Preliminary thoughts on building safe recommendation systems