The Oracle That Contradicts Itself

The Wall Street Journal reports a finding so perfectly absurd it almost feels scripted: several advanced AI models, asked to rank which jobs are most vulnerable to automation, each produced a different list. One highlighted administrative roles. Another singled out customer service. A third spread the risk more evenly. The models could not agree on which occupations they themselves were supposed to eliminate. It is the intellectual equivalent of a weather app that forecasts sun, rain and a meteor strike within the same 24-hour window, then expects you to pack accordingly.

Now, you might ask: why should we expect these systems to agree? They are trained on different data, use different architectures, and encode different biases. That is precisely the problem. An entire industry has spent the last two years telling us that AI will reshape the labour market with surgical precision, that executives need to prepare for mass displacement, that workers should retrain yesterday. Yet when you ask the actual technology for a straight answer, it offers a shrug in multiple languages. The oracle cannot decide which prophecy to deliver, so it delivers all of them.

Two weeks ago I wrote about Cognizant's 'AI-led restructuring' — a phrase that allowed the company to cut five percent of its workforce while claiming technological inevitability. The week before that, Sam Altman warned against 'AI washing' even as his own products powered the layoff narrative. Now we have a fresh layer of absurdity: the same models that justify the cuts cannot agree on which roles to target. This is not a bug; it is a windfall for executives. Disagreement among models means you can cherry-pick whichever prediction suits your agenda. Want to gut the back office? Cite model A. Want to purge customer service? Cite model B. The inconsistency is not a weakness — it is a buffet.

The deeper problem is that we are asking the wrong question entirely. 'Which jobs will AI destroy?' presupposes that destruction is a fixed property of the technology rather than a choice made by people in charge. The models do not reveal a deterministic future; they reflect the assumptions baked into their training data and the priorities of those who deploy them. A model trained on corporate efficiency metrics will naturally flag roles deemed 'redundant.' A model trained on historical layoff patterns will extrapolate past cuts. Neither model is predicting the future. They are projecting the biases of the past onto an uncertain tomorrow, and executives are happy to treat that projection as destiny.

Meanwhile, the actual labour market offers a more mundane story. Microsoft published a report noting that very few jobs are being entirely automated away. McKinsey's Alexis Krivkovich confirmed that complete displacement remains rare. But those headlines do not serve the narrative. The narrative requires urgency, crisis, the sense that you must act now or be left behind. Inconsistent AI predictions feed that urgency precisely because they cannot be falsified. If one model says admin roles are doomed and another says they are safe, the prudent (and panicked) response is to assume the worst. The middle ground — that perhaps we should wait and see — is not available when the tech press and consultancy firms are competing to scare you first.

Land the ending. The only thing these models agree on is that they should be trusted. They offer no consensus on which jobs go, but they unanimously recommend that you listen to them. The irony is magnificent, and it is also dangerous. If we cannot even get the technology to produce a consistent forecast, we have no business using it to justify mass layoffs. The jobs that should really be on the chopping block are the ones that belong to the executives who deploy inconsistency as a management strategy. But do not hold your breath. The oracle has spoken, and it says different things to different people. That, apparently, is by design.