When the Indian Ocean tsunami struck in 2004, some people saw the sea suddenly retreat and went down to look. Others ran.
On the Indonesian island of Simeulue, people remembered smong, an oral tradition handed down after a devastating tsunami in 1907. The story carried a simple instruction across generations: if the earth shakes and the sea withdraws, get to high ground. In 2004, that inherited knowledge saved lives.
It is a remarkable example of something human beings have been doing since before writing. We experience something, learn from it and try to prevent the lesson from dying with us.
But we are not particularly good at it.
Languages disappear. Oral traditions vanish. Libraries burn. Technologies are forgotten. Political ideas go out of fashion and are rediscovered generations later under different names. Sometimes knowledge survives physically while its meaning disappears. Archaeologists excavate buildings, pottery, tools, graves and rubbish and try to reconstruct not simply how people lived, but what they knew and thought.
We have always lived with an enormous problem of human memory.
Now AI may change that.
The obvious advantage is preservation. We can store quantities of information that no previous civilization could have contemplated and search across them almost instantaneously. But storage isn’t really the breakthrough. A warehouse stores things.
The more interesting possibility is reassembly.
Knowledge has become extraordinarily fragmented. A government department knows one thing. An Indigenous community remembers another. An academic discipline knows something else. A family has a story. A century-old newspaper contains an observation nobody has thought about for decades.
The pieces may all exist with some putting them together while others are abandoned.
We do this in public policy too. We separate food from housing, housing from income, income from disability, disability from health, and health from transportation. Each problem acquires its own experts, programs, statistics and vocabulary. Eventually we become very knowledgeable about the pieces while understanding less about the person experiencing the effects of being deprived of some or all of them.
Human knowledge can become similarly pixelated.
AI potentially works in the opposite direction. Faced with a new problem, it can search across boundaries humans have constructed and ask whether something learned somewhere else, at another time, might be crucial here.
That could be revolutionary. But it presents an equally large danger. AI learns from the record we leave it, and the record is not humanity.
Powerful societies wrote things down. Poor communities often did not. Governments created archives about themselves. Ordinary people left fewer records. Colonial powers sometimes dismissed local knowledge precisely because it wasn’t written in forms they recognized as knowledge.
Imagine an extraordinarily capable AI built in 1900 and instructed to preserve only humanity’s best-established knowledge. Would it have retained every Indigenous warning, folk observation and unfashionable scientific hypothesis? Or would it have confidently curated some of tomorrow’s truths out of existence?
That is why asking AI to distinguish the important from the unimportant is more dangerous than it first appears. Today’s trinket can become tomorrow’s Rosetta Stone.
Nor should we ask AI simply to preserve what is true and discard what is false. False ideas matter too. Humanity needs to remember the medical treatment that killed patients, the economic theory that failed, the prejudice once regarded as common sense and the government policy that produced the opposite of what its designers intended.
Otherwise, we don’t preserve learning. We preserve only the winners’ version of it.
An intelligent archive should be able to tell us: People believed this. Here is why. Here is the evidence they had. Here is what happened. Here is what later evidence showed. Here is what remains uncertain.
That is much harder than declaring something true or false. It is also much more valuable. And it suggests that our task is not simply to build larger artificial memories.
We need to preserve the argument around the knowledge.
We need minority reports, abandoned theories, oral histories and unsuccessful experiments. We need to distinguish evidence from assertion without deleting assertion. We need provenance: Where did this idea come from? Who believed it? What interests were involved? What evidence contradicted it? What happened when somebody acted upon it?
Most importantly, we cannot outsource remembering to AI.
Knowledge has to exist outside the machine as well. Languages must still be spoken. Children must still hear stories. Archives need protection. Institutions need institutional memory. Indigenous and local knowledge must be preserved by the communities to which it belongs, not merely harvested into somebody else’s database. People must still learn history rather than assuming a machine will retrieve it when required.
Because information available somewhere is not the same thing as knowledge available when it is important.
The people of Simeulue didn’t survive because somewhere there existed an archive describing the tsunami of 1907. They survived because the knowledge had travelled across a century and was present in people’s minds at the moment the sea disappeared.
That may be the standard we should apply to AI.
Not: How much can it remember?
But: Can it help us remember what matters when it’s crucial, without deciding for us what humanity is permitted to forget?
That requires work both inside and outside AI. Inside it, we need systems that preserve uncertainty, competing interpretations, provenance, forgotten knowledge and the history of failure—not machines that merely reproduce whatever today’s majority believes most confidently.
Outside it, we need human institutions capable of challenging the machine, adding what it cannot see, preserving knowledge it cannot capture and deciding what lessons deserve to be carried forward.
The ambition should not be to create a machine that remembers so that humanity no longer has to. It should be to create a civilization that forgets less, connects more and learns better.
For thousands of years, we have left fragments behind and hoped somebody in the future could make sense of them. We can now do something different.
We can preserve the fragments. We can preserve the arguments. We can preserve the mistakes. And we can preserve the lessons.
But ultimately, we still have to do what the people of Simeulue did.
We still have to know which way to run.