Of the hundreds of z ≳ 6 quasars discovered to date, only one is known to be gravitationally lensed, despite the high lensing optical depth expected at z ≳ 6. High-redshift quasars are typically identified in large-scale surveys by applying strict photometric selection criteria, in particular by imposing non-detections in bands blueward of the Lyman-α line. Such procedures by design prohibit the discovery of lensed quasars, as the lensing foreground galaxy would contaminate the photometry of the quasar. We present a novel quasar selection methodology, applying contrastive learning (an unsupervised machine learning technique) to Dark Energy Survey imaging data. We describe the use of this technique to train a neural network which isolates an ‘island’ of 11 sources, of which seven are known z ∼ 6 quasars. Of the remaining four, three are newly discovered quasars (J0109−5424, z = 6.07; J0122−4609, z = 5.99; J0603−3923, z = 5.94), as confirmed by follow-up and archival spectroscopy, implying a 91 per cent efficiency for our novel selection method; the final object on the island is a brown dwarf. In one case (J0109−5424), emission below the Lyman limit unambiguously indicates the presence of a foreground source, though high-resolution optical/near-infrared imaging is still needed to confirm the quasar’s lensed (multiply imaged) nature. Detection in the g band has led this quasar to escape selection by traditional colour cuts. Our findings demonstrate that machine learning techniques can thus play a key role in unveiling populations of quasars missed by traditional methods.