Imagine somebody tells you this:
A magnitude 7.1 earthquake will strike near Los Angeles next Tuesday at 14:17.
That sounds useful.
It also sounds like the kind of thing modern science should be able to do.
We can track hurricanes across oceans.
Forecast temperature days ahead.
Watch volcanoes deform.
Detect gravitational waves from collisions billions of light-years away.
So why can we not tell people exactly when the ground beneath a city is about to break?
The answer is not that seismologists know nothing.
It is almost the opposite.
They know enough to understand why precise earthquake prediction is an extraordinarily hard problem.
Prediction requires three things at once
The U.S. Geological Survey uses a strict definition.
A genuine earthquake prediction has to specify:
time
location
magnitude
Not:
“There could be a large earthquake on the West Coast soon.”
Not:
“Seismic activity is elevated.”
Not:
“A major earthquake will happen somewhere around the Pacific.”
Those statements may sound predictive.
They are too broad to fail cleanly.
And that matters.
A prediction becomes scientifically interesting only when reality has a real chance to prove it wrong.
Earthquakes begin where we cannot watch directly
Most damaging tectonic earthquakes begin kilometers beneath the surface.
Stress accumulates on faults.
Rock deforms.
Fault surfaces interact.
Fluids may change conditions locally.
Nearby earthquakes can alter stress.
But the final transition from a fault that is locked to a fault that suddenly ruptures happens inside a complex underground system we cannot instrument point by point.
This is not like watching a storm cloud on radar.
The decisive process is largely hidden inside rock.
Scientists infer what is happening from indirect measurements:
seismic waves
surface deformation
geology
GPS
strain
historical earthquake patterns.
Those tools are powerful.
They do not produce a countdown clock.
The precursor problem
For more than a century, researchers have looked for reliable earthquake precursors.
Small earthquakes.
Changes in groundwater.
Radon.
Electromagnetic anomalies.
Animal behavior.
Ground deformation.
Unusual seismic patterns.
The problem is not that nothing interesting ever happens before a large earthquake.
Sometimes it does.
The problem is specificity.
A signal is only useful for prediction if it happens reliably before large earthquakes and does not happen constantly when no large earthquake follows.
Many proposed precursors fail that test.
A swarm of small earthquakes may occur before a major event.
It may also end without one.
An anomaly that appears ten times and is followed by a large earthquake once is not a reliable alarm.
The foreshock paradox
Some major earthquakes have foreshocks.
That sounds promising.
Why not recognize the foreshock and issue a warning?
Because a foreshock is usually called a foreshock after the mainshock happens.
Before the larger event, it is simply an earthquake.
Most small earthquakes are not followed by a much larger one.
So the same event can have two identities:
Before the mainshock: ordinary earthquake.
After the mainshock: foreshock.
The label contains hindsight.
Forecasting is different from prediction
This distinction is crucial.
Scientists cannot currently say:
“A magnitude 7.1 will happen here at 14:17 next Tuesday.”
They can estimate probabilities.
A 2024 Reviews of Geophysics paper on operational earthquake forecasting explains that deterministic prediction of exact time and place remains impossible, while forecasting models can estimate the probability of earthquakes within a region and time window.
That is a completely different type of knowledge.
Prediction says:
This event will happen.
Forecasting says:
The probability has changed.
Weather science uses both probabilistic and deterministic tools too, but earthquake forecasting works with a system where the underlying event is far less directly observable.
Aftershocks are one place probability becomes useful
After a large earthquake, the chance of additional earthquakes nearby rises.
Scientists can use well-established statistical relationships to produce aftershock forecasts.
The USGS does this operationally.
But even here, the result is not:
“Another magnitude 6 will happen at 18:43.”
It is:
“There is an estimated probability of earthquakes above certain magnitudes over the next day, week or month.”
That may sound less dramatic.
It is far more honest.
Early warning is not prediction either
Earthquake early warning creates another source of confusion.
Systems such as ShakeAlert can detect an earthquake after rupture has already begun and send alerts before the strongest shaking reaches locations farther away.
That can create seconds of warning.
Sometimes more.
But the earthquake is already happening.
The system is racing seismic waves.
It is not predicting the event in advance.
Think of it as:
detection at machine speed
rather than
foreknowledge.
Why machine learning has not solved the problem
Artificial intelligence has revived public excitement around earthquake prediction.
Machine learning can find patterns in enormous datasets.
It can improve earthquake detection.
Classify seismic signals.
Estimate properties.
Assist forecasting research.
But pattern recognition does not automatically solve the physical predictability problem.
A model can find correlations that look impressive in historical data and still fail prospectively.
This is why modern forecasting research emphasizes:
benchmark comparison
prospective testing
reproducibility
transparency.
A prediction system has to work on earthquakes it has not already seen.
Otherwise hindsight can quietly leak into the model.
The enormous base-rate problem
Earthquakes happen constantly.
Most are small.
Large destructive earthquakes are rare relative to the number of possible places and times where they could occur.
That creates a statistical trap.
Imagine an alarm that frequently detects “earthquake-like conditions.”
If major earthquakes are rare, even a detector that seems impressive can produce huge numbers of false alarms.
A practical prediction system needs more than sensitivity.
It needs enough specificity that societies are not constantly evacuating for earthquakes that never arrive.
Prediction is not useful merely because it occasionally looks right.
It has to outperform chance and alternative models reliably.
Why vague earthquake predictions seem successful
A prediction says:
“A significant earthquake will strike the Pacific region soon.”
There are thousands of earthquakes around the Pacific.
Eventually something happens.
Then the prediction is reposted.
The misses disappear.
The hit survives.
This is classic selection bias.
The scientific fix is simple:
Write the prediction down beforehand.
Define the magnitude.
Define the region.
Define the time window.
Track every prediction.
Count every failure.
Now the magic often disappears.
Could precise prediction become possible someday?
Science should be careful with the word never.
Better sensors may reveal patterns we currently cannot measure.
New physics or better fault models may improve forecasting.
Machine learning may identify useful predictive information.
Dense sensor networks may change what is possible.
But current scientific evidence does not support precise deterministic prediction of major earthquakes.
The frontier is forecasting and rapid detection.
That may sound less exciting than a prophecy.
It is actually a more sophisticated achievement.
What science can do extremely well
Earthquake science can identify:
active faults
high-hazard zones
building vulnerabilities
probabilistic risk
ground-motion expectations
aftershock probabilities
and shaking once an earthquake has begun.
This changes the practical question.
If we cannot know the exact minute of the next earthquake, where should the effort go?
Toward making the earthquake less catastrophic when it arrives.
Engineering.
Codes.
Preparedness.
Monitoring.
Early warning.
Risk communication.
Prediction would be wonderful.
Resilience does not have to wait for it.
The DarkBrain conclusion
Earthquakes cannot currently be predicted precisely because the final transition to fault rupture occurs inside a complex, partially hidden physical system and no precursor has proved reliable enough to identify exact time, place and magnitude.
But “cannot predict” does not mean “cannot know anything.”
Science can forecast probabilities.
Map hazard.
Detect rupture rapidly.
Estimate aftershock risk.
And design structures around expected shaking.
That distinction matters because certainty is seductive.
A confident prediction feels more powerful than a probability.
But one of science's greatest strengths is knowing when the evidence supports a number instead of a prophecy.