add caveat about risks of multiple data releases

yoid2000 · yoid2000 · commit b65c5424fbfd · 2023-12-17T07:28:30.000+01:00
diff --git a/_pages/syndiffix-mostlyai-ctgan.md b/_pages/syndiffix-mostlyai-ctgan.md
@@ -114,5 +114,5 @@ Synthetic data products are marketed as one-size-fits-all: the promise is that a
 
 ### The privacy of synthetic data products erodes with multiple syntheses
 
-Unlike SynDiffix, the anonymity of other synthetic data products weakens with each data synthesis. As such, a comparison of multiple SynDiffix synthetic datasets against a single Mostly AI or CTGAN synthetic dataset is fair from a privacy perspective.
+Unlike SynDiffix, the anonymity of other synthetic data products weakens with each data synthesis, at least in theory. As such, a comparison of multiple SynDiffix synthetic datasets against a single Mostly AI or CTGAN synthetic dataset is fair from a privacy perspective. Whether generating a large number of synthetic datasets with other products results in real risks in practice is unknown, though it does seem unlikely that generating a relatively small number of datasets leads to practical risks.
 

Original file line number	Diff line number	Diff line change
`@@ -114,5 +114,5 @@ Synthetic data products are marketed as one-size-fits-all: the promise is that a`
`114`	`114`
`115`	`115`	`### The privacy of synthetic data products erodes with multiple syntheses`
`116`	`116`
`117`		`-Unlike SynDiffix, the anonymity of other synthetic data products weakens with each data synthesis. As such, a comparison of multiple SynDiffix synthetic datasets against a single Mostly AI or CTGAN synthetic dataset is fair from a privacy perspective.`
	`117`	+Unlike SynDiffix, the anonymity of other synthetic data products weakens with each data synthesis, at least in theory. As such, a comparison of multiple SynDiffix synthetic datasets against a single Mostly AI or CTGAN synthetic dataset is fair from a privacy perspective. Whether generating a large number of synthetic datasets with other products results in real risks in practice is unknown, though it does seem unlikely that generating a relatively small number of datasets leads to practical risks.
`118`	`118`