Can ChatGPT compare two technical data sheets reliably?
Summary
Only if every value is traced to its source line and both sheets use the same test method, units and conditions; otherwise ChatGPT's comparison misleads. The tool lines up two technical data sheets in seconds, but the PDF alone does not show whether a gap comes from the product or from a different method, temperature or typical-value basis.
ASTM D92-24 (Cleveland open cup) covers petroleum products with flash points above 79 °C and below 400 °C except fuel oils, and points to ASTM D93 for lower flash point substances that D92 may miss (ASTM D92-24, 2024).
ASTM D445-26 reports kinematic viscosity in mm2/s, where 1 mm2/s equals 1 cSt (ASTM D445-26, 2026).
ASTM E691-23 says test method precision matters when comparing results "between data sources (different laboratories, instruments, etc.)" (ASTM E691-23, 2023).
On ChatGPT plans other than Enterprise, file retrieval extracts the digital text of a PDF, not its embedded images (OpenAI, 2026).
When can ChatGPT's comparison of two data sheets be trusted, and when not?
ChatGPT's comparison of two technical data sheets can be trusted for a value only when both sheets measured it the same way and you have checked the number on the page; every other case needs a check or cannot be compared. The table below decides, row by row, whether an AI side-by-side of two data sheets can be used as is, needs a source check, or cannot be compared at all.
Condition | Verdict | Why |
|---|---|---|
Same test method, edition, unit and test temperature; value checked on the page | Usable | Like for like; still weigh the gap against the method's precision (ASTM E691-23, 2023) |
Flash point by ASTM D92 open cup on one sheet, ASTM D93 closed cup on the other | Not comparable | Different apparatus; D93 is the method for low flash point substances D92 may miss (ASTM D92-24, 2024) |
Kinematic viscosity at 40 °C on one sheet, 100 °C on the other | Not comparable | Each result belongs to its test temperature; ask for the same temperature |
Same property in different units (cSt and mm2/s, °F and °C) | Check the conversion | 1 mm2/s equals 1 cSt (ASTM D445-26, 2026); converted °F values may be rounded |
Typical value on one sheet, specification limit on the other | Not comparable as a limit | Typical properties "do not constitute a specification" on one maker's sheet (ExxonMobil, 2026) |
Value with ">", "<", "min" or a footnote | Check on the page | ">210" °C is a bound, not a measured 210 (ExxonMobil, 2026) |
Scanned table or table saved as an image | Check every value | Outside Enterprise, retrieval reads digital text, not embedded images (OpenAI, 2026) |
The flash point row shows the problem: a Cleveland open cup result and a Pensky-Martens closed cup result are two different tests, and no AI comparison turns them into a difference between products. Read the table rows before a summary line such as "the second product has the higher flash point", because a summary that hides the method row also hides why it may be wrong.
Why do test methods and typical values break an AI comparison of two TDS?
Test methods and typical values break an AI comparison of two technical data sheets (TDS) because the model compares numbers, while the meaning of each number sits in its method, its conditions and its basis. ASTM D92-24 measures flash point in an open cup and ASTM D93-26 in a closed cup, and D92 itself says D93 "should also be used when it is desired to determine the possible presence of small but significant concentrations of lower flash point substances that may escape detection by Test Method D92" (ASTM D92-24, 2024). Kinematic viscosity under ASTM D445-26 or ISO 3104:2023 is measured at a chosen temperature, so a 40 °C value and a 100 °C value describe the fluid in two different states. Typical values add a second trap: one maker's data sheet says its typical properties "are typical of those obtained with normal production tolerance and do not constitute a specification" (ExxonMobil, 2026). A typical value set against a guaranteed maximum compares an average with a limit. Even a like-for-like gap needs context. A gap smaller than the method's published reproducibility may be lab-to-lab scatter, and ASTM D3244-24 uses that reproducibility when a supplier and a receiver dispute conformance with a specification (ASTM D3244-24, 2024).
What table-reading errors does ChatGPT make on data sheet PDFs?
ChatGPT can misread data sheet PDFs in four ways that matter for a comparison: it can miss values held in images, drop a qualifier such as ">", attach a value to the wrong row or unit, and state a wrong value with confidence. On plans other than Enterprise, retrieval "extracts digital text from the file rather than its embedded images" (OpenAI, 2026), so a scanned table may never reach the model. A 2024 benchmark of GPT-3.5 and GPT-4 on tables found that performance "varied with different input choices, including table input format, content order, role prompting, and partition marks" (Sui et al., 2024). NIST calls the wider failure confabulation: "the production of confidently stated but erroneous or false content" (NIST AI 600-1, 2024). Qualifiers are easy to lose. One product data sheet lists a Cleveland open cup flash point as ">210" °C (ExxonMobil, 2026), and a comparison that drops the ">" turns a lower bound into a measurement. Footnotes that set the method or test temperature often sit below the table, away from the value they change. The same rule holds for any AI answer drawn from company files, covered in can ChatGPT answer from my company's technical documents: trust only what you can trace to the page.
How do you check an AI comparison of two technical data sheets?
You check an AI comparison of two technical data sheets by tracing each compared value to its line on each sheet, in a fixed order, before you draw any conclusion. Use this order for every row that matters to the decision:
Ask for a table with property, value, unit, test method and page number for each sheet, and reject rows without a page.
Open each page and confirm the number, its qualifier (">", "<", min, max, typical) and any footnote.
Confirm that method, edition and test temperature match; mark mismatched rows "not comparable" instead of converting them.
Redo unit conversions yourself; 1 mm2/s equals 1 cSt (ASTM D445-26, 2026).
For matched rows, compare the gap with the method's published reproducibility before calling it a difference.
Ask both suppliers for results on the same method and temperature for every row that still decides the choice.
In Lubinpla, you can attach both data sheets to the main agent or @mention them from your records, ask for the comparison, and check each value against its numbered source before you rely on it.
Frequently asked questions
Can ChatGPT convert a flash point from ASTM D92 to ASTM D93?
The public scope pages of ASTM D92-24 and ASTM D93-26 give no conversion between the two methods, and D92 refers users to D93 for low flash point contaminants (ASTM D92-24, 2024). Treat an AI-converted value as unsupported and ask the supplier for a result on the method your specification names.
Is a ChatGPT comparison table enough for a purchasing decision?
Not on its own. The table is a list of leads until each value carries its page and has been checked there. Use it to find the rows that differ, then decide on the checked values and, where methods differ, on matched test data from the suppliers.
Does asking ChatGPT to cite page numbers solve the problem?
Page numbers make the check faster, but a cited page can still be the wrong one or hold a different value. The check is opening the page and reading the line, not the presence of a citation in the answer.