“We know how to build machines that do math (augmentation); we don’t know how to build machines that can read tax codes (replacement).”
Our reading, the sentence that was checked:
We do not currently know how to build machines that can read and understand tax codes well enough to replace accountants.
Assessment: Unresolved Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Three baffling claims about AI and machine learning in four days, statistical errors in top journals, and claims from Yann LeCun that you should not believe., paragraph 23.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-too-much-benefit-of-the-doubt-0b495a-a1
We recorded 13 claims from this post. See every claim from this post.
The wording above is the version this site observed on 2026-10-29. The source may have been edited after 2022-10-16. This page does not claim the wording was captured then.
What the claim commits to
- Kind
- Observation
- Deadline, as the speaker wrote it
- None stated. This site does not add one.
- Deadline state when assessed
- No horizon
- Deadline state on 2026-10-29
- No horizon (the release’s reference date)
From the speaker
- We know how to build machines that do math
- we don’t know how to build machines that can read tax codes
- replacement
Our interpretation
- In context, read tax codes means interpret and apply tax law reliably enough to perform an accountant's tax-judgment work, including deductibility decisions.
- It does not mean displaying text, recognizing characters, finding passages or executing rules already translated by people.
- The reference to replacement does not require automation of unrelated accounting duties.
- I use the current-year reporting window through publication for this ongoing capability claim, rather than a later model's capabilities.
Counts as supported if: Evidence covering the contemporary state of the field establishes that available systems lacked reliable autonomous interpretation and application of tax law at the stated replacement standard.
Counts as contradicted if: A system available by the post date demonstrates reliable autonomous tax-law interpretation and application across representative accountant judgment tasks, without case-by-case human translation of the governing rules.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Was it true for the period described?
Outcome: Unresolved
The record was checked, and the evidence available by the cutoff does not settle it.
Scored on a plain reading. The speaker named no measure for this claim, so it was judged on the reading an informed reader would take, stated below under our interpretation. Both reviews had to find that reading fair.
- Evidence as of
- Deadline state when assessed
- No horizon
- Assessment recorded
- Policy version
- 2026-10-28.7
Rationale
The [statutory-reasoning research](https://arxiv.org/abs/2105.07903) documents serious earlier machine-reading limitations. The [original benchmark](https://arxiv.org/abs/2005.05257) also shows why human-authored rule execution must be distinguished from reading tax law autonomously. These sources are consistent with Marcus's contextual claim, but they are earlier, restricted evaluations. They do not establish the field-wide October 2022 absence of replacement-level capability.
Limitations
- The cited tests are restricted tax-law tasks, not accountant replacement evaluations.
- Earlier failures cannot automatically be carried forward to the post date.
- The claim leaves jurisdiction and acceptable reliability unspecified.
Evidence (2 items)
Context only: arXiv,
nearly at-chance
Why it matters here: The sentence describes the earlier benchmark's weak neural baselines; the 2021 paper also reports improvements from added structure. These results motivate the concern but do not establish the capabilities of every relevant system in October 2022.
- Publisher
- arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: arXiv,
hand-constructed Prolog-based system
Why it matters here: A symbolic system could solve this restricted task after people encoded the law. This is material evidence against interpreting Marcus as denying all automated tax reasoning, but it does not meet the autonomous reading and replacement standard.
- Publisher
- arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
What would settle it
A contemporary capability survey or representative evaluation of systems available by October 16, 2022, testing autonomous tax-law judgment at a professional reliability standard.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Three baffling claims about AI and machine learning in four days, statistical errors in top journals, and claims from Yann LeCun that you should not believe.
We know how to build machines that do math (augmentation); we don’t know how to build machines that can read tax codes (replacement).
- Where
- paragraph 23
- Published
- 2022-10-16
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-too-much-benefit-of-the-doubt-0b495a-a1: Current. Outcome: Unresolved. Evidence as of 2026-10-28. Public date not recorded.
Recorded events
Assessment publishedBy this site
First assessment published after both reviews and the final-text check passed.
About assessment marcus-too-much-benefit-of-the-doubt-0b495a-a1.