Can We Train Machines to Be Better Than Us?

To think clearly about whether we can train better AI, it helps to understand specifically how bias enters the training process, because there is more than one pathway and they require different responses. The most discussed form is historical bias, where the training data reflects past discrimination that we would now consider unjust. The classic example is a hiring model trained on a company's historical hiring decisions, which may systematically underrepresent certain groups not because those groups are less qualified but because they faced barriers that kept them out of the applicant pool or caused their applications to be screened out by biased human reviewers. The model learns to replicate that outcome because replication of historical patterns is, in a narrow technical sense, exactly what it is optimized to do.

But there are subtler forms that get less attention. Representation bias arises when the training data simply does not include enough examples from certain groups for the model to learn accurate patterns about them. Medical AI trained primarily on data from well-resourced healthcare systems in high-income countries may perform significantly worse for populations whose health presentations, comorbidities, and treatment responses differ systematically from that training population. Facial recognition systems trained predominantly on lighter-skinned faces have demonstrated dramatically higher error rates for darker-skinned faces, with real consequences in contexts where those systems are used for identification. The model did not learn to discriminate intentionally. It learned from data that was unrepresentative, and the gaps in its performance track the gaps in its training.

There is also what I think of as measurement bias, which is perhaps the most insidious because it hides inside the very metrics we use to evaluate whether a model is performing well. If we measure the success of a predictive policing model by how accurately it predicts arrests, and arrests themselves reflect biased policing patterns, then a model that is highly accurate by that metric is accurately predicting bias rather than accurately predicting criminal behavior. The evaluation looks rigorous. The outcome replicates and potentially amplifies the original injustice. This is why the question of what we are measuring and whether that measurement captures what we actually care about is not a secondary technical concern. It is the central ethical question in AI development, and it does not have a technical answer.

Here is where I want to push back against a kind of resigned fatalism that sometimes settles over this conversation. Yes, the data we have is compromised. Yes, the historical record we are training on is shot through with human failure. But the conclusion that AI is therefore inevitably biased and there is nothing to be done is not supported by the evidence of what is actually possible with deliberate effort. There are research teams and organizations doing serious work on bias detection, dataset curation, algorithmic fairness constraints, and evaluation frameworks that measure outcomes rather than just accuracy. This work is hard, it is often underfunded relative to its importance, and it is genuinely technically complex. But it is producing results, and those results suggest that the gap between the AI we are building and the AI we could build with genuine commitment is not fixed.

The most promising approaches tend to involve treating fairness not as a post-hoc correction but as a design requirement from the beginning. That means interrogating training data before it enters a model, asking explicitly whose experiences are represented and whose are missing, and making deliberate choices about how to address those gaps. It means defining success metrics that reflect real-world equity rather than just statistical performance on a benchmark. It means involving communities that are likely to be affected by a system in its design and evaluation, not as a consultation exercise but as a genuine accountability mechanism. And it means building the institutional capacity to monitor deployed systems over time, because bias can emerge or shift as the world changes in ways that the training data did not anticipate.

None of this is simple, and I want to be honest about that. Training AI to be better than us in the relevant sense is not primarily a machine learning problem. It is a problem of human values, human governance, and human willingness to invest in getting this right even when getting it wrong is commercially cheaper and less contentious. The technology could do better. Whether we will choose to make it do better is a question about us, not about the machines. And the answer to that question, which is still genuinely open, depends on whether enough people in enough positions of influence come to understand that the cost of biased AI at scale is not just a technical error rate. It is, in the most direct possible sense, the perpetuation of injustice with greater speed and at greater scale than the biased humans it was supposed to replace.

I want to close with something more constructive than a catalogue of problems, because I think the constructive case is both real and underappreciated. The aspiration to train machines that are better than us is not naive. In fact, there are specific domains where AI is already demonstrating the ability to be more consistent, more equitable, and less subject to the worst human tendencies than human decision-makers in the same contexts. Studies comparing AI-assisted medical diagnosis to unassisted diagnosis have found that in certain conditions, the AI shows less variance across demographic groups than human physicians, who bring unconscious assumptions about which patients are likely to have which conditions. That is a real form of improvement that is worth pursuing deliberately rather than hoping will emerge accidentally.

The path toward AI that is genuinely better than us in morally meaningful ways runs through a few clear commitments. Radical transparency about training data, about what is in it, who it represents, and what it excludes, so that the biases baked in are at least visible rather than hidden. Genuine diversity in the teams building and evaluating these systems, because the blind spots in AI development track the blind spots of the people doing the developing, and those blind spots are narrower when the team is broader. Regulatory frameworks that hold deployed systems accountable for disparate outcomes, not just for input data quality, because the test of whether a system is fair is ultimately what it does to real people in real situations. And a cultural shift in the field toward treating fairness not as a constraint on performance but as a dimension of it.

The question of whether we can train machines to be better than us is ultimately a question about whether we are willing to be better than we have been. AI learns from what we give it. If we give it our history uncritically, it will replicate our history. If we give it our aspirations, our hard-won understanding of where we have failed and what fairness actually requires, we create at least the possibility of something that reflects the best of what we are capable of rather than the average of what we have done. That is a harder project than building a model that performs well on a benchmark. But it is the project that actually matters. And given how much influence these systems are going to have over how decisions get made in the world, it is the project we cannot afford to treat as optional.