Here’s a solution I’ve come up with that expands a bit on Erik Gillespie’s. It does exactly what you asked for and it worked for me.

Using Jackson 2.9

@JsonDeserialize(using = CustomDeserializer.class)
public abstract class BaseClass {

    private String commonProp;
}

// Important to override the base class' usage of CustomDeserializer which produces an infinite loop
@JsonDeserialize(using = JsonDeserializer.None.class)
public class ClassA extends BaseClass {
    
    private String classAProp;
}

@JsonDeserialize(using = JsonDeserializer.None.class)
public class ClassB extends BaseClass {
    
    private String classBProp;
}

public class CustomDeserializer extends StdDeserializer<BaseClass> {

    protected CustomDeserializer() {
        super(BaseClass.class);
    }

    @Override
    public BaseClass deserialize(JsonParser p, DeserializationContext ctxt) throws IOException, JsonProcessingException {
        TreeNode node = p.readValueAsTree();
        
        // Select the concrete class based on the existence of a property
        if (node.get("classAProp") != null) {
            return p.getCodec().treeToValue(node, ClassA.class);
        }
        return p.getCodec().treeToValue(node, ClassB.class);
    }
}

// Example usage
String json = ...
ObjectMapper mapper = ...
BaseClass instance = mapper.readValue(json, BaseClass.class);

If you want to get fancier, you can expand CustomDeserializer to include a Map<String, Class<?>> that maps a property name that, when present, maps to a specific class. Such an approach is presented in this article.

Update

Jackson 2.12.0 gets Polymorphic subtype deduction from available fields which adds @JsonTypeInfo(use = DEDUCTION)!

AsDeductionTypeDeserializer implements inferential deduction of a subtype from the fields. As a POC not intended for merging, tere’s an amount of cut’n’paste code etc but I thought a functional PR would be the best basis for discussion of something I write out of interest.

It works by fingerprinting the full set of possible fields of each subtype on registration. On deserialisation, available fields are compared to those fingerprints until only one candidate remains. It specifically only looks at immediate-child field names as is immediate-child values are covered by existing mechanisms and deeper analysis is a much more imposing ML task not really part of Jackson’s remit.

By the way, there’s a (now closed) Github issue requesting this here: https://github.com/FasterXML/jackson-databind/issues/1627