Smart AI, responsible AI

Skewed data, skewed AI

Earlier we said AI learns from data.
Show it many examples
and it decides by the patterns inside them.
But what if the examples shown
are leaning to one side?
The AI learns that lean too.
If it sees only one kind of example a lot,
it recognizes other kinds poorly.
Not human prejudice but the imbalance in data
flows straight into the AI's decisions.

Feed the examples to one side. When the data skews to one color, the AI's decision line tilts that way and keeps getting the smaller side wrong.

When we piled the data to one side,
the AI's judgment tilted along with it.
The side it saw a lot it gets right,
but the side it saw little it keeps missing.
It is not that the AI is bad.
Its learning material was skewed.
We call this bias.
So when building an AI,
people must check carefully
whether the data was gathered evenly.

It confidently makes things up

Earlier we learned an LLM picks the next word
by probability and strings it along.
Its goal is to connect words plausibly,
so it does not separately know what is true.
So even when asked what it does not know,
rather than saying it does not know
it sometimes spins out a plausible sentence.
The tone looks very confident
but the content is wrong.
We call this hallucination.

Tap a question the AI knows and one it does not. Even for what it does not know it does not stop, but confidently makes up a plausible answer, that is hallucination.

For a question it knew, the right answer came,
but for one it did not know it did not stop
and produced a plausible answer anyway.
On top of that, an LLM only knows material
up to when it was trained,
so it is shaky on the latest events after that.
So an AI's answer
must always be checked for truth.
Especially numbers, sources, and recent events
are safest when a person looks again.

Same technology, for good or for bad

AI can make pictures, mimic voices,
and write text without a hitch.
This ability is truly useful in many ways.
It can restore a lost old photo to be clear,
read text aloud for someone who cannot see,
and interpret a foreign language on the spot.
But with the very same ability
one could make fakes that look real
and use them to deceive others.
The tool is one, but the intent can be two.

Pick one technology and light its helpful use and its harmful use side by side. You can see that even the same tool, how it is used decides good or bad.

The same technology could help on one side
and harm on the other.
The technology itself has no good or bad.
What decides good or bad
is the choice of the person who uses it.
So the more powerful the tool,
the more we need agreements on how to use it.
If a fake is made, mark it as a fake,
and draw a line so it does no harm to others,
that kind of agreement.

Check by people, transparent, and fair

So how do we use AI responsibly?
It helps to remember just three things.
First, people check.
Do not believe the AI's answer as is;
for important decisions a person looks again.
Second, be transparent.
With what data and how it was built,
show it without hiding.
Third, watch for fairness.
Take care that no particular side loses out.

Tap the three checks in order to pass them. Only when human check, transparency, and fairness are all on does the AI become safe to rely on.

With all three turned on,
at last it became safe to use.
Looking back, the things we learned
gather into one here.
AI learns from data,
stacks neurons and layers to catch patterns,
and an LLM strings words by probability.
As powerful as it is, its limits are clear too.
So the last square is always the human's part.
Check, disclose, and be fair.

Let's wrap up

Gathered on one line, it is this.
AI is powerful but not perfect.
Skewed data skews decisions (bias),
it confidently makes up what it does not know (hallucination),
and the same technology can be used badly (misuse).
So the responsibility to use it with people checking,
transparently, and fairly must always go along.
The smarter the AI,
the more it matters to use that smartness responsibly.

Tap the AI journey in order to close it out. (learns from data -> patterns from neurons and layers -> LLM speaks by probability -> limits are bias, hallucination, misuse -> use it responsibly)

We have come truly far.
Starting from data,
to one neuron, a neural network stacked in layers,
an LLM that strings words by probability,
and on to AI that creates new things.
Now at the end we gained one more thing.
Not only how to use AI cleverly
but how to use it responsibly.
From the seat of a person holding a powerful tool,
let's seek better answers together with AI.