A spreadsheet that looks neat, a JSON answer that passes schema validation, a table with tidy columns: it looks reliable. Yet several studies from the first half of 2026 describe that this formal correctness says little about the substantive accuracy of the figures. By our analysis, this makes a strong case for treating every numerical statement and every calculated field from AI as a claim that deserves separate scrutiny, especially for professionals working with sensitive or high-impact data.
The immediate occasion is the paper StructHallu-Drift: Benchmarking Structured Hallucinations Under Schema Evolution in LLMs (July 2026). It introduces a benchmark for structured output — SQL, JSON and record format — and reports that between 39 and 54 per cent of those outputs contain at least one semantic hallucination. In this benchmark, schema validation caught most syntactic errors, but around 8 per cent of semantic errors remained, particularly in type coercion and relational inconsistencies. In other words: the structure is correct, but the number or the relationship between fields is not.
Checking structure is not the same as checking figures
This is the crux. A JSON or schema check tests form: are the required fields present, are the types correct, is the structure valid? That is useful, but it says nothing about whether the number in a cell is the right number, whether an aggregation is correct, or whether rows and columns are consistent. The StructHallu-Drift benchmark shows that it is precisely there that the errors remain.
A June 2026 study, When LLMs Read Tables Carelessly: Measuring and Reducing Data Referencing Errors, makes the same problem concrete at the table level. The study describes that large language models systematically extract the wrong numbers from tables: they read the wrong cell, ignore relevant values or refer to values that do not exist. In that study, accuracy in data-referencing tasks rose by up to 12 percentage points by adding a specialised critic model that checks answers against the underlying table. In our estimation, this underscores above all that a single model without a checking layer is vulnerable here.
Even AI research itself is full of errors in figures and tables
That error-proneness is not a fringe phenomenon is shown by Systematic Quantification of Errors in Published AI Papers (December 2025). An automated Correctness Checker found an average of 4.7 objective errors per paper across 2,500 published AI papers — including incorrect calculations in tables and inconsistencies between text and table. Human experts confirmed 83.2 per cent of the errors reported by the system. This explicitly concerns errors in published AI papers, not a direct measurement of production systems; in our estimation it serves above all as an indication of the error-proneness of formulas and tables.
Two things stand out. First: automated checking of formulas, calculations and tables is feasible and finds a great deal. Second: the system itself is not infallible — 83.2 per cent of the reports were confirmed, so about 16.8 per cent fell away under human scrutiny (a calculated approximation based on that percentage). This indicates that automatic validation is one layer in a broader chain, not a final verdict.
How critical users already do it in practice
The CHI 2026 study "I'm Always a Little Skeptical of It" tracked, in a setup with twelve blind spreadsheet users, how they verify AI output in spreadsheets. All participants stated that they never fully trust generative AI output. They combined five strategies: manual checks in the spreadsheet, follow-up questions to the same model, checking questions to other AI systems, verification by sighted people and testing against their own domain knowledge.
In that study, about half of the observed errors went unnoticed, especially in visual elements such as charts, conditions and formatting. By our analysis, this is a sobering lesson: even multiple verification is no assurance, and precisely for that reason it seems sensible to us to make the checks that have been carried out explicit and visible rather than leaving them in the user's head.
Validation becomes a product feature — but responsibility stays with the user
Providers are now building this kind of checking in. In the official Google Workspace update Troubleshoot formula errors quickly with Gemini in Sheets (June 2026), a feature is announced that lets users have formula errors analysed and corrected in one click. According to that announcement, Gemini can analyse formula errors, explain the cause and propose a corrected formula.
The announcement is clear about the limit: it is a tool for error detection, not an assurance of correct outcomes. Responsibility for correct use and interpretation lies, according to the update, explicitly with the user. In our estimation, this means that organisations themselves must determine how an AI suggestion for a calculation is weighed before it counts in decision-making.
A layered, visible verification chain
From these sources, by our analysis, a consistent picture emerges. Reliable validation of figures, calculations and tables from AI requires multiple layers: schema validation for form, semantic and relational checks for content, execution tests that actually run formulas and queries, a critic model that checks data references against the source, and finally human review of the outcomes. This indicates that none of these layers is sufficient on its own.
A verification console such as I am Vera fits here. Vera is a privacy-focused verification layer — not a model that produces figures itself and not a chatbot — that can route a task via selected independent AI models and make the checking steps that have been carried out visible. This supports review and control, but does not assure correctness. For professionals working with confidential information, it is also relevant that the product architecture is designed so that pre-processing and anonymisation take place on EU infrastructure and only anonymised content is sent to the selected AI models; if that privacy check fails, nothing is forwarded (see the Semantic Privacy Shield). This is a description of design and architecture, not a legal guarantee. With Vera Office, documents can be viewed and edited within the same secure environment.
Vera does not assure correctness and does not remove the risk of hallucinations. What a verification console can do is support the checking and provide more visibility: showing which model generated which table or calculation, which validation steps were carried out and where something was corrected, and recording that for audit and accountability. The professional final judgement on the figures always remains with the user.