I've been watching something unsettling unfold across boardrooms and leadership teams.
AI tools are getting faster, more fluent, and more confident in their outputs. And that confidence is becoming dangerous.
Because here's what most executives don't realise: GPT-4o has a hallucination rate of approximately 45% in certain setups. OpenAI's reasoning-focused o3 series? Between 33-51% on factual benchmarks. More than double earlier models.
The problem isn't that AI makes mistakes. The problem is that it makes them confidently, fluently, and persuasively.
Knowledge workers now spend an average of 4.3 hours per week fact-checking AI outputs. Yet 47% of enterprise AI users admitted to making at least one major business decision based on hallucinated content in 2024.
This isn't a technical problem waiting for a technical solution. This is a judgement problem that's getting worse as the technology improves.
The Illusion of Accuracy
AI-generated content arrives with all the aesthetic markers of quality.
Clean structure. Confident tone. Professional formatting. Complete sentences that flow beautifully into one another.
Your brain registers these signals and makes a snap assessment: this looks right, so it probably is right.
But here's what's actually happening beneath the surface. Current AI training and evaluation procedures reward guessing over acknowledging uncertainty. The models are incentivised to provide an answer, any answer, rather than admit they don't know.
The result? Fiction presented as fact, delivered with the same polish as genuine insight.
In 2025 alone, judges worldwide issued hundreds of decisions addressing AI hallucinations in legal filings. That accounts for roughly 90% of all known cases of this problem to date. Judges said these errors waste scarce time and resources.
If the legal system is struggling with this, your organisation is too.
Speed Without Substance: The Productivity Paradox
Nearly 90% of firms report that AI has had no impact on employment or productivity over the last three years, despite widespread adoption.
Let that sink in.
About two-thirds of executives reported using AI, but that usage amounted to only about 1.5 hours per week. And 25% of respondents reported not using AI in the workplace at all.
This mirrors a historical pattern. Following the advent of transistors and microprocessors in the 1960s, productivity growth actually slowed. It dropped from 2.9% between 1948 and 1973, to 1.1% after 1973.
Nobel laureate Robert Solow observed: "You can see the computer age everywhere but in the productivity statistics".
We're living through the same paradox now. Research shows that AI adoption tends to hinder productivity in the short term. Organisations that adopted AI for business functions saw a drop in productivity of 1.33 percentage points.
The verdict from enterprise leaders is clear: AI is creating a productivity paradox. It's speeding up work whilst quietly adding more of it, with overall difficulty proving out real productivity benefits.
Why This Happens
AI generates outputs faster than humans can evaluate them.
You ask for a strategic brief. It arrives in 30 seconds. Looks professional. Reads well. So you move forward.
But you didn't spend the time you would have spent if a human had written it. You didn't question the assumptions. You didn't test the logic. You didn't verify the facts.
The speed of generation inversely correlates with the time humans spend evaluating quality. That creates a systematic blind spot where bad outputs slip through because they "feel right".
Human Judgement as Competitive Infrastructure
New research shows that human experience and judgement are still critical to making decisions, because AI can't reliably distinguish good ideas from mediocre ones or guide long-term business strategies on its own.
A Harvard Business School study with 640 entrepreneurs found no statistical difference in business performance when given access to AI assistants. This suggests that investment in support, training, and decision-making frameworks may be as important as access to AI tools themselves.
The pattern is consistent across industries.
A randomised controlled trial with 5,000+ agents at a U.S. tech support desk delivered a 35% throughput lift for bottom-quartile reps but almost no gain for veterans. This demonstrates that AI's productivity gains are highly context-dependent, varying significantly by user skill level and task complexity.
Even more revealing: human-AI collaboration often underperforms either agent working independently, except in creative tasks.
For executives, the implications are clear. Only 20% of respondents believe their organisations excel at decision-making. The majority state that the time they devote to decision-making is ineffective.
For managers at an average Fortune 500 company, this could translate into more than 530,000 days of lost working time and roughly $250 million of wasted labour costs per year.
The Deskilling Spiral
Here's the hidden risk nobody's talking about.
Typically, the better someone gets at using a technology, the more efficient they become. But with AI, as you get more proficient, you start to understand more about the ways in which the tech can go wrong.
This creates a competence trap. AI tools seem to disproportionately benefit stronger users who can better judge which of the AI's suggested angles are more likely to prevail.
But what happens when your team stops practising the craft because AI does it faster?
They lose the expertise needed to judge quality. You end up in a deskilling spiral where automation erodes the very judgement it depends on.
The executive mandate is becoming clear: people often confuse leadership intuition with full AI substitution. However, the most effective leaders understand that AI does not change the nature of accountability and judgement.
The winning formula? Let AI handle the speed and scale. Allow your leadership to manage meaning, values, and direction.
What This Means for Your Organisation
Although many organisations are investing in AI, only 13% of companies are fully prepared to deploy it. This means that many systems are still experimental, siloed, or insufficiently embedded in workflows.
The risk is real. Hiring practices, compliance decisions, and crisis response cannot be left entirely to automated systems. These critical situations require judgement informed by ethics, empathy, and a thorough understanding of the context.
Your competitive advantage isn't in how fast you can generate content. It's in how well you can evaluate it.
Three Practical Shifts
1. Reverse the interaction pattern
Stop asking AI for answers. Start asking it to ask you questions first.
Before you generate that strategic document, have the AI gather requirements. What's the actual objective? Who's the audience? What constraints matter? What assumptions need testing?
This simple reframe transforms AI from an answer machine into a requirements-gathering tool. Better questions always precede better solutions.
2. Build evaluation infrastructure
Organisations that invest in developing evaluation frameworks, quality standards, and review processes will outcompete those focused solely on generation capacity.
The infrastructure of judgement becomes the competitive moat. You need taste platforms, not just content platforms.
Create rubrics. Establish testing protocols. Implement peer review systems. Make evaluation a discipline, not an afterthought.
3. Protect domain expertise
The ability to evaluate quality requires domain expertise, contextual understanding, and subjective assessment that cannot be automated.
If your team stops doing the work, they stop building the expertise needed to judge whether AI's work is any good.
Maintain practice. Keep your people sharp. Ensure they can still do the thing they're now evaluating.
The Real Competition
In a world where generation becomes commoditised, the ability to evaluate quality becomes the scarce resource.
Knowing whether output is "good," "accurate," or "worth putting your name on" requires judgement that can't be automated.
This positions human discernment as the bottleneck that determines value.
Your competitors are racing to generate faster. You should be racing to evaluate better.
Because in the end, the organisation that ships the highest-quality decisions wins. Speed without substance is just noise.
And your judgement is what separates signal from noise.
