You're assuming the AI will not be given any instructions to produce the report with a certain data bias.

That only matters to the extent the report is read, though.

If so, that would be auditable (in theory; whether it would be audited in practice, probably not).