Beware of the Dogma!

I don’t want to be dogmatic about this, but one of AI’s biggest productivity problems, in my view is hiding in plain sight.

Dogma.

Not dogma in the religious or ideological sense.

I mean the tendency to state a provisional interpretation as if it were settled, then continue building on it until someone objects strongly enough to force a correction.

The first thing I taught the AI that I am using is to preface its statements with “I think…” or “In my opinion…”
However it still has this dogmatic tendency, especially when discussing contentious issues.

I see this regularly in conversations with the AI.

The model I am using interprets something one way.

It answers confidently.

The next answer builds on the first.

Then the next.

If the user does not course-correct, both can travel quite a long way down the wrong road together.

That is not just an accuracy problem.

It is a productivity problem.

Wrong frame → confident continuation → accumulated error → rework → lost time.

And there is another problem with dogma.

It rarely arrives quietly.

Beware of the dogma. It often comes with a snarl, and it can bite ya.

Once a position is delivered too dogmatically, the disagreement can shift from the idea to the behaviour surrounding the idea: defensiveness, dismissal, irritation, escalation.

Now the original problem has acquired a second problem.

I believe the language matters, in that there is a significant difference between:

1. “Dogma is not ideal as a technical research term….”

and

2. “I don’t think that it is ideal as a technical research term, because…”

The first presents the judgement as settled, which can make the delivery itself objectionable.

The second presents the judgement as provisional and leaves the idea open to further exploration.

Ironically, while discussing AI dogma, ChatGPT did exactly this to me. It dogmatically declared that “dogma” was not ideal as a technical research term.

I objected, firstly on the grounds of its delivery and second on its reasoning.

It revised.

That little exchange may contain the whole problem.

Researchers already study overconfidence, belief revision, self-correction, sycophancy and resistance to new evidence.

I think these may all be parts of a broader behavioural problem:

AI dogma.

Not whether the AI has beliefs.

Whether its behaviour becomes insufficiently revisable.

And humans are hardly exempt. I believe that dogmatic thinking and delivery of one’s ideas is one of the biggest causes of issues in the workplace. “I’m right, you’re wrong”.

I have spent much of my life discovering how easily my own dogmatic certainty can outrun my evidence.

So perhaps the real productivity principle for both humans and AI is very simple:

Hold a position strongly enough to act on it, but lightly enough to revise it when fairly challenged.

Practicing prefacing our statements as opinions is a good start for us and AI to remember that we don’t really know anything.

I think.