Shield yourself: Judgement is the last human bastion against AI bots ready to take jobs away

The fact is that wherever there is no definitive way to verify what is what, we are forced to depend on the consensus of experts to know what the truth is.

Earlier this month, in a paper I wrote for Observe Research Foundation titled The Absorption Frontier, I made the argument that the reliability of AI systems was critically dependent on the extent to which their outputs can be verified.

While verifiability is often deterministic and empirical, there are situations—such as was the case with the Prague Consensus—that call for it to be done by humans. If AI absorption depends on the verifiability of outputs, the domains where machines will struggle are those in which humans will remain relevant.

Coding is the best example of deterministic verification—where things either work or do not. Since software outputs can be checked instantly, AI can learn from its mistakes as it codes, improving the code it generates until it eventually excels at programming.

In other, more empirical domains, truth is only observable long after the decision has been made—as with the creditworthiness of borrowers or the efficacy of medical treatment. But even when verification is delayed, it should be possible to design AI to learn and improve.

This is a lot harder to do in consensus domains, where truth is whatever qualified experts agree it is (as with the demotion of Pluto) or in adversarial domains, where outcomes are reached in response to the arguments of counterparties.

In 1950, the mathematician John Nash showed that whenever two parties stand to gain by agreeing, there is always a bargain they can arrive at—but what exactly this negotiated solution ultimately ends up being depends on inputs only the parties themselves can supply.

There is, as a result, no ground truth against which to check what a group of experts will agree on tomorrow or what the other side of a negotiation will be willing to accept.

That said, we must not be so bold as to presume that AI systems will never be able to do this. In 2017, a model trained on two centuries’ worth of US Supreme Court decisions predicted case outcomes with over 70% accuracy. But as good as these systems are, they are, at best, making a calculated guess about a decision that has not yet been made. In every instance, it will be the judges who actually decide, and since we have no way to look into their minds, they could still rule otherwise.

It is also tempting to believe that it should be possible to train AI agents to negotiate transactions on our behalf. For as long as there have been lawyers, we have engaged them to do precisely that—negotiate commercial and other agreements on our behalf. Surely, it should be possible to get AI to do the same.

But as anyone who has done this will tell you, lawyers have to keep going back to clients for instructions, and every time the other side makes an unanticipated offer, it is the client who decides whether to accept it or push back. If the AI agents that negotiate for me behave the same way, or end up doing so, the inputs necessary for a Nash bargain will still come from me.

There will be domains where a conclusion must be reached afresh each time by the people in the room, based on what each of them brings to it. And since that conclusion cannot exist until it is reached, no AI system can learn from it in advance. Since neither consensus nor adversarial domains offer ground truths that AI can rely on to train and improve, might these be the domains where human judgment will always remain relevant?

Last month, I pointed out in this column that we train young professionals by allowing them to learn by doing. This is why skilling programmes in most disciplines are designed to have them perform relatively mundane work until they acquire the skills senior professionals need. Since this is also the sort of checkable work that AI does well, the more we incorporate AI into our professional lives, the fewer opportunities our young professionals will get to gain the experience they need to forge careers.

The solution, as I presented it in that article, was to ensure that we reserve some of this sort of work for the young professionals in our organizations—even if that means forgoing some of the efficiency gains that AI makes possible—because we know of no other way to train them.

But if it is clear that there are some domains in which AI will, for the foreseeable future, be unable to match what experienced humans bring to the table, should we not focus our training on building skills in those domains? We should design our training programmes so that young professionals can learn how experts reach agreement on contentious issues and how to hold their ground against pushback.

AI will inevitably get better at anything that can be verified, and will probably do so faster than most of us expect. But the work that will continue to elude AI is the kind that has to be agreed upon or argued over, where people still have to decide among themselves what is true. To remain relevant in the age of AI, these are the skills we will need.

The author is a partner at Trilegal and the author of ‘The Third Way: India’s Revolutionary Approach to Data Governance’. His X handle is @matthan.