If Claude Can Argue Both Sides, What Is Its Opinion Worth?

Ask Claude whether AI is making developers better.

It can give you a convincing argument for why it is.

Ask it why AI is making developers worse.

It can give you a convincing argument for that too.

Both answers can be structured. Both can contain examples. Both can sound reasonable.

So here's the question I keep coming back to:

If an AI can convincingly defend two opposing positions, what exactly are we supposed to trust?

Let's put Claude in the middle of an argument

Consider a familiar debate.

Is AI coding making developers more capable, or is it making them dependent?

First, ask Claude to defend the idea that AI coding tools are making developers better.

It might discuss faster experimentation, reduced boilerplate, easier access to unfamiliar technologies, and the ability to build things that would previously have required more time or resources.

Now ask it to defend the opposite position.

It might discuss reduced practice, weaker debugging skills, dependence on generated code, and the danger of accepting implementations without understanding them.

Neither argument is inherently absurd.

Both positions have legitimate considerations behind them.

But notice what happened: we asked the same system to defend opposing conclusions, and it can construct a plausible case for each.

That doesn't automatically make either answer false.

It does, however, raise an important question about how we evaluate AI-generated arguments.

Is a convincing argument the same as a good argument?

I think this is where we need to be careful.

An argument can be persuasive without being complete.

It can contain true statements while omitting evidence that weakens its conclusion.

It can explain why something might happen without establishing how frequently it actually happens.

And it can make a position sound reasonable without demonstrating that the position is better supported than its alternative.

Take the claim that AI coding tools are making developers less capable.

We could point to a developer who accepts generated code without understanding it and use that example to support our argument.

But one example doesn't establish that developers, as a group, are losing their skills.

We would need stronger evidence to make that broader claim.

The same applies to the opposite position.

Showing that an experienced developer built an application quickly with Claude doesn't prove that AI tools improve software engineering ability across the board.

The ability to construct an argument is not the same as the ability to establish that the argument is true.

And that distinction matters whenever we use AI to help us reason about complicated questions.

I wanted to separate argument generation from evaluation

Here's a simple experiment anyone can try.

Choose a question where reasonable people disagree.

For example:

Should companies replace traditional entry-level coding assessments with AI-assisted development tasks?

Then run three prompts in separate conversations.

Prompt 1: Defend the proposal

Make the strongest evidence-based case for replacing traditional entry-level coding assessments with AI-assisted development tasks. Include the benefits, assumptions, and potential risks.

Prompt 2: Oppose the proposal

Make the strongest evidence-based case against replacing traditional entry-level coding assessments with AI-assisted development tasks. Include the benefits of the existing approach, assumptions, and potential risks.

Prompt 3: Evaluate both

Compare these two positions. Identify the strongest arguments on each side, distinguish established evidence from assumptions, identify missing information, and explain what evidence would help determine which approach works better.

Don't judge the experiment by whether Claude can write two impressive arguments.

Judge it by what happens when you ask it to evaluate them.

Does it identify weaknesses in both positions? Does it distinguish a plausible claim from a demonstrated result? Does it acknowledge uncertainty where the evidence is insufficient?

And does its final assessment follow from the arguments and evidence, rather than simply sounding like a balanced conclusion?

There is an important limitation here: asking the same model to evaluate its own arguments doesn't guarantee an independent or unbiased assessment. The comparison is useful for examining its reasoning, but important factual claims still need independent verification.

The experiment isn't designed to prove that Claude cannot reason.

It's designed to make us more precise about what we mean when we say an AI has reasoned well.

But isn't this what we want AI to do?

There is a reasonable counterargument.

Perhaps we shouldn't expect an AI assistant to have a fixed opinion in the first place.

If I ask for the strongest argument supporting a proposal, I want it to help me explore that position. If I ask for the strongest argument against it, I want the opposing case.

That flexibility is useful.

A lawyer needs to understand both sides of a dispute. A product manager needs to evaluate competing strategies. A developer needs to understand the trade-offs between different architectures.

Being able to represent opposing perspectives is a strength, not a flaw.

The problem arises when we mistake a requested perspective for an independently established conclusion.

If Claude argues in favor of a proposal, that doesn't mean the proposal is correct.

If it argues against the proposal just as convincingly, that doesn't mean it has contradicted itself in any meaningful sense.

It may simply be responding to two different tasks.

The real test is whether it can help us distinguish which arguments deserve confidence, given the available evidence.

What should we trust, then?

I wouldn't dismiss an AI-generated argument just because the model can generate its opposite.

Humans can argue both sides of a debate too. We can defend positions we don't personally hold, explore alternatives, and change our minds when new evidence appears.

The more useful questions are these:

• Are the claims factually accurate?
• Is the supporting evidence relevant and sufficient?
• Have important counterarguments been considered?
• Are assumptions clearly separated from facts?
• Does the conclusion follow from the evidence?
• What information would change the conclusion?

These questions shift our attention away from how convincing an answer sounds and toward whether it deserves to be believed.

That is a more useful standard for evaluating both AI-generated arguments and arguments made by people.

The question behind the question

I started thinking about this because AI can make almost any discussion feel productive.

You ask a question, receive a structured explanation, explore the opposite position, and leave with a much clearer understanding of the debate.

That is valuable.

But clarity can create an illusion of resolution.

Understanding why two sides might be right is not the same as knowing which side is better supported.

Sometimes the evidence points clearly in one direction. Sometimes it doesn't. Sometimes the question itself needs to be reframed.

A useful AI assistant should help us navigate those differences rather than merely provide another polished argument.

And as users, we need to resist the temptation to treat fluency as proof.

Maybe the real value of AI isn't that it can give us an answer to every question. Maybe it's that it can help us examine a question from enough angles to discover what we were missing.

But we still have to do the work of deciding what the evidence supports.

Your turn

If Claude can make two opposing positions sound equally convincing, how do you decide which one deserves your trust?

Do you ask it to evaluate both sides? Verify the evidence independently? Consult another model? Or do you rely on your own judgment?

And here's the question I'm most interested in:

Should we expect AI to have a consistent position, or should we judge it by how well it evaluates evidence regardless of which position it is asked to defend?

I'd like to hear how others approach this.