Today's Jagged AI Is Good Enough to Fix Our Worst Decisions

One thought keeps nagging at me: even if AI models stopped improving tomorrow, they could still do a massive amount of good. That has less to do with machine brilliance and plenty to do with how reliably flawed human judgement can be.

Daniel Kahneman spent decades documenting how poorly we make decisions under uncertainty, and the findings are sobering. A famous study of Israeli parole boards found judges were far less likely to grant parole before a food break than straight after one, even if later analysis questioned whether hunger or case scheduling explained the drop. Doctors routinely anchor on the first diagnosis they consider and under-weight contradictory evidence. Recruiters favour candidates who remind them of themselves. None of this is rare; it is everyday professional life.

Kahneman split cognition into System 1, which is fast, intuitive, and pattern-matching, and System 2, which is slow, deliberate, and analytical. Plenty of professional work claims to run on System 2 while actually leaning on System 1. A tired parole judge running on instinct is hardly weighing every nuance of a file, and a GP squeezed into a seven-minute appointment slot barely has the breathing room to deliberate. Working conditions push well-meaning people into mental shortcuts that Kahneman showed to be systematically skewed.

Put current AI models alongside that reality. An algorithm never gets hungry, never has a grim Monday morning, and brings the same energy to a case at 4pm as it did at 9am. Models are far from flawless, of course. They inherit biases from their training data and human labellers, their outputs can vary from run to run, and a 2024 randomised study showed that GPT-4, Claude 2, Gemini Pro, and GPT-3.5 can all fall into anchoring traps of their own. Even so, imperfect software can still be much less noisy and erratic than the tired humans it supports.

Look at how quickly the evidence is moving. A 2024 JAMA Network Open trial, which feels like a hundred LLM-years ago, found that GPT-4 alone scored higher on diagnostic reasoning across clinical vignettes than both unaided doctors and doctors using the model as a support tool. According to an AMA survey of US physicians, clinical adoption of AI tools jumped from 38% in 2023 to 66% in 2024, while the FDA had authorised more than 1,300 AI-enabled medical devices by the end of 2025. Beyond medicine, OpenAI reported that GPT-5.2 Thinking won or tied 70.9% of comparisons against human professionals across its GDPval benchmark of knowledge-work tasks.

I have written before about the double standard we apply when comparing machines with people. In the US alone, NHTSA estimated 39,345 traffic deaths in 2024, most involving human drivers, yet a single autonomous vehicle crash triggers immediate outrage. We demand near-perfection from algorithms while shrugging at chronic human error. Kahneman would have spotted the culprit straight away: loss aversion and availability bias dressed up as caution.

Ironically, the professions fighting hardest against AI disruption are often the ones that need the help most. Law is a cognitive trade fenced off by guild rules. Medicine runs under relentless time pressure that forces System 1 snap judgements onto problems crying out for System 2 care. Both sectors also sit on huge unmet need, from legal deserts where ordinary people cannot afford advice to clinics where a seven-minute GP slot is all a patient gets.

There is a strong Jevons Paradox case here. If AI makes routine legal work faster and cheaper, pent-up demand could bring representation to people currently locked out of the system. If models shoulder more of the initial diagnostic legwork, doctors have a shot at spending proper time on the knotty cases that genuinely rely on human judgement and empathy. The real prize is freeing skilled people to do the parts of the job that actually require a human being.

Psychology remains the biggest hurdle. As Kahneman proved over and over, we are awful at spotting our own biases. Someone who has spent a decade training for a profession will naturally bristle at the suggestion that software might handle routine calls more reliably. That pushback rarely comes from sheer arrogance; it is just how blind we all are to our own mental blind spots.

We shouldn't let great become the enemy of good. Even with all their jagged edges, today's models can help strip some of the noise and bias out of high-stakes decisions. Measuring AI against perfection misses the point; the real test is whether it beats the status quo, which Kahneman spent a lifetime showing is much shakier than we like to pretend.