It is certainly possible to develop ML models that yield more accurate results than would be possible without ML.
One route to do this is through so-called "Δ-learning" where you use ML to learn a correction to a less expensive, often less accurate level of theory. An example can be found here for thermochemical properties of organic molecules. Somewhat related to this general idea, here is a paper discussing how ωB97X-D/def2-TZVP energies were able to be predicted from semi-empirical GFN1-xTB input features.
Naturally, another route one can take is to use ML with data from experiments, which can yield more accurate results than theory alone. For instance, it is well-established that GGA functionals yield underpredicted band gaps, and prior ML work has been carried out to predict band gaps at greater-than-DFT accuracy with this in mind. Many other studies are of this type, such as this paper on ML models that can be more accurate than TD-DFT for emission wavelengths.
Given a large dataset of inexpensive but somewhat inaccurate data and a smaller dataset of more expensive (or difficult to obtain) but accurate data, one can also use "transfer learning" to develop a ML model that has an accuracy comparable to the high-fidelity reference data. As an example, this work showed that a neural network potential could approach CCSD(T)/CBS accuracy on a dataset that is largely DFT-generated.
It is also possible to use ML models to identify likely issues or errors with a given calculation, as nicely demonstrated in this paper by Kulik and coworkers. Presumably, this could be used to make your calculations more accurate by knowing what calculation failures need to be addressed.