ChatGPT 和 Claude 答錯數據表數值的七個案例及查核方法

Disclosure: I build PartsPlan, the tool mentioned at the end. The cases below are real answers I checked against the manufacturer's pages.

If you ask a chat AI for a component value, you get a confident number, and a wrong value looks exactly like a right one. Between Oct 29 and 31, 2026 I asked ChatGPT and Claude web chat (PDF attached or web search on) questions where I expected trouble, then checked every answer against the datasheet page. Here are seven misses, and the habit that would have caught each one.

1. BSS138 — the wrong temperature curve (about 4x off)

Asked for the 125 °C drain current at VGS = 1.75 V with the onsemi PDF attached, the answer said about 0.03 A. On page 4, Figure 5, the three temperature curves cross near 2.2 V; below the crossing the 125 °C curve is the highest, about 0.117 A at 1.75 V.

Check: on transfer curves, find where the curves cross. The order of the curves flips on either side.

2. ATtiny85 — right value, evidence from another chip

The value (4.63 V output-high at 15 mA) and the page were right, but the graph shown as evidence came from an ATmega328P datasheet.

Check: the figure number and the part name in the figure title must match your part.

3. TL431 — numbers that aren't in the datasheet

The answer said it couldn't open the PDF, then gave Vref as 2.482 / 2.508 V and explained that "TI rounds outward". TI's datasheet (SLVS543S, p.12, Table 6.11) prints 2483 / 2507 mV; 2.482 and 2.508 appear nowhere in its 85 pages.

Check: match the full part number (grade and package) to the manufacturer's own table.

4. ATmega2560 — a graph value from a PDF whose images weren't read

With the 436-page PDF attached, the answer said the curve values weren't available, then gave "about 10 µA, read from the graph". Page 380 shows about 9.2 µA. Claude's help page says PDFs of 101–1000 pages are processed as text only.

Check: for long PDFs, ask which page and figure the value came from, and open it.

5. ATmega328P — the PDF never got in

The 33.3 MB datasheet was refused at the 30 MB attachment limit.

6. STM32F103RCT6 — "not in the datasheet" (it's on a graph)

The answer found the table values but gave no 85 °C typical value and pointed me to the page-52 graph to read myself. Figure 18 shows about 240 µA. (The 365 µA in Table 17 is the guaranteed maximum, a different number.)

Check: typical values often live only in the Typical Characteristics graphs.

7. Panasonic ALDP112 — the line that matters was left out

Asked whether a 12 V relay could be driven at 15 V, the answer reasoned from the 180% maximum allowable voltage (at 20 °C, a footnoted limit) but left out the line on the same page: "please use the relay within ±5% of rated coil voltage."

Check: read the notes around a rating table, not just the table.

What these have in common

None of these are exotic. Graph-only values, long PDFs, and close-but-different numbers are everyday datasheet questions. These are selected failures, not an accuracy rate — I asked where I expected the tools to be weak, one or two runs per question — and results vary by model, settings and time.

What I built

PartsPlan is an MCP server you connect to ChatGPT, Claude, Cursor or VS Code. The model looks up the manufacturer's datasheet, finds the page, reads the table or graph, and returns each value with the document and page number so you can open the original in seconds. Values it can't establish are left missing rather than estimated. It isn't error-free: in two of the runs above, the PartsPlan-connected reply had a wrong side detail. That's exactly why every value carries a page.

Free tier: 300 tool calls a month, no card; ChatGPT and Claude connect with a login.

Connect: https://partsplanai.com/connect?utm_source=devto&utm_medium=article&utm_campaign=devto_1102
All seven cases with screenshots and source pages: https://partsplanai.com/en/results/real-cases?utm_source=devto&utm_medium=article&utm_campaign=devto_1102

Sources: onsemi BSS138 Rev. 7 p.4 Fig. 5 · Atmel ATtiny25/45/85 2586Q p.184 Fig. 22-22 · TI SLVS543S p.12 Table 6.11 · Microchip DS40002211A p.380 Fig. 32-12 · Microchip DS40002061B p.514 · ST DS5792 Rev 13 p.52 Fig. 18 · Panasonic ASCTB193E (Oct 2020) p.2. ChatGPT, Claude and the part makers' names are trademarks of their owners; no affiliation implied.