Australia’s climate is highly variable, which poses considerable challenges for climate assessments over regional and local scales. This thesis focuses on the effects of internal variability on temperature extremes over Australia, including historical evaluation, future projections, and attribution. Conclusions are mainly based on state-of-the-art climate models in Phase 6 of the Coupled Model Intercomparison Project (CMIP6), in which several initial-condition large ensembles (LEs) are analysed to assess the effects of internal variability. This thesis investigates whether model estimates of internal variability are robust and how this can affect future risk assessments. First, evaluation of temperature extremes indicates modest improvement in CMIP6 compared to previous generation models (CMIP5). Model ranges for some extreme indices in CMIP6 tend to be narrower, implying a reduction in model uncertainty. For different LEs, model differences in internal variability indicate that the metrics used to examine model performance in simulating climatology are not suitable to examine intrinsic variability. Second, in addition to projected changes, the time of emergence (TOE) concept is introduced which indicates when the trend in a climate variable emerges above the “noise” of natural climate variability. A “warm-gets-warmer” pattern exists for some extremes over Australia and tropical regions usually show the largest warming. In comparison with CMIP5, partly due to the higher climate sensitivity in some models, CMIP6 shows stronger warming and exhibits earlier TOE. The results also signify the role of internal variability in the noise, further influencing the spread of TOE. Third, compared to hot extremes, it is found that the long-term changes of cold extremes are less evident to changes in anthropogenic aerosols. Moreover, based on a non-stationary generalised extreme value distribution, results indicate that greenhouse gases (anthropogenic aerosols) can increase (decrease) the occurrence probability of hot extremes while the converse is true for cold extremes. Uncertainty in attribution statements can be further reduced by better representing internal variability. To enable more robust climate assessments, associated physical mechanisms and observationally-based Large Ensembles, in which internal variability from observations is sampled to create surrogate realizations, need to be further developed and understood.