AI Guides › Trend Watch
Why Everyone's AI Predictions From Last Year Were Mostly Wrong
By Nigel Guy · 3 min read
Every year brings a fresh round of predictions: what AI will do next, which
jobs will change, which companies will win. They're confident and specific,
they travel well, and almost nobody goes back to check them. If you do, you
tend to find that the misses aren't random. They fall into a small number of
patterns that the prediction format itself encourages — which means you can
use last year's misses to read this year's crop more carefully.
The rule: predictions fail in patterns — wrong timing, straight-line
extrapolation, capability mistaken for adoption, the wrong mechanism, and
incentives dressed up as forecasts — so read any new prediction by asking
which pattern it's most exposed to, and hold it that loosely.
The Five Ways Predictions Miss
- Right direction, wrong timing. The capability arrives, but uptake
lags behind it — habits, budgets, regulation, trust and plain
organisational inertia all slow things down. Occasionally the reverse
happens and something arrives sooner, somewhere nobody was looking.
- Straight-line extrapolation. Recent pace gets assumed to continue
indefinitely. Progress in any field tends to come unevenly — bursts,
plateaus, and sideways moves.
- Capability mistaken for adoption. "AI can now do X" quietly becomes
"everyone will be doing X by next year". Those are very different claims;
see "The Difference Between Adoption And Hype, Measured Honestly".
- The wrong mechanism. The change does happen, but through different
products, different players or different uses than the prediction named.
- Incentive-shaped forecasts. Predictions from people who sell, invest
in, or build an audience around the outcome. That doesn't make them
dishonest, but it does select for boldness over accuracy.
The Prediction Audit
Try this on any set of predictions you saved or remember — including your
own.
| Prediction |
Specific enough to check? |
What actually happened |
Which miss pattern, if any |
|
|
|
|
- Mark whether each one was checkable at all. A surprising number won't
be.
- Score only the checkable ones. Right, partly right, wrong.
- Tag each miss with a pattern. You'll usually see one or two patterns
dominate.
- Note which sources did better. That's the real payoff: calibrated
trust in who to listen to next time.
Keep it private. The point is calibration, not scoring points off anyone.
Reading this year's crop
For each new prediction, ask: Is it specific and dated? What would show it
was wrong? Who benefits if people believe it? Does it confuse "can" with
"will"? A prediction that survives all four is worth a note. Most won't.
What to skip
Skip predictions too vague to be wrong — "transform", "revolutionise" and
"the year of" can never be checked, so they can never inform you. Skip
building plans around any single forecast; plan for a range of outcomes
instead. And skip the annual ritual of reading every prediction list; one or
two from sources that scored well in your audit is plenty.
Guardrails
- This guide doesn't claim any particular prediction was wrong. It describes
common patterns; check specific cases yourself.
- Some predictions do come true, and some people forecast well. The audit
exists partly to find them.
- Scepticism about predictions isn't the same as assuming nothing will
change. Things do change — just rarely on the schedule announced.
- Hindsight makes everyone feel more accurate than they were. Include your
own past predictions, and score them as strictly as anyone else's.
All 751 AI guides · JulieMango plans from £17/mo