Hook
Four months. That is the entire window. Epoch AI reportedly found that AI acknowledgments in mathematics papers jumped to 25% within four months, and the trade press repeated it as proof that academic norms are shifting. No baseline. No sample size. No coding rule. No link to the primary report. A number, a timeframe, and a conclusion.
I have audited signals like this before, and the pattern never changes. In August 2020, while finishing my MS in Economics, I found an integer overflow in the early Compound Finance governance module. I did not wait for someone else to surface it. I compiled a standardized bug-bounty submission, modeled the exploit path from first principles, and filed it to the protocol's GitHub. It returned a $5,000 bounty and a formal acknowledgment. The lesson was never about the bug. It was about the discipline: a number without a methodology is not data. It is a narrative with a decimal point attached.
A 25% acknowledgment rate does not measure AI's role in mathematics. It measures how many authors typed a sentence. Those are different objects. Confusing them is the most expensive error available in this story, and most of the coverage has already made it.
Context: What "AI in Math" Actually Means
Start with the structure of the claim. The signal is not a new model architecture. It is not a benchmark score. It is a self-reported behavior metric โ the frequency with which authors mention an AI tool in the acknowledgments section of a paper. Epoch AI, for those who do not track the space, positions itself as a provider of AI trend data and benchmarks: training compute curves, frontier capability trackers, and evaluation frameworks. Its discoveries function as content products. When it publishes a number, the number becomes a citation, and the citation becomes authority.
That is a legitimate operation. It is also a machine that runs on narrative selection. Choosing to publish "25% in four months" is an editorial decision, and editorial decisions carry incentives. None of that makes the number false. It makes the number incomplete, and incomplete data is the most dangerous kind because it survives casual scrutiny.
The real subject here is not mathematics. It is infrastructure. AI has moved from a research tool to a layer inside the research pipeline โ literature search, LaTeX polishing, translation, code generation, proof-strategy exploration, and, at the frontier, formal verification inside Lean, Isabelle, and Coq. These are not the same activity. They sit at completely different levels of technical maturity, reliability, and verifiability. The single number flattens all of them into one bucket labeled "AI involvement."
I run the same flattening test on DeFi dashboards every day. A protocol reports $800 million in total value locked. The dashboard shows one number. The number does not distinguish between organic deposits, mercenary capital chasing a subsidy, and recursive leverage looping the same dollar three times. The label says "TVL." The reality is a mixture of three incompatible things. When you treat the aggregate as a single signal, you are not analyzing the protocol. You are analyzing the marketing.
The 25% figure has the same defect. It is a headline aggregate over a heterogeneous population, and the heterogeneity is the only part that matters.
Core: Auditing the Signal Layer by Layer
Layer One โ The Tool Layer
This is where the vast majority of the 25% almost certainly lives. An author uses an LLM to polish English prose, translate a draft, format references, or generate a plotting script. The cognitive contribution remains human. The AI is a word processor with a better autocomplete. Adoption here is near production-grade because the cost of failure is trivial and the time savings are immediate.
Acknowledging this costs one sentence. That asymmetry โ trivial cost to acknowledge, meaningful credit to signal compliance โ is the engine behind the number.
Layer Two โ The Assistant Layer
The author uses an LLM to explore proof strategies, construct counterexamples, or search for a known result they suspect exists. This is the proof-of-concept-to-production transition zone. Terence Tao has publicly described using AI assistants this way. The output is not trusted directly; it is treated as a candidate that a human then verifies. Reliability is real but bounded. This layer is growing fast and is genuinely interesting.
Layer Three โ The Formal Layer
This is where AI generates a contribution that can be machine-checked inside a proof assistant. Systems in the AlphaProof and AlphaGeometry class reached silver-medal-adjacent performance on olympiad problems in 2024. The Lean and mathlib ecosystem is heavily open source. This is the only layer where the phrase "AI contributed mathematics" is literally verifiable, because a proof either compiles or it does not.
Now run the arithmetic. Acknowledgments are dominated by Layer One. The number is being cited as if it described Layer Three. That is not a rounding error. It is a category error, and it is the core failure of the entire story.
The Disclosure Rate Is Not the Usage Rate
Here is the second audit finding, and it is the one the coverage missed completely. The 25% is a disclosure rate, not a usage rate. Those two quantities move for different reasons.
A disclosure rate rises when a journal or preprint server mandates a disclosure. Springer Nature, arXiv, and most major publishers have layered AI-use policies onto authors since 2023. A four-month spike is exactly the shape you would expect from a policy change followed by author compliance โ authors retroactively adding the sentence because the submission form now has a field for it. A usage rate spike would require AI adoption itself to multiply several-fold in four months. That is not what the curve describes.
I watched the identical confusion destroy people in the Terra collapse of May 2022. When the anchor yield on Anchor broke, the dashboard still showed a deposit number. The deposit number was real. The meaning of the deposit number had already inverted โ what looked like sticky liquidity was exit liquidity waiting for a queue slot. The number did not change. The label did. Readers who trusted the label instead of the mechanism lost everything.
A disclosure rate that rises because the disclosure form appeared is a compliance metric, not an adoption metric. Treating the two as the same variable is the analytical equivalent of marking your own homework.
Verifiability Inverts the Headline
Mathematics is the worst possible field to make this error in, because mathematics is where AI hallucination is most easily exposed. A fabricated proof does not survive a competent referee for an afternoon. This creates a counterintuitive result: the ratio of AI-generated mathematical content that is actually accepted into the literature is very likely far below the ratio of AI acknowledgments. The gap between the two is the honest measure of AI's real contribution, and it is a gap the headline erases.
Efficiency is the only honest validator. In a proof assistant, efficiency is literal: the proof compiles or the session fails. In an acknowledgments section, there is no validator at all. You are reading a self-report with no third-party audit, no coding rule disclosed, and no way to distinguish "thanked ChatGPT for grammar" from "used a reasoning model to find the key lemma."
The methodology questions that would make this number usable are not hard to list. What is the operational definition of an AI acknowledgment โ keyword match or human annotation? What is the baseline โ 0% to 25%, or 20% to 25%? What is the sample size โ dozens of papers or thousands? Is the corpus all of arXiv, a subfield, or peer-reviewed journals only? Does it separate AI-generated content that was adopted from AI used on non-core tasks? Every one of those answers changes the interpretation. None of them are in the report as circulated.
The Same Inflation Mechanism as Liquidity Mining
Here is where the crypto reader should lean in, because the pattern is not foreign. It is our own. Liquidity mining APY is not a measure of organic demand. It is a measure of how much a protocol is willing to pay to manufacture a TVL number. Cut the subsidy, and the number collapses, because the number was never demand โ it was incentive response wearing demand's clothes.
The AI acknowledgment rate behaves the same way. Mandate the disclosure, and the number jumps, because the number was never adoption โ it was compliance response wearing adoption's clothes. Both metrics are subsidies dressed as signals. Both are read by people who want the headline to mean something larger than the mechanism supports. Liquidities trapped in code, not in trust โ and the acknowledgment line is code that reports whatever the author chooses to write.
Who Benefits From the Ambiguity
Follow the incentive. A research organization that publishes trend data benefits when its trend data gets cited. A number like "25% in four months" is more citable than "25% with a 0% baseline and unclear coding rules." A crypto media outlet that has no Web3 angle on an academic AI story benefits from running it anyway, because AI is the traffic magnet of the cycle and the aggregator needs volume. Neither party needs the number to be wrong. They need it to be dramatic. Dramatic and wrong are different failure modes, but they produce the same reader.
Standard-setting is the prize underneath. Whoever defines "AI contribution" defines the evaluation market that follows โ the audit tools, the compliance logs, the citation-credit frameworks. That is a real commercial surface, and it is currently empty. The first credible party to publish a defensible methodology captures the default. I have watched this exact dynamic play out in the AI-agent trading space, where I built a standardized protocol for agents to interact with DeFi and published it as a whitepaper on automated compliance. Standards are not documentation. They are market structure. Whoever writes the schema writes the rules everyone else implements against.
The Three-Layer Maturity Map
| Capability | Representative Progress | Maturity | |---|---|---| | Math text reasoning (contest problems) | Frontier LLMs at olympiad level | Production-scale | | Formal proof (Lean, Isabelle) | Neural theorem provers | Research to POC | | Conjecture and counterexample search | LLM-assisted, human-verified | POC | | Writing, literature, code support | General-purpose LLMs | Full-scale |
The 25% number collapses all four rows into one. Three of the four rows are not mathematics. One of them is the only row that can be verified. The headline describes the bottom row and implies the second.
Contrarian: The Number Is Probably a Floor, Not a Ceiling
Everyone is debating whether 25% overstates AI's role in mathematics. They are arguing about the wrong direction. The more defensible claim is that 25% understates AI usage and overstates AI contribution โ simultaneously.
Understates usage, because self-reported disclosure always lags actual behavior. Authors who used AI and did not disclose are invisible to the metric. The true usage rate is higher than the disclosure rate, always, because disclosure carries a reputational cost and usage often does not. The measurement only captures the compliant slice of the population.
Overstates contribution, because the compliant slice is dominated by tool-layer use, and the tool layer contributes nothing to the mathematics. So the same number is simultaneously too low on adoption and too high on significance. That is not a paradox. It is what happens when you measure one variable and report another.
The blind spot for retail readers is the substitution. They will mentally replace "acknowledgment rate" with "AI contribution rate," and every downstream conclusion they draw โ about academic integrity, about AI capability, about research funding โ inherits that substitution error. Fear is a bad indicator, data is a leader, and this is not even data yet. It is a press release with a percentage sign.
Leverage magnifies character, not just capital. Apply that to metrics. A weak, unaudited signal amplified by leverage โ by media repetition, by policy citation, by fund allocation โ does not become strong. It becomes a larger version of its own weakness. The 25% is being levered right now by every outlet that repeats it without the methodology, and the liquidation event, when the primary report finally lands with different coding rules, will be a credibility drawdown rather than a price drawdown. Same mechanism. Different asset.
Takeaway
Do not trade the headline. Trade the methodology.
The single event that resolves this entire story is the release of Epoch AI's primary report: the sample size, the baseline, the coding rule, and the definition of an AI acknowledgment. Until that document exists, the 25% is a claim, not a measurement, and any position โ intellectual or financial โ built on it is a position built on an unaudited line item.
The signals worth tracking are mechanical. Whether arXiv and the major publishers tighten disclosure policy, because that predicts the next disclosure-rate jump regardless of actual usage. Whether an independent evaluator replicates the number with a disclosed coding rule. Whether the formal-verification ecosystem produces a result that a referee accepts as genuine mathematics, because that is the only layer where the word "contribution" survives contact with a validator.
Audit the logic before you trust the label. The label says 25%. The logic, so far, says nothing at all. Red candles do not negotiate with hope โ and neither does a methodology you have not read.