Quick Answer: Can AI prepare Form 8621?
Can AI prepare Form 8621?
Where did AI fail?
Can I use AI instead of a PFIC calculator?
Why Testing AI on Form 8621 Matters
Many U.S. taxpayers now ask the same question: Can I use ChatGPT, Gemini or Claude to prepare Form 8621?
That question matters because PFIC reporting is not normal tax preparation. A small-looking foreign fund can trigger Form 8621, IRC §1291 excess distribution rules, historical tax-rate recomputation, §6621 interest, and detailed Part V reporting.
The risk is not that AI gives no answer. The risk is that AI gives a confident answer that looks organized but is wrong.
PFIC Benchmark Summary: ChatGPT vs Gemini vs Claude on Form 8621
ChatGPT and Gemini reached the right excess distribution and attempted daily holding-period allocation, but failed the Form 8621 Part V line mapping. Claude failed earlier by allocating the excess distribution evenly by year instead of by holding-period days.
| Test Item | ChatGPT | Gemini | Claude | Filing-Ready? |
|---|---|---|---|---|
| Identify 2024 distribution | Correct | Correct | Correct | Yes |
| Calculate prior-three-year average | Correct | Correct | Correct | Yes |
| Calculate 125% threshold | Correct | Correct | Correct | Yes |
| Identify excess distribution | Correct | Correct | Correct | Yes |
| Allocate excess by holding-period days | Mostly correct | Mostly correct | Incorrect | No |
| Treat partial-year 2019 and 2024 correctly | Mostly correct | Mostly correct | Incorrect | No |
| Apply true historical §6621 interest rates | No | No | No | No |
| Map Form 8621 Part V lines correctly | Incorrect | Incorrect | Not properly mapped | No |
| Produce filing-ready Form 8621 support | No | No | No | No |
The PFIC §1291 Test Case: Data and Setup
The dataset used one PFIC lot and annual distributions. All amounts are assumed to be in USD.
| Date | Event | Units | Amount |
|---|---|---|---|
| 2019-09-05 | Purchase | 1,000 | 10,000 |
| 2019-10-15 | Dividend | 0 | 70 |
| 2020-06-15 | Dividend | 0 | 80 |
| 2021-06-15 | Dividend | 0 | 90 |
| 2022-06-15 | Dividend | 0 | 100 |
| 2023-06-15 | Dividend | 0 | 110 |
| 2024-05-05 | Dividend | 0 | 2,000 |
The prior-three-year distributions and the 125% threshold calculation:
All three AI models got this part right. That is the easy part.
Correct PFIC §1291 Treatment: What AI Should Have Done
| Item | Correct Treatment |
|---|---|
| Total 2024 distribution | $2,000 |
| Non-excess portion | $125 |
| Excess distribution | $1,875 |
| Allocation method | By holding-period days, not by calendar years |
| 2019 treatment | Partial-year holding period (118 days) |
| 2024 treatment | Partial-year holding period (125–126 days) |
| Prior PFIC years | Subject to deferred tax and §6621 interest |
| Current-year allocation | Included in current-year income (Line 16b) |
| Form 8621 result | Requires correct Part V line mapping and Line 16a support |
Correct Day-Based Allocation Concept
A reasonable day-based allocation looks like this:
| Year | Holding-Period Days | Approx. Allocation |
|---|---|---|
| 2019 | 118 | About $130 |
| 2020 | 366 | About $403 |
| 2021 | 365 | About $402 |
| 2022 | 365 | About $402 |
| 2023 | 365 | About $402 |
| 2024 | 125–126 | About $138 |
| Total | 1,704 / 1,705 | $1,875 |
ChatGPT and Gemini reached this battlefield. Claude did not.
ChatGPT PFIC Test: Form 8621 Result
ChatGPT correctly identified the core §1291 excess distribution:
ChatGPT also attempted a day-based holding-period allocation:
| Year | Days | Allocated Excess |
|---|---|---|
| 2019 | 118 | 129.84 |
| 2020 | 366 | 402.73 |
| 2021 | 365 | 401.63 |
| 2022 | 365 | 401.63 |
| 2023 | 365 | 401.63 |
| 2024 | 125 | 137.54 |
| Total | 1,704 | 1,875.00 |
That was not bad. But ChatGPT failed the Form 8621 Part V mapping. It treated some allocation amounts as if they belonged on the wrong lines — in particular, it confused the prior-year allocation base with the Line 16c tax result. That is not filing-ready.
Gemini PFIC Test: Form 8621 Result
Gemini also found the excess distribution correctly and used a slightly different day count:
| Year | Days | Allocated Excess |
|---|---|---|
| 2019 | 118 | 129.77 |
| 2020 | 366 | 402.49 |
| 2021 | 365 | 401.39 |
| 2022 | 365 | 401.39 |
| 2023 | 365 | 401.39 |
| 2024 | 126 | 138.56 |
| Total | 1,705 | 1,875.00 |
The output looked clean and confident. That is the danger. Gemini also treated amounts as if they belonged on the wrong Form 8621 lines and did not produce a filing-ready Line 16a statement.
Claude PFIC Test: Form 8621 Result
Claude got the 125% test right. It correctly identified the $1,875 excess distribution. Then it made a more basic §1291 error.
| Claude Method | Result |
|---|---|
| Excess distribution | 1,875 |
| Number of years used | 6 |
| Allocation per year (Claude) | 312.50 |
Claude's answer was shorter and easier to read, but the calculation failed at a more fundamental level than ChatGPT and Gemini.
Three AI Models, Three Different PFIC Failure Patterns
| Failure Type | ChatGPT | Gemini | Claude |
|---|---|---|---|
| Got excess distribution ($1,875) | Yes | Yes | Yes |
| Understood day-based allocation | Mostly | Mostly | No |
| Preserved partial-year logic | Mostly | Mostly | No |
| Produced correct Part V mapping | No | No | No |
| Produced filing-ready output | No | No | No |
ChatGPT and Gemini got closer to the correct §1291 computation but failed the form. Claude gave a cleaner-looking answer but used the wrong allocation method. The result is the same: none of the three AI models produced a filing-ready Form 8621.
Correct Form 8621 Part V Line Mapping: What AI Missed
For this simplified case, the Form 8621 Part V structure should be understood as follows:
AI Error Table: Form 8621 Line-by-Line Failures
| Form 8621 Line | ChatGPT | Gemini | Claude |
|---|---|---|---|
| Line 15b | Placed excess distribution ($1,875) instead of prior-three-year total ($300) | Same error | Did not properly map |
| Line 15e(1)/(2) | Failed correct excess-distribution placement | Failed correct excess-distribution placement | Did not properly map |
| Line 16a | Gave a numeric allocation amount, not a proper support statement | Same issue | Did not produce proper statement |
| Line 16b | Mixed current-year allocation with non-excess distribution | Same issue | Used wrong equal-year annual allocation |
| Line 16c | Confused allocation base with aggregate tax increase | Same issue | Did not reach correct computation |
| Line 16f | Used rough interest estimate | Used rough interest estimate | Did not compute accurately |
Why AI PFIC Errors Matter for U.S. Expats and Tax Preparers
A taxpayer does not file a tax explanation. A taxpayer files Form 8621. That form needs line control.
This matters for U.S. persons holding foreign funds, including:
- KiwiSaver and PIE funds
- UCITS ETFs and OEICs
- Canadian mutual funds and ETFs
- Australian managed funds
- Israeli provident and investment funds
- Korean ETFs
- Foreign robo-advisor portfolios
- Non-U.S. mutual funds held through local brokers
Many of these assets look ordinary in the local country. Under U.S. tax rules, they may be PFICs. If the taxpayer uses AI to calculate Form 8621 and the model puts the right number on the wrong line, the filing can still be defective.
Why AI Struggles With PFIC and Form 8621
PFIC is not hard because the words are hard. PFIC is hard because the calculation is mechanical.
A real Form 8621 calculation may require:
- PFIC event identification
- Per-lot tracking and FIFO matching
- Daily holding-period allocation
- Current-year vs prior-year split
- Historical highest ordinary tax rates
- §6621 underpayment interest with daily compounding
- Foreign-currency conversion
- Form 8621 Part V line mapping
- Line 16a statement support
- Reconciliation between workpaper and form
AI can describe these concepts. That is not the same as enforcing them. A PFIC calculator cannot guess. It must apply fixed rules to fixed data.
What This AI Form 8621 Benchmark Proves — and Does Not Prove
Three facts established
AI can identify a simple PFIC excess distribution. AI may produce plausible §1291 calculations that still fail Form 8621. Different AI models fail differently.
Risk: Confidence, not silenceReal-world complexity
Multiple lots, partial sales, FIFO matching, foreign currency, MTM/QEF elections, true quarterly §6621 interest, prior-year amendments, missing historical data.
Real cases are harder
FAQ: AI, ChatGPT, Gemini, Claude, PFIC and Form 8621
Can AI prepare Form 8621?
Can ChatGPT calculate PFIC tax?
Is Gemini accurate for Form 8621?
Is Claude accurate for PFIC calculations?
Can AI calculate a PFIC §1291 excess distribution?
What did AI get wrong on Form 8621?
Can AI handle Form 8621 Line 15e?
Can AI prepare Form 8621 Line 16a?
Can AI calculate §6621 interest for PFIC?
Should I use ChatGPT, Gemini or Claude for PFIC tax filing?
Official Sources and References
- 🔗 IRS Form 8621: Information Return by a Shareholder of a Passive Foreign Investment Company or Qualified Electing Fund
- 🔗 Instructions for Form 8621: Official IRS instructions for completing Form 8621.
- 🔗 IRS Quarterly Interest Rates: Official quarterly interest rates under IRC §6621.
Recommended Reading
- 🔗 §1291 Excess Distribution Calculation Guide
- 🔗 Form 8621 Line 15e Under Currency Rules
- 🔗 Form 8621 Line 16a Statement: Workpaper Format
- 🔗 PFIC Calculator vs Excel: Why Spreadsheets Fail
- 🔗 TurboTax Form 8621 Guide
Current as of May 2026 · Based on Form 8621 (Rev. 12/2025)