diff --git a/NEWS.md b/NEWS.md
index e26cf65c1e..fe861cebcf 100644
--- a/NEWS.md
+++ b/NEWS.md
@@ -86,6 +86,8 @@
         out_col_name = "sum_x"
       )]
     ```
+    
+11. `DT[, if (...) .(a=1L) else .(a=1L, b=2L), by=group]` now returns a 1-column result with warning `j may not evaluate to the same number of columns for each group`, rather than error `'names' attribute [2] must be the same length as the vector`, [#4274](https://github.com/Rdatatable/data.table/issues/4274). Thanks to @robitalec for reporting, and Michael Chirico for the PR.
 
 ## BUG FIXES
 
diff --git a/R/data.table.R b/R/data.table.R
index fd0e9fd1cd..3a3cf1f29d 100644
--- a/R/data.table.R
+++ b/R/data.table.R
@@ -921,8 +921,12 @@ replace_dot_alias = function(e) {
               if (is.name(thisq)) nm[jj] = drop_dot(thisq)
               # TO DO: if call to a[1] for example, then call it 'a' too
             }
-            if (!is.null(jvnames) && any(idx <- nm != jvnames))
-              warning("Different branches of j expression produced different auto-named columns: ", brackify(sprintf('%s!=%s', nm[idx], jvnames[idx])), '; using the most "last" names', call. = FALSE)
+            if (!is.null(jvnames)) {
+              if (length(nm) != length(jvnames))
+                warning("j may not evaluate to the same number of columns for each group; if you're sure this warning is in error, please put the branching logic outside of [ for efficiency")
+              else if (any(idx <- nm != jvnames))
+                warning("Different branches of j expression produced different auto-named columns: ", brackify(sprintf('%s!=%s', nm[idx], jvnames[idx])), '; using the most "last" names', call. = FALSE)
+            }
             jvnames <<- nm # TODO: handle if() list(a, b) else list(b, a) better
             setattr(q, "names", NULL)  # drops the names from the list so it's faster to eval the j for each group; reinstated at the end on the result.
           }
@@ -1367,7 +1371,10 @@ replace_dot_alias = function(e) {
         setattr(jval,"names",NULL)  # discard names of named vectors otherwise each cell in the column would have a name
         jval = list(jval)
       }
-      if (!is.null(jvnames) && !all(jvnames=="")) setattr(jval, 'names', jvnames)  # e.g. jvnames=="N" for DT[,.N,]
+      if (!is.null(jvnames) && any(nzchar(jvnames))) {
+        if (length(jvnames) > length(jval)) jvnames = jvnames[seq_along(jval)]  #4274
+        setattr(jval, 'names', jvnames[seq_along(jval)])  # e.g. jvnames=="N" for DT[,.N,]
+      }
       jval = as.data.table.list(jval, .named=NULL)
     }
 
diff --git a/inst/tests/tests.Rraw b/inst/tests/tests.Rraw
index 7959217d97..3001616f90 100644
--- a/inst/tests/tests.Rraw
+++ b/inst/tests/tests.Rraw
@@ -3058,7 +3058,7 @@ test(1034, as.data.table(x<-as.character(sample(letters, 5))), data.table(V1=x))
   # na.rm=TRUE with list column value, PR#4737
   test(1035.016, melt(data.table(a1=1, b1=list(1:2), b2=list(c('foo','bar'))), na.rm=TRUE, measure.vars=list(a="a1", b=c("b1","b2"))), data.table(variable=factor(1), a=1, b=list(1:2)))
   test(1035.017, melt(data.table(a1=1, b1=1, b2=2), na.rm=TRUE, measure.vars=list(a="a1", b=c("b1","b2"))), data.table(variable=factor(1), a=1, b=1))#this worked even before the PR.
-  
+
   ans1 = cbind(DT[, c(1,2,8), with=FALSE], variable=factor("l_1"))
   ans1[, value := DT$l_1]
   test(1035.02, melt(DT, id.vars=c("i_1", "i_2", "l_2"), measure.vars=c("l_1")), ans1)
@@ -10890,7 +10890,7 @@ test(1743.217, sapply(fread("a,b,c,d,e,f\na,b,c,d,e,f", colClasses = list(factor
 test(1743.218, sapply(fread("a,b,c,d,e,f\na,b,c,d,e,f", colClasses = list(factor = c(1, 2, 4), factor = 3), select = c(5, 4, 2, 3)), class), y = c(e = "character", d = "factor", b = "factor", c = "factor"))
 
 test(1743.22, fread("a,b,c\n1999/01/01,2,f", colClasses=list(Date=1L), drop="a"), data.table(b=2L, c="f"))
-test(1743.231, fread("a,b,c\n2,1,4i", colClasses=list(complex="c", integer=2L), drop="a"), data.table(b=1L, c="4i"), 
+test(1743.231, fread("a,b,c\n2,1,4i", colClasses=list(complex="c", integer=2L), drop="a"), data.table(b=1L, c="4i"),
      warning=paste0(base_messages$coerce_na, ".*left as type 'character'"))
 test(1743.232, fread("a,b,c\n2,1,3+4i", colClasses=list(complex="c", integer=2L), drop="a"), data.table(b=1L, c=3+4i))
 test(1743.241, fread("a,b,c\n2,2,f", colClasses = list(character="c", integer="b"), drop="a"), data.table(b=2L, c="f"))
@@ -17531,7 +17531,7 @@ test(2183.40, names(melt(iris.dt, measure.vars=patterns("[.]"))), c("Species", "
 # measure with pattern=
 test(2183.41, melt(DTid, measure.vars=measure(value.name, istr="bar", pattern="([ab])([12])")), error="each ... argument to measure must be a function with at least one argument, problem: istr")
 test(2183.42, melt(DTid, measure.vars=measure(value.name, istr=function()1, pattern="([ab])([12])")), error="each ... argument to measure must be a function with at least one argument, problem: istr")
-test(2183.43, melt(DTid, measure.vars=measure(value.name, istr=interactive, pattern="([ab])([12])")), error="each ... argument to measure must be a function with at least one argument, problem: istr")    
+test(2183.43, melt(DTid, measure.vars=measure(value.name, istr=interactive, pattern="([ab])([12])")), error="each ... argument to measure must be a function with at least one argument, problem: istr")
 test(2183.44, melt(DTid, measure.vars=measure(value.name, istr=function(x)1, pattern="([ab])([12])")), error="each ... argument to measure must be a function that returns an atomic vector with same length as its first argument, problem: istr")
 test(2183.45, melt(iris.dt, measure.vars=measure(value.name, dim, baz, pattern="(.*)[.](.*)")), error="number of ... arguments to measure =3 must be same as number of capture groups in pattern =2")
 test(2183.46, melt(iris.dt, measure.vars=measure(function(x)factor(x), dim, pattern="(.*)[.](.*)")), error="each ... argument to measure must be either a symbol without argument name, or a function with argument name, problems: 1")
@@ -17591,3 +17591,8 @@ DT = data.table(id=c(1,1,2,2), x=c('y','y','y','z'), v=c('a','b','c','d'))
 test(2185, dcast(DT, formula=id~x, fun.aggregate=agg, value.var='v'),
            data.table(id=c(1,2), y=c('a','c'), z=c(NA,'d'), key="id"))
 
+# compatible branches might seem incompatible if the condition is global, #4274
+DT = data.table(a=1L)
+test(2186, DT[, if (TRUE) .(a=1L) else .(a=1L, b=2L)], DT,
+     warning='j may not evaluate to the same number of columns for each group')
+