ARE SKILLS & SENSE-MAKING EBBING AWAY?

In Part 1, we described competitive absorption – the tendency for AI productivity gains to flow to customers rather than to the firms that invested in the technology.

We looked at fee pressure, margin erosion and hiring collapse. The evidence found is quantitative and the pattern is well-documented: innovators historically captured 2.2% of the value their innovations created. For generative AI, the ratio appears even more extreme.

And, of course, competitive absorption is uncomfortable. But it’s visible. It shows up in contracts, in hiring data and in frozen salaries. A leadership team that understands it can plan around it.

Two other costs of AI abundance don’t show up in any dashboard. Which makes them harder to measure, slower to surface and, ultimately perhaps, more consequential for the organisations that miss them.

---

Capability erosion – human skills degrade

You’re driving somewhere you’ve been dozens of times. The sat-nav is on. You follow it. And you realise, halfway through, that you have no idea where you actually are.

Eleanor Maguire at University College London spent years studying why. Her team scanned the brains of London taxi drivers – people who spend three to four years memorising 25,000 streets - ‘The Knowledge’ – and found their posterior hippocampi were measurably larger than non-drivers. A longitudinal study tracked 79 trainees over four years and confirmed the direction of causation: the training caused the growth, not the other way around. Only qualified trainees showed the structural change.

Then Louisa Dahmani and Véronique Bohbot at McGill University studied what happens when GPS replaces that process. They measured GPS users’ hippocampal-dependent spatial memory, including a subset tracked over three years. The finding was dose-dependent: the more someone relied on GPS, the worse their spatial memory became. GPS doesn’t just spare us the effort of navigating. It shifts the brain from a spatial strategy to a stimulus-response strategy, actively suppressing the formation of cognitive maps.

Which means we’re trading the ability to navigate for the convenience of being navigated. For most of us, that’s a bargain we’re happy to take again. Nobody mourns their declining sense of direction when the alternative is arriving on time.

The question is what happens when the same dynamic operates in domains where the eroded skill actually matters.

In 2023, Thomas Dratsch and colleagues at three German university hospitals published a study in Radiology – one of the discipline’s top journals – testing what happens when AI gives radiologists the wrong answer.

The revealed numbers create thoughts worth sitting with.

Very experienced radiologists – more than fifteen years in the field on average – scored 82.3% accuracy when the AI suggested the correct category. When it suggested the wrong one, their accuracy collapsed to 45.5%.

Inexperienced radiologists went from 79.7% to 19.8%.

Roll that around your head a bit.

Done? Well, a follow-up study by Rezazade Mehrizi and colleagues made it worse. Neither providing explanations for the AI’s reasoning - nor making radiologists aware of the possibility of AI error - overcame the effect. Knowing that AI could be wrong didn’t help. Understanding why the AI reached its conclusion didn’t help. The bias persisted.

You might reasonably object: this is about unreliable AI. As AI improves, the problem resolves. But a 2025 study in the Lancet Gastroenterology and Hepatology suggests otherwise. Krzysztof Budzyń and colleagues tracked 19 experienced endoscopists – each with more than 2,000 colonoscopies – across four centres. After a period of routine AI-assisted colonoscopy, they measured the endoscopists’ performance without AI. Adenoma detection rates had fallen from 28.4% to 22.4%. A 20% relative decline – from normal use of reliable AI.

The authors described it as ‘the first real-world clinical evidence for the phenomenon of deskilling, potentially affecting patient-related outcomes.’ The AI wasn’t wrong. It was working as designed. And the humans relying on it were quietly losing the ability to do the job without it.

Eric Topol, the cardiologist and digital health researcher at Scripps, has been tracking a parallel pattern in gastroenterology. He describes a ‘never-skilled’ cohort – clinicians who trained with AI from the start and never developed the foundational diagnostic capability their predecessors built through years of unassisted practice. His analogy comes from an unexpected place: Quebec. ‘You have to have deliberate intent to preserve your language,’ he said. ‘Otherwise, it’s crowded out.’ The same applies to clinical skill.

Nita Farahany, a neuroethicist at Duke, frames the mechanism more precisely. The distinction that matters, she argues, is between the impact on the receiver and the impact on the contributor. If AI writes something that is identical to that which a junior analyst would have written, the client receiving the report doesn’t notice. The work product is the same. But the analyst who didn’t write it loses the developmental process. The act of contributing, for Farahany, is the act of becoming. ‘The brain needs that act,’ she says.

Stephen Casner and colleagues at NASA Ames studied airline pilots and found a distinction that sharpens this for enterprise leaders. Automation preserved what they called ‘by hand’ skills – the motor and procedural tasks of flying. What it degraded were the ‘by mind’ skills – situational awareness, judgement, the ability to recognise when something has gone wrong. Pilots who flew in maximum automation 99.6% of the time – fewer than five hours of manual flying per year – could still operate the controls. But they struggled to know when to.

It is, therefore, reasonable to posit that AI in knowledge work is currently displacing ‘by mind’ skills. Judgement. Diagnosis. Evaluation. Pattern recognition under uncertainty. And here is where the trap would close: the skill required to oversee AI – to catch its errors, to know when its confident output is wrong – would be the same skill that atrophies through AI use.

Every metric that measures performance with AI present shows improvement. But is anyone measuring performance with AI is then absent? Is the gap between the two widening silently?

---

Sense-making collapse – more information, worse decisions

This section which follows has the weakest evidence of this series, so let's bear that caveat in mind. What follows points toward a pattern rather than proving one. But the pattern, if it holds, also carries the highest stakes.

In 1525, Erasmus wrote: ‘Is there anywhere on earth exempt from these swarms of new books?’ The printing press had been operating for roughly 75 years. The complaint was not about the existence of books. It was about the volume of unreliable, contradictory and low-quality material that the new technology had made cheap to produce.

Elizabeth Eisenstein, the pre-eminent historian of printing, documented how the press amplified contradictory and erroneous material alongside the scientific works we now celebrate.

The numbers are striking. In 1640, England produced 848 publications. In 1641, it produced 2,042. But by the end of the first the Reformation’s first decade six million had proliferated. The information environment transformed faster than the institutions could filter it.

And filtering did not catch up quickly. Widespread printing arrived in the 1450s. The Royal Society was founded in 1660. Philosophical Transactions began in 1665. Formalised referee reports appeared in the 1830s. The term ‘peer review’ wasn’t coined until the 1970s. Roughly 150 to 200 years between information abundance and effective institutional filtering.

This does not prove that AI will produce a comparable gap. It suggests that information abundance is not necessarily well-correlated to an abundance of understanding and effective knowledge. And that the distance between them can persist for longer than anyone might reasonably expect.

The contemporary signals of history rhyming are suggestive rather than conclusive. Security operations centres receive an average of 4,484 alerts per day, with 67% ignored due to false positives. Clinicians override 90–96% of all clinical decision support alerts. A 2014 study found 66 ICU beds generating more than two million alerts in a single month – 187 per patient per day. The information is abundant. But the signal is at risk of being hidden inside it.

Iain McGilchrist, the psychiatrist and philosopher, made a provocation on the Jim Rutt Show that is worth holding lightly rather than gripping tightly. In 1965, he observed, America had 9% college graduates and no internet. It managed to produce functional collective discourse on complex topics. In 2026, with 35% college graduates and universal information access, the quality of that discourse has, by many measures, declined. ‘Information,’ he said, ‘has nothing whatever to do with intelligence or wisdom.’

McGilchrist’s observation has a hundred confounding variables and should be treated accordingly. And the printing press analogy, which rests on firmer historical ground, contains its own honest complication. The same period of epistemic chaos that saw pamphlet wars, conspiracy and information overload also produced the Scientific Revolution. Abundance displaced some forms of sense-making and enabled others. AI may well do the same.

The defensible claim is narrower than ‘abundance degrades reasoning.’ The claim is that there is a temporal gap between old forms of sense-making breaking down and new ones emerging to make functional use of the new technology. We are, it seems likely, early in that gap. Organisations have sense-making processes – reporting structures, analytical workflows, decision-support systems – designed for information scarcity. AI is making information abundant faster than organisations can redesign the way they make sense of it. What happens during the gap matters.

---

The underlying pattern

Three costs, then – plus the demand creation channel that Jevons described and everyone is still talking about. They are all ways in which the new abundance of intelligence may deliver negative consequences.

Each one operates through a different mechanism in a different domain. Competitive absorption is economic. Capability erosion is cognitive. Sense-making is social and institutional. Demand creation is behavioural. They are not, in any strict sense, the same phenomenon.

But they do share a structural feature that, we believe, makes them worth examining together.

Each one describes AI creating measurable gains in one dimension while degrading something in a different dimension that the gain metric doesn’t capture. Productivity metrics don’t capture margin erosion. AI-assisted accuracy metrics don’t capture unassisted human accuracy. Information volume metrics don’t capture decision quality. The trap tightens precisely as the dashboards improve.

Of course, they are not fixed in impact and each can be mitigated.

Competitive absorption would weaken if AI’s consumer-surplus ratios compressed sharply over time – that would mean the 14:1 ratio captured in 2024 was an unusual moment rather than a structural feature of the technology.

Capability erosion would weaken if AI systems can be reliably designed to preserve rather than degrade skill. Early evidence from radiology suggests this may be possible where AI is deliberately built as a scaffold rather than a replacement. If that design pattern generalises, the clinical and aviation evidence may prove less universal than one may suppose. And it would be weakened if new sense-making institutions emerge rapidly – that would mean the printing press analogy overstates the timeline.

Finally, there are three things we looked for and didn’t find - but we feel are also worth naming.

There are no peer-reviewed study that directly measures the speed of AI competitive absorption in enterprise settings.

There is also no controlled study we could find that measures willingness to pay premiums for human-maintained expertise versus AI-only delivery.

And no organisation we found is explicitly pricing what might be called the Baumol inversion – the dynamic where AI makes automatable work cheaper and non-automatable work relatively more valuable.

Regardless, we believe there is enough evidence to suggest the pattern is real. The question is how much of the current AI moment it explains. The answer is almost certainly more than conventional AI business cases assume, and less than a pessimist would like to believe.

And we'll gaze into that soon – it’s the territory of Part 3.