What the AI hype gets wrong
Separating what the tools do from what the marketing says they do.
If you've followed AI coverage for more than a year, you've seen the pattern. Every few months, a new claim that AI is about to change everything. Every few months, that claim quietly recedes into "well, it's still useful for some things."
This isn't an argument that AI is useless. It really isn't. It's an argument that the hype gets specific things wrong, in ways worth being clear about.
Why the claims keep receding
The pattern behind almost every overstatement is the same: a real result on a narrow, well-defined task gets generalised to work that is neither. Coding benchmarks become "10x productivity". A standardised test score becomes "smarter than the experts". An impressive demo becomes "agents will run your business". The underlying result is usually true. The generalisation almost never is.
It matters because the two are sold together. AI genuinely is faster at drafting, summarising, structuring and translating. It is not better at judgement, at holding context across months, or at being accountable for a decision, and those are the parts of knowledge work that actually decide outcomes.
The claim
"AI is going to replace knowledge work"
"AI agents will run your business"
"Everything gets 10x more productive"
"AI is now smarter than the experts"
"Imagine where it'll be in five years"
"AI will solve [hard problem]"
What's actually true
It takes the drafting parts, not the judgement ones
Bounded automation with a human reviewing works. Autonomy doesn't
10 to 30% on real knowledge work, not 10x
Better at benchmarks, not at novel problems or edge cases
Nobody knows. Five-year predictions are for fundraising decks
Useful inside the work experts do. Not a substitute for them
Stay sceptical of the hype. Stay open to the actual usefulness. Both are true at the same time.
What the hype doesn't mention
Four things that never make it into the marketing. The failure modes are predictable and recurring, hallucination, sycophancy, confident wrongness, stale information, and none of them are going away this year. Most teams adopt AI badly, using free tiers for sensitive work and skipping the verification, so adoption is harder than it sounds. The boring use cases are the valuable ones: email drafts, meeting summaries, restructuring notes, none of which make a good demo. And the competitive advantage is small, because everybody has access to the same tools.
What honest assessment sounds like
Here is the whole pitch. It's a good tool for specific tasks, drafts, summaries, brainstorming, translation, restructuring, and it saves real time on those. It's bad at judgement, current information, and anything requiring lived experience. Used carelessly it produces confident-looking work that is subtly wrong. Used well it makes a real difference to a team's output.
Nothing about transformation. Nothing about the future. Nothing about falling behind. If a piece of AI content can't survive being summarised like that, the content is selling something.
Three things to do this week
01
Discount any number that came from a benchmark
The 10x figures come from narrow coding tasks. They don't survive contact with a real job full of interruptions and judgement.
02
Assume the failure modes are permanent
Hallucination, sycophancy and confident wrongness are not going away this year. Build the checking in.
03
Compete on integration, not on access
Everyone has the same tools. The advantage goes to the team that works them into real workflows.
Conclusion
01
The hype is wrong about scale, speed, autonomy and productivity. Not about usefulness.
02
For most knowledge work the realistic gain is 10–30%, not 10x.
03
The boring use cases are the valuable ones. The demos show the wrong things.
Go to next