In the lite version of Cramer-Shoup we have a group $G$ with generators $g_1$ and $g_2$, private key $a_1, a_2, b_1, b_2$, and public key $A = g_1^{a_1} g_2^{a_2}$, $B = g_1^{b_1} g_2^{b_2}$. Encryption works by picking a random integer $r$ and sending the tuple $(g_1^r, g_2^r, A^r m, B^r)$. When decrypting we check the $B^r$ portion of the ciphertext is correct and, if so, obtain $m$ by calculating the inverse of $A^r$ and multiplying $A^r m$ with it.
Consider the following formulation of DDH: for random group elements $g_1, g_2, u_1, u_2$ and random integer $r$, $(g_1, g_2, u_1, u_2)$ is indistinguishable from $(g_1, g_2, g_1^r, g_2^r)$. Cramer-Shoup lite can be shown to be IND-CCA1 secure under this assumption. The proof goes by employing a hypothetical CCA1 attacker $\mathcal{A}$ which breaks Cramer-Shoup lite to break DDH. Our challenge instance is a quadruple $(g_1, g_2, u_1, u_2)$. We simulate the IND experiment for $\mathcal{A}$ with our own generated private and public keys; thus we can also simulate its decryption oracle since we have full knowledge of the scheme's private key. When $\mathcal{A}$ commits itself to a message pair $m_0, m_1$, we send it one of this messages (randomly chosen) encrypted, but with $g_1^r$ replaced with $u_1$ and $g_2^r$ with $u_2$. Thus $\mathcal{A}$ receives $(u_1, u_2, u_1^{a_1} u_2^{a_2} m, u_1^{b_1} u_2^{b_2})$, which is a valid encryption iff $u_1 = g_1^r$ and $u_2 = g_2^r$ for the same exponent $r$. $\mathcal{A}$'s response (ciphertext is valid or not) is correct with non-negligible probability, so we can distinguish the DDH quadruples with non-negligible probability.
I am having trouble understanding why the argument fails for IND-CCA2. Is it because $\mathcal{A}$ can use $u_1$ and $u_2$ in subsequent decryption queries?