If they only used live PhD-level human discourse, over topics that had been, or can be peer-reviewed in advance[0] before finding its way into a database, it might be more likely to seem more intelligent than training on average everyday data.<p>Don't get me started on how much less advanced you should expect if random internet data is given any weight at all.<p>[0] Rehearsal might be good, it could even include PhD's in Education since teaching the machine could almost be considered equivalent to professors preparing a curriculum for teaching regular students.