diff --git a/man/assign.Rd b/man/assign.Rd
index 91c988cd8..ed5ed8be1 100644
--- a/man/assign.Rd
+++ b/man/assign.Rd
@@ -92,6 +92,7 @@ Since \code{[.data.table} incurs overhead to check the existence and type of arg
     \code{DT[a > 4, b := c]} is different from \code{DT[a > 4][, b := c]}. The first expression updates (or adds) column \code{b} with the value \code{c} on those rows where \code{a > 4} evaluates to \code{TRUE}. \code{X} is updated \emph{by reference}, therefore no assignment needed.  Note that this does not apply when \code{i} is missing, i.e. \code{DT[]}.
 
     The second expression on the other hand updates a \emph{new} \code{data.table} that's returned by the subset operation. Since the subsetted data.table is ephemeral (it is not assigned to a symbol), the result would be lost; unless the result is assigned, for example, as follows: \code{ans <- DT[a > 4][, b := c]}.
+    Note that \samp{:=} modifications are cumulative. When reusing a \code{data.table} in loops or multi-level tests, use \code{\link{copy}} to ensure a fresh state.
 }
 \value{
 \code{DT} is modified by reference and returned invisibly. If you require a copy, take a \code{\link{copy}} first (using \code{DT2 = copy(DT)}).
@@ -180,4 +181,3 @@ system.time(for (i in 1:1000) set(DT, i, 1L, i))
 
 }
 \keyword{ data }
-
diff --git a/man/test.Rd b/man/test.Rd
index 651ef1d35..fea7309be 100644
--- a/man/test.Rd
+++ b/man/test.Rd
@@ -25,7 +25,7 @@ test(num, x, y = TRUE,
 \item{env}{ A named list of environment variables to set for the duration of the test, much like \code{options}. A list entry set to \code{NULL} will unset (i.e., \code{\link{Sys.unsetenv}}) the corresponding variable. }
 \item{context}{ String, default \code{NULL}. Used to provide context where this is useful, e.g. in a test run in a loop where we can't just search for the test number. }
 \item{requires_utf8}{ \code{FALSE} (default), \code{TRUE}, or a character string. When set, the test is skipped if UTF-8 characters cannot be represented in the native encoding. Use \code{TRUE} for default UTF-8 test characters or provide a custom string of test characters. }
-\item{optimize}{ A vector of different optimization levels to test. The code in \code{x} will be run once for each optimization level, with \code{options(datatable.optimize=optimize)} set accordingly. All optimization levels must pass the test for the overall test to pass. If no \code{y} is supplied, the results from the different levels are compared to each other for equality. If a \code{y} is supplied, the results from each level are compared to \code{y}. }
+\item{optimize}{ A vector of different optimization levels to test. The code in \code{x} will be run once for each optimization level, with \code{options(datatable.optimize=optimize)} set accordingly. All optimization levels must pass the test for the overall test to pass. If no \code{y} is supplied, the results from the different levels are compared to each other for equality. If a \code{y} is supplied, the results from each level are compared to \code{y}. Note that since \code{x} is evaluated multiple times, side effects like \code{:=} are cumulative; use \code{\link{copy}} if a fresh state is required for each level. }
 }
 \note{
    \code{NA_real_} and \code{NaN} are treated as equal, use \code{identical} if distinction is needed. See examples below.
diff --git a/vignettes/datatable-reference-semantics.Rmd b/vignettes/datatable-reference-semantics.Rmd
index 5353440e7..c6d6b3c0e 100644
--- a/vignettes/datatable-reference-semantics.Rmd
+++ b/vignettes/datatable-reference-semantics.Rmd
@@ -388,6 +388,22 @@ To select a single column as a vector, remember:
 - `DT[, mycol]` is safer as it always returns a new, independent copy.
 - `DT$mycol` is fast but may return a reference. Use `copy(DT$mycol)` to guarantee independence.
 
+### d) Side effects and testing
+
+Because `:=` modifies by reference, changes are cumulative. If the same *data.table* is reused—for example, in a loop or when using `test()` with multiple `optimization` levels—subsequent runs will start with the modified table from the previous run. Use `copy()` to ensure each run starts with the same data.
+
+```{r}
+DT = data.table(a = 1L)
+# Subsequent runs are cumulative
+DT[, a := a + 1L][]
+DT[, a := a + 1L][]
+
+# Use copy() to isolate runs
+test_expr = function(x) copy(x)[, a := a + 1L][]
+test_expr(DT)
+test_expr(DT)
+```
+
 ## Summary
 
 #### The `:=` operator