Checklist to verify an AI answer about chemical product data

The checklist covers AI answers that state facts about a named chemical product: identity, property values, test methods and the documents behind them. It does not judge formulation advice, hazard classification work or regulatory interpretation, which need a qualified person, and it cannot confirm a value that no primary document records.

When should you use a checklist to verify an AI chemical answer?

Use a checklist to verify an AI chemical answer before any value, identifier or test method from it reaches a customer, a specification, a safety decision or a purchase order. A generative AI tool can state a wrong value as confidently as a right one. NIST's generative AI profile describes "Confabulation" as a phenomenon in which systems "generate and confidently present erroneous or false content in response to prompts" and notes that outputs may include "confabulated logic or citations that purport to justify or explain the system's answer" (NIST AI 600-1, 2024).

The checklist is for technical service, quality, regulatory and purchasing staff at chemical makers, distributors and users who receive AI answers from any tool, including answers pasted in by a colleague or a customer. Use it at three moments: when an answer will be quoted outside the company, when it will change a specification or an order, and when two answers or two documents disagree.

The checklist has 16 items in four groups: identity, values and units, test methods, and sources. Each item names how to verify it against a primary document and the condition that rejects the answer. The result is one of three decisions for each answer: use it, use it after correction, or reject it. A Word version of the checklist, with a blank form and an illustrative filled example, is at the end of the page.

Which identifiers should you check in an AI answer about a chemical?

Check four identifiers in an AI answer about a chemical: the product name and grade, the CAS Registry Number, the CAS number's link to the named substance, and the molecular formula. An answer about the wrong substance or grade fails every later check, so identity comes first.

Compare the product name and grade with the product identifier on the label and in section 1 of the safety data sheet (SDS), which a US SDS must carry (OSHA Appendix D, 2024). An answer that names a product family or a different grade is about something else.

Check the CAS Registry Number's format. A CAS RN "includes up to 10 digits which are separated into 3 groups by hyphens," and the final part is a single check digit (CAS, 2026). To compute it, multiply the digits before the check digit by 1, 2, 3 and so on from the right, add the products, and take the last digit of the sum. A number that fails is not a valid CAS RN.

A valid format does not prove the number belongs to the named substance. CAS describes a CAS RN as an identifier that "designates only one substance" (CAS, 2026), so look it up in CAS Common Chemistry or in section 3 of the SDS. The molecular formula is a useful cross-check but, as the American Chemical Society's guide warns, "the formula is not enough information by itself to uniquely identify a substance" (ACS CINF, 1993).

How do you check units and values in an AI chemical answer?

Check units and values in an AI chemical answer by confirming that every number carries a unit, a basis, a test condition and a status (typical, specification limit or lot result), and that it appears in the cited document. A number missing any of these cannot be compared with a specification.

The unit comes first. NIST's guide to the SI states that "the value of a quantity is its magnitude expressed as the product of a number and a unit" (NIST SP 811, 2008). An answer that gives "viscosity 46" or "contains 500" has no value to check.

The basis comes next. A content stated as a percentage or in ppm is ambiguous until mass, volume or amount is named; NIST notes the ambiguity of large-number names "is one of the reasons why the use of ppm, ppb, ppt, and the like is deprecated" (NIST SP 811, 2008). Compare a limit only with a value on the same basis.

The test condition follows. A kinematic viscosity means little without its temperature, and ISO 3104:2023 covers a range from 0.2 mm2/s to 300 000 mm2/s over -20 °C to +150 °C (ISO 3104:2023, 2023). Two values at different temperatures cannot be compared.

Status is the last check. Technical data sheet (TDS) values are typical: one maker states its "values are not intended for use in preparing specifications" (Dow, 2025). An AI answer that presents a typical value as a guaranteed limit or a lot result needs correction.

How do you confirm a test method an AI answer cites?

Confirm a test method an AI answer cites by checking that the standard exists under that number, that the edition is the one the answer or the specification names, that its scope covers the material and range, and that it measures the property claimed. Each check uses the publisher's own page.

Search the number on the publisher's site. A standard number that returns nothing, or a title that does not match the property, rejects the citation.

Check the edition. Publishers mark old editions: on astm.org, ASTM D445-06 is marked "Historical" while ASTM D445-26 is marked "Active" (ASTM, 2026). The national store EVS lists ISO 3104:2020 as withdrawn from 7 November 2023, when ISO 3104:2023 became valid (EVS, 2023). An answer citing an old edition is not wrong by that fact alone, because a specification may fix an edition, but the edition must exist and be the one the comparison needs.

Check the scope. ISO 3104:2023 is intended for liquids in which shear stress and shear rate are primarily proportional (Newtonian flow behaviour), and its note says different viscometers can give different results when viscosity varies with shear rate (ISO 3104:2023, 2023). A result for a strongly shear-thinning product is outside that scope.

Check the property. ISO 3104:2023 measures kinematic viscosity and obtains dynamic viscosity by multiplying by density (ISO 3104:2023, 2023). An answer that quotes a dynamic viscosity in mPa s as a kinematic value in mm2/s has mixed two properties.

How do you check that the sources in an AI answer exist and say what it claims?

Check that the sources in an AI answer exist and say what it claims by opening each one, finding the sentence or table cell that holds the value, and confirming the document is the current revision for the same product and grade. A citation that looks complete can still be invented.

Fabricated citations are measured, not hypothetical. In a 2023 study of literature reviews written by ChatGPT, "55% of the GPT-3.5 citations but just 18% of the GPT-4 citations are fabricated," and 43% of the real works cited by GPT-3.5 had at least one substantive citation error (Walters and Wilder, 2023). That study covered academic citations; product documents were not tested, so treat the rates as a warning, not a measure for chemical answers.

Existence is the first test: find the document itself, not a page that mentions it. The second test is agreement: the cited document must contain the value with the same unit, basis and condition. A source that is real but silent on the claim fails as surely as one that does not exist.

Revision is the third test. A US SDS must state the date of preparation or last change in section 16 (OSHA Appendix D, 2024), and new significant hazard information must be added within three months (OSHA 29 CFR 1910.1200(g)(5), 2024). Compare the date on the cited document with the revision you hold, and the product identifier and grade with the product in question.

The full AI chemical answer checklist with reject conditions

The full AI chemical answer checklist lists 16 items, each with how to verify it and the condition that rejects or corrects the answer. Work down the list in order, because a failed identity item makes every later item meaningless, and record the decision for each item in the form.

The rows that reject most answers in practice are the source rows: an AI answer whose cited document cannot be found, or does not contain the value, is rejected whatever its other qualities. A failed edition or status item usually leads to correction rather than rejection, because the right value can often be read from the current document. Items marked not applicable need a reason in the form, such as "no test method cited", so that a reviewer can tell a skipped check from a passed one.

The checklist shows how to verify each item and when to reject the answer.

How do you fill in the AI chemical answer checklist form?

Fill in the AI chemical answer checklist form one answer at a time: record the answer and its context, mark each of the 16 items pass, fail or not applicable with a note, then record one decision and who made it. The form keeps the answer, the checks and the decision together so that someone else can repeat them.

The table below previews every section of the Word form, with an illustrative example entry for each.

Write the corrected value next to the failed item, with its source, rather than editing the AI answer in place; the record then shows what changed. Where the AI answer came from Lubinpla, sentences that rest on a record or a web page usually carry a numbered source that opens the original record, file or page, so items 14 to 16 start from those sources. Keep the completed form with the answer it checked, and file it where the decision it supported is filed.

What does an illustrative check of an AI chemical answer look like?

An illustrative check of an AI chemical answer shows how the checklist turns one confident paragraph into a decision. The answer, product, CAS number, values and document below are invented for this example.

The illustrative AI answer reads: "Product X (CAS 12345-67-8) has a kinematic viscosity of 46 cSt per ASTM D445-06, so it meets your 50 maximum. Source: Product X TDS, 2019."

Item 2 fails first. The digits before the check digit are 1234567; multiplied from the right by 1 to 7 they give 7, 12, 15, 16, 15, 12 and 7, which sum to 84, so the check digit should be 4, not 8. The CAS number is invalid, and identity is unconfirmed until SDS section 3 supplies the real one.

Items 5, 7 and 8 fail next, and item 6 is not applicable because no content value is given. The value has a unit but no temperature, the "50 maximum" has no unit, and a TDS value is typical, not a lot result. Item 11 fails: D445-06 exists but is Historical, so the specification's edition decides whether it can be used. Item 16 fails last and items 14 and 15 stay open: the TDS cited is dated 2019, and the current revision held by the reader must be opened to see whether the value and the method are still there.

The decision is reject. The identity is wrong, the source is unconfirmed, and the comparison with the limit cannot be made from the answer as written.

When should you reject an AI answer about a chemical product?

Reject an AI answer about a chemical product when its identity is wrong, when a cited source does not exist or does not contain the claim, or when a value needed for a decision has no unit, basis or condition. These failures cannot be fixed by editing the answer, because the fact itself is unconfirmed.

Correct the answer instead of rejecting it when the fact is confirmed but stated imperfectly. Typical cases are an old edition named where the current one gives the same method, a unit written in a non-SI form, a typical value labelled as a limit, and a source that is right but cited by an older revision date. Write the corrected value with the primary document as its source.

Use the answer only after all 16 items pass or are marked not applicable with a reason. Use does not make the AI the source: the record should cite the SDS, TDS, certificate of analysis or standard where each value was found, so that a reader can check it without the AI.

Escalate instead of deciding when the question needs judgment the checklist does not test: hazard classification, regulatory status in a jurisdiction, compatibility with a customer's material, or a value that two primary documents give differently. NIST's profile notes that risks "may be especially important to monitor when integrating GAI into applications involving consequential decision making" (NIST AI 600-1, 2024), and these are such decisions.

Where the AI answer checklist does not apply

The AI answer checklist does not apply to questions without a primary document to compare against, to judgment calls, or to tasks where the AI output is a draft for a qualified person rather than a stated fact. Use it to check facts, not to approve advice.

Facts without a document come first. If no SDS, TDS, certificate of analysis or published standard records a value, the checklist can show that the answer is unsupported but cannot confirm it. Ask the maker for data or run the test.

Judgment is the second limit. Hazard classification, the legal status of a substance in a jurisdiction, product selection for a process and root cause of a failure all depend on reasoning beyond the 16 items. The checklist checks the facts such work rests on; a qualified person still makes the call.

Sector rules are the third. Pharmaceutical, food contact and medical device work have their own requirements for data integrity and document control, and those requirements take precedence over this checklist.

The checklist also does not assess the AI tool. A clean result for one answer says nothing about the next one, so apply the items to every answer that will be used, not to a sample.