on a single scenario. None would have been correct regardless;
excluding them entirely, phi3.5 scores 51/91, or 56%. The headline 51% is the
conservative figure.
A limit worth stating. This is four small models on one machine with one
bracket structure. It does not establish how larger or hosted models behave, and
it should not be read as doing so. What it does show is a specific, reproducible
mechanism — applying every rule presented, regardless of applicability — that is
visible in the models' own working rather than inferred from a score.
Disclosure. The models are the subject being measured, not the author. None
of them wrote any part of this post. The commentary was drafted with AI
assistance; the experiment design, the ground-truth computation and the
conclusions are mine.
Sujit Karki||6 min read