CSV from AI: ordinary text can become a formula
Separate CSV field structure from spreadsheet interpretation and test the actual recipient workflow without assuming one universal mitigation.
Article language: English

In this article
Quotation marks in CSV organise the representation of a field. They are not an agreement with a spreadsheet program that its contents will remain plain text. This distinction can surprise a team accepting an AI-prepared export: a file passes structural checks, yet one cell starts behaving like a formula when opened. A valid file and an appropriate way to use it require separate checks.
Imagine a fictional training team collecting exercise suggestions. A model organises descriptions and returns them in CSV. One suggestion is simply =2+3 because participants are meant to explain that notation. In the "exercise content" column, we expect literal text. The numerical result 5 does not fulfil that task, however mathematically correct it may be. The spreadsheet has rushed to answer before the trainer could ask the question.
We use only this harmless arithmetic to explain the boundary. We do not describe executing commands or retrieving data through formulas. No test of a particular Excel or Calc version was performed here. We show how to prepare a controlled acceptance test that establishes the behaviour of the program actually used by the company.
Two agreements: representing fields and interpreting them later
CSV stores fields separated by an agreed delimiter. An exercise description may contain a comma, a quotation mark or a newline. The exporter should keep them in the correct cell. If the text spills into an extra column, there is a writing or import error before formulas even come into question.
Correctly separating columns does not establish the content type, however. The spreadsheet program makes further decisions: it recognises numbers, dates or expressions. This happens after the fields are read. Quoting an entire field can correctly preserve a comma in the description while still not deciding the intended cell behaviour.
OWASP describes CSV Injection and the limitations of neutralisation measures. It warns, among other things, that a protected representation can change when Excel saves and reopens a file. It also emphasises that there is no single strategy suitable for every spreadsheet and downstream consumer. A maliciously prepared formula may also attempt to expose spreadsheet data. The effect depends on the application, its version and security settings, so assess the specific path used to open the file.
Do not reduce this to whether the model meant well. A value can come from a form, a customer's file or material copied by the model. Automation adds a processing stage, but does not make the content trusted. The exporter is responsible for the conditions under which data reaches its recipient, regardless of who composed the sentence in the cell.
The Polish-language article on JSON and the meaning of AI data (in Polish) distinguishes syntax from value correctness. Here there is another distinction: valid text in a model response may be treated as an instruction by the next program. Checking the response before export does not automatically cover that boundary.
Who is actually receiving the file?
The training team needs to read suggestions, add comments and return a selection. Another application may ingest an identical CSV as a data source. These uses need not share one file preparation method. A version convenient for a human reader might contain an extra character that an automated importer treats as part of the description.
Start by identifying the specific path. Which program, version and opening method will acceptance cover? Double-clicking a file, choosing an import command and pasting text are different actions. Do not group them all as "opening in a spreadsheet" when different settings can change the result.
Microsoft explains that Excel uses current default data-format settings when opening a CSV. Its documentation distinguishes this from other import paths. The CSV extension alone therefore does not describe every condition under which columns are read.
In our example the recipient is the trainer. She should not have to remember how the author's computer was configured. She receives the specified file version and a short import path matching the tested environment. If the organisation permits several programs, each needs its own result or a clearly limited use. Testing one does not justify a claim about all of them.
If you are considering a conversation with Aeneia Audit, prepare an outline of the file's journey from the model to the application your team uses. Agree whether this workflow should be included in a review and what you expect to learn. A specialist security assessment of the exporter's code needs a separately agreed scope and an appropriate provider.
The process owner should also decide whether the next recipient may edit the file. Read-only material and a table returned for import have different requirements. If the export must return to the system, record identifiers need to be preserved and source text distinguished from any representation prepared for a spreadsheet.
A harmless fixture instead of testing company data
Prepare a separate file using only agreed fictional values. In the description column, use separate rows for short ordinary text, the literal expression =2+3, and descriptions containing a comma, quotation mark and newline. Give each row a fictional identifier. These cases test different things. The first is a reference point, the second reveals interpretation, and the others check field boundaries.
Write the expected result for each before opening the file. In our task the expression must remain text, not be calculated. A description containing a comma must occupy one cell, and multiline text must not create an extra record. These expectations allow the result to be assessed without guessing from a tidy-looking table.

LibreOffice 26.8 help describes the Evaluate formulas option, which determines whether expressions beginning with an equals sign are calculated or imported as text. It separately describes the Text column type. These are documented settings of that product, not universal instructions for every spreadsheet.
The test record should include the selected import path and actual settings. Then compare the file's text with the cell contents. Seeing a number alone does not reveal whether it is a constant or the result of an expression. The person testing should be able to explain how they recognised the distinction in the program used. Do not add a result that was not checked.
Add two separate expectations for the same value to the test description. In the "exercise content" column, the expression is text. In a separate control column intended for calculations, the addition result may be exactly what we want. The issue is agreement with the field's role, not a ban on all formulas in every workbook. The process owner decides where calculation is needed and where its definition may come from.
There is no need to begin with confidential data or a file downloaded from a random place. The exporter owner can prepare a fixture from agreed values. It must, however, pass through the same generation path as an ordinary result. A hand-corrected demonstration does not prove that the actual AI export works.
Also retain a description containing a delimiter followed by the literal expression. Checking only the first character of the whole row is not enough. The check concerns the value in each field after correct separation. The implementer should demonstrate that both cell boundaries and the expected interpretation are preserved. These are two parts of evidence that can be inspected separately.
A fix should not quietly change the content
The easiest reaction is to remove every problematic character. In our dataset that would destroy the exercise. Elsewhere, a minus can denote a valid negative number, and an initial plus can belong to a textual identifier. A protective rule needs to know the column type, not just have a list of characters to delete.
We do not recommend a single prefix as a solution for every case. The implementer must choose data preparation appropriate to the agreed recipient, preserve the original and demonstrate the expected result. If the spreadsheet representation differs from the source, describe that explicitly. A downstream application should not receive a changed value merely because someone previously viewed the file.
A prompt to the model saying "return safe CSV" is not enough either. The appropriate part of the process should enforce the expected representation. The model can organise text, while a controlled exporter takes responsibility for passing it on. Acceptance covers the actual file, not a declaration in a chat response.
When designing a table export, name the finished file's intended use; see the Polish-language article on table exports (in Polish). "For viewing in the agreed spreadsheet" carries a different expectation from "raw data for reimport". The download interface should help choose the right variant if the process actually offers both.
Similarly, previewing AI-prepared HTML or SVG requires a defined display method, as discussed in the Polish-language article on HTML and SVG previews (in Polish). In both situations content enters a program that gives it behaviour. That does not mean identical protections: CSV requires assessment of cell import, while a web document has a separate preview context. Do not transfer a result from one check to the other format.
Replacing CSV with a workbook also requires acceptance testing. Another format may allow cell types to be set explicitly, but a new file extension alone does not confirm correct implementation. If the material contains active Office document elements, separate questions about its contents arise; see the Polish-language article on active Office content (in Polish). Do not silently expand tests by adding features unnecessary for simply reading the suggestions.
The file comes back. Acceptance continues
The trainer adds a comment, saves the CSV and returns it to the coordinator. It is a second version of the file, even if its name remains the same. If that exchange is normal work, the test should include saving and reopening in the agreed program. The first import result does not describe the whole process.
Then compare both the cell's interpretation and the content returned to the system. The literal expression must still mean the same exercise. Also check that the quotation mark and newline in a description were not lost. If the program changed the representation, preserve examples from before and after rather than correcting them manually in the report.
The Polish-language article on spreadsheets in AI processes (in Polish) also discusses why record identity matters. After the trainer reorders rows, the importer should recognise the exercise by its retained identifier and assign the comment to the correct entry. This is separate from formula interpretation, but both criteria must be met for the returned file to support further work.
A short record should retain a sample source record, a file exported through the actual process, import settings and the read result. For each deviation identify the stage: field separation, content interpretation or resaving. If only the exporter was fixed, the repeat should use a new export of the same values. Changing the exercise text to something easier to accept does not prove the program was fixed.
Acceptance may end with a limited decision: a given export version is for reading through a specified path, and resaving CSV is outside the accepted use. If exchanging comments is necessary, the process needs another tested option. The limitation should match real work; otherwise users will bypass it at the first training deadline.
For our team, the result should be clear without understanding the mechanism: the trainer sees the literal exercise, can add a comment through the intended path, and the coordinator receives the same content with the correct record. Only that evidence justifies accepting the export. The mere fact that a table opened and can add numbers is a rather modest achievement.







