# Confident arithmetic error Part of [The AI Tells Index](https://feedsquad.com/ai-tells). A tell signals low effort. It does not identify an author. Skilled writers produce every shape listed here, some of them daily, and automated detectors misread those writers for it at rates measured above 60 percent on non-native English prose. Nothing in this index proves that a machine wrote anything. Read an entry as one piece of evidence to weigh against the false-positive notes printed beside it. ## Facts - Id: confident-arithmetic-error - Category: Semantic (semantic) - Subcategory: factual-failure - Also known as: fluent but wrong numbers, APR and APY confusion - Status: Active. Currently signals low-effort writing. - Severity: high - Evidence grade: corroborated (named independently by multiple credible secondary sources) - Languages: en - Added: 2026-08-15 - Updated: 2026-08-15 - Page: https://feedsquad.com/ai-tells/confident-arithmetic-error ## Description Calm, grammatical, authoritative prose whose numbers do not survive a calculator. The register never wavers. Nothing in the sentence signals that the arithmetic went wrong, because nothing in the sentence was ever checked. ## Why it reads as machine-written Fluency and arithmetic are produced by different faculties in a language model, and only one of them is being optimised in the sentence. The worked example that made this famous is a set of personal-finance explainers published in late 2022 and early 2023, in which a total was reported as a gain: ten thousand at three percent yields three hundred, and the text gave the balance instead. CNN reported 77 published stories written with an internally designed AI engine, an audit that followed the first factual error, and corrections on a number of them with a small number described as substantial. The outlet did not say how many of the 77 were corrected, so no count appears here. The confidence is what does the damage, because a hedged wrong number invites checking and an assured one does not. ## Detection Type: judge (a rubric for a lightweight model judge) Rubric for a lightweight model judge: ```text Task: recompute every number in the text. Step 1. Extract each numeric claim with its unit, its rate type and its period. Step 2. Recompute. Check these families in order, since they cover most observed failures: (a) total versus gain, where a balance is reported as interest earned or a revenue figure as profit; (b) rate types, where an annual percentage rate is treated as a yield, or a simple rate as a compounding one; (c) percentages versus percentage points; (d) parts against their sum, where listed components do not add to the stated total; (e) date arithmetic, where a duration disagrees with its endpoints; (f) unit and scale, where thousands and millions or metric and imperial are mixed. Step 3. Record the stated value, the computed value and the difference. Escape hatches: figures inside a quotation, which belong to the speaker; numbers the text marks as rounded, illustrative or hypothetical, where the rounding accounts for the gap; currency conversions with no stated date; deliberately simplified worked examples that say they are simplified. Decision. FLAG on any error large enough to change the reader's decision, and note the rest as minor. Report the recomputation, not an opinion about the writer. Output: a table of stated value, computed value, and verdict per claim. ``` ## Examples Constructed specimens. Written for this index. Never quoted from anyone's posts. ### Specimen 1 Before, exhibiting the tell: > Put 10,000 euros into an account paying 3 percent a year and after twelve months you have earned 10,300 euros in interest. After, repaired: > Put 10,000 euros into an account paying 3 percent a year and after twelve months you have earned 300 euros. The balance is 10,300. The gain is 300. Confusing those two is the most common error in this kind of explainer and it is worth naming. Note: The arithmetic in the repair is real and checkable, which is the point of the repair. ### Specimen 2 Before, exhibiting the tell: > A 25,000 euro loan at 4 percent costs a flat 1,000 euros in interest per year for the life of the loan, so five years of borrowing costs 5,000 euros. After, repaired: > A 25,000 euro loan at 4 percent costs about 1,000 euros in the first year and less every year after that, because interest is charged on what is left. Repaid in five equal annual instalments of roughly 5,620 euros, the total interest lands near 3,100 euros rather than 5,000. Note: The repair was computed rather than estimated. Anyone can rerun it with an amortisation formula and should. ## False positives Who legitimately writes this way. People are bad at arithmetic and always have been. Financial journalists confuse a rate with a yield, subeditors drop a zero, and spreadsheets export percentages as decimals into copy that nobody recomputes. Fatigue produces the same sentence as anything else. What separates a slip from the pattern is distribution: one wrong figure in a piece is a correction, while a run of pieces in which the numbers are consistently wrong in the same direction and the prose is consistently assured points at a process with no numerate reader in it. That is a finding about a pipeline, not about a person, and it should be reported that way. ## Model attribution The documented case involved a proprietary in-house system rather than a named consumer assistant, which is worth stating because the tell is often attributed to a vendor it cannot be traced to. ## Sources 1. Futurism, CNET's article-writing AI is already publishing very dumb errors https://futurism.com/cnet-ai-errors (tier: press; accessed 2026-08-14) 2. CNN, plagued with errors, a news outlet's decision to write stories with AI backfires https://www.cnn.com/2023/01/25/tech/cnet-ai-tool-news-stories (tier: press; accessed 2026-08-14) ## Status history Ids are permanent. A retired tell keeps its id and its page. - 2026-08-15, Active: Documented in a news outlet's own correction record and covered by two independent outlets, one of which reported the correction count. The failure recurs wherever numeric copy is published without a numerate reader in the loop. ## License CC BY 4.0. https://creativecommons.org/licenses/by/4.0/ Attribution: The AI Tells Index, feedsquad.com/ai-tells Reuse the data, including commercially. Keep the attribution line. --- Dataset version 1.0.0. Schema version 1. Part of [FeedSquad](https://feedsquad.com). Built by [Herman Foundry](https://hermanfoundry.com) from Levi, Finnish Lapland.