You Cannot Hold AI Accountable—Why Knowing Its Weaknesses Improves the Quality of Your Decisions

Why is it that the more you consult with AI, the less capable you become of making decisions?
The structural reason for this is that AI is fundamentally poor at 'assuming responsibility' and 'maintaining long-term context'.
By accurately understanding these two limitations, you can set appropriate expectations for AI and regain your own decision-making power.

Why a map of strengths alone is insufficient

This week, we are covering the areas where AI excels.
Generation, exploration, summarization, and evaluation
If you use these four correctly, the quality and speed of your work will increase.

However, a map of strengths alone is not enough.

If you remain unaware of its weak areas, you will end up delegating tasks it is bad at along with the tasks it is good at, even if you are delegating the good tasks correctly.
If you continue to have expectations for areas where it is weak, not only will the quality of the output drop, but your own decision-making ability will be gradually eroded.

Today and tomorrow, we will cover the areas where AI is weak.
Today, it is 'assuming responsibility' and 'maintaining long-term context'.

AI's Weak Area #1: Assuming Responsibility

AI cannot take responsibility.
This is not a matter of capability, but a matter of structure.

If you ask, 'Should I change jobs?', AI will organize the options and criteria for you.
If you ask, 'Should I invest in this business?', it will list the risks and returns.
If you ask, 'What should I do about this relationship?', it will provide patterns for how to handle it.

You get an output.
You get a plausible-sounding answer.
But AI cannot take responsibility for that answer.

If you change jobs and fail, AI loses nothing.
If an investment goes wrong, AI suffers no loss.
If a relationship breaks down, AI feels no pain.

An 'answer' from an entity that bears no responsibility can serve as a reference, but it cannot be the basis for a decision.
If you confuse these, when you act based on AI output and fail, you will fall into the trap of thinking, 'It's because the AI said so.'
In that state, you cannot learn from failure.
You will repeat the same mistakes.

The assumption of responsibility always lies with the human side.
AI is merely a tool to assist in that judgment.
This recognition protects the quality of your decisions.

AI's Weak Area #2: Maintaining Long-Term Context

AI does not know your context.

Information shared during a conversation is only valid within that conversation.
If you start a different conversation the next day, the AI starts from zero again.
Your past experiences, the skills you have built up, your family situation, your vision for five years from now...
AI does not possess these.

What does this mean?
Consultation without context will only return answers without context.

To the question, 'I want to start a side job, what would be good?', AI will return a generic answer.
But that answer does not have the context that
you are in your 40s, work in sales, have two children, only have time on weekends, are good at writing, and want to become independent in three years.

If you adopt an answer without context as is, you will end up taking actions that do not fit your situation.
You take action, but get no results.
Because you get no results, you don't continue.
Because you don't continue, you give up.

It is not uncommon for the lack of long-term context to be at the root of this sequence.

How to use AI correctly once you know its weak areas

Once you understand that accountability and long-term context are AI's weak points, your way of using it will change.

Regarding taking responsibility
Use AI output as material for decision-making. You make the decision itself.
Maintain a structure where you say, "Based on the information organized by the AI, I have decided this," rather than "Because the AI said so."

Regarding maintaining long-term context
Supplement the context yourself every time.
Include the premise in every question to the AI, such as "I am in this situation, I have this goal, and I have these constraints."
Questions that include context elicit answers that fit that context.

It is the human's role to compensate for weak areas.
Leave the areas of strength to the AI. Humans compensate for the areas of weakness.
This division of roles is the essence of designing how to use AI correctly.

Today's 5-minute experiment

Recall one answer you received recently after consulting with an AI.
Then, try answering the following two questions.

Did that answer include your long-term context?
If you acted based on that answer, who would take responsibility?

If either of these is ambiguous, the result of that consultation may not be usable.
Try supplementing the context and re-posing the same question to the AI after shifting the responsibility back to yourself.
The output should change.

Tomorrow (3/13), I will write about the "area of value judgment where AI is weak."
What happens if you keep asking AI for value judgments?
I will break down that structure.

This article is included in the serial magazine "Thinking OS in the AI Era."