This study set out to establish how well the artificial intelligence protein-structure prediction program, AlphaFold2, could recognise and model leucine zipper (L-zip) dimers. The ability to use AI programs like AlphaFold2 to accurately predict protein structures has the potential to revolutionize the structural biology field (Evans et al., 2022; Jumper et al., 2021). However, like any new tool, it needs to be tested and properly understood to produce meaningful results. We used the Activator Protein 1 (AP-1) transcription factors FosB and JunD as a case study. AP-1 transcription factors are a ubiquitous family of homo or heterodimers with a coiled-coil L-zip dimer interface. A basic fork-shaped region at the N-terminus of the L-zip binds into the major groove of DNA, recognising palindromic consensus sequences (Yin et al., 2017; Bejjani et al., 2019). FosB and JunD have both been very well studied and characterised structurally (Eckert et al., 2013; Kumar et al., 2022; Yin et al., 2017), which makes them an excellent case study to test and analyse the quality of leucine-zipper constructs produced in AlphaFold2. We found that AlphaFold2 could accurately identify and model L-zip dimers, but could not recognise when clashing charges within the L-zip would make the dimer unstable. We designed synthetic L-zip peptide pairs to be highly energetically unstable, and found that AlphaFold2 still predicted they would form L-zips with high confidence scores (mean: 98.04%). This study shows that AlphaFold2 excels at pattern recognition, but is weaker at using “physics-based methods” to accurately refine protein structures (Bouatta et al., 2021).