We consider a Cahn–Hilliard equation in a bounded domain Ω in {\mathbb{R}^{n}} over a time interval {(0,T)} and discuss the backward problem in time of determining intermediate data {u(x,\theta)}, {\theta\in(0,T)}, {x\in\Omega} from the measurement of the final data {u(x,T)}, {x\in\Omega}. Under suitable a priori boundness assumptions on the solutions {u(x,t)}, we prove a conditional stability estimate for the semilinear Cahn–Hilliard equation\lVert u(\,\cdot\,,\theta)\rVert_{L^{2}(\Omega)}\leq C\lVert u(\,\cdot\,,T)%
\rVert_{H^{2}(\Omega)}^{\kappa_{0}},and a conditional stability estimate for the linear Cahn–Hilliard equation\lVert u(\,\cdot\,,\theta)\rVert_{H^{\beta}(\Omega)}\leq C\lVert u(\,\cdot\,,T%
)\rVert_{H^{2}(\Omega)}^{\kappa_{1}},where {\theta\in(0,T)}, {\beta\in(0,4)} and
{\kappa_{0},\kappa_{1}\in(0,1)}. The proof is based on a Carleman estimate with the weight function {\mathrm{e}^{2s\mathrm{e}^{\lambda t}}} with large parameters {s,\lambda\in\mathbb{R}^{+}}.