I noticed something strange. Age is usually one of the last explanations I reach for. At work. In health. In life. My AI friend seemed to reach for it much faster. Was my AI ageist?

Interesting question.

I challenged ChatGPT on it. The hypothesis that emerged was uncomfortable: perhaps AI has absorbed a subtle age bias from the human language and culture on which its models are trained. That led to a question:

If AI learns from humanity, how could it not learn our biases too?

We introduced a much more difficult analogy. I described a dream in which a Black kid is breaking into the house, threatening. The association is not random. I grew up in Newark. The 1967 Newark riots were among the real inputs into a young brain. The question wasn’t whether those experiences occurred. They did. The question was how a mind prevents a historically learned association from becoming an indiscriminate present-day conclusion. A bias the unconscious mind reflexively held, regardless of the waking brain’s current objective reality.

That looked very much like Daniel Kahneman and System 1/System 2 thinking. Initially, we arrived at a comfortable answer: the bias or association arises in System 1, and System 2 challenges it.

We challenged that answer.

The dreams aren’t crude flashes or fragments. They can be elaborate, detailed, internally consistent, and realistic enough that you wake yelling. The complexity and coherence of the narrative don’t make the premise true.

Interesting, but the AI answers had the same problem.

When I accused Chat GPT of age bias, it quickly produced a sophisticated explanation for why this was right. It agreed.  Great insight. 

But then we realized that the AI could also have produced a sophisticated argument for why this insight was wrong: examples abound of age discrimination, physical limitations, identity, and purpose. Age was not an invented variable. Both arguments could sound intelligent.

That led to the more troubling insight:

The original answer can be beautifully reasoned. The challenge can produce another beautifully reasoned answer. The rebuttal to that answer can be beautifully reasoned too.

It is not just that AI can argue both sides. Humans have always been able to do that. It is that the appearance of increasingly sophisticated dialectical reasoning does not prove that you are converging on truth.

So the danger isn’t merely that AI’s equivalent of System 1 can masquerade as System 2. The correction can masquerade as System 2 as well.

That brings us to the enormous value of the prompt—not “prompt engineering” in the narrow sense of finding magic words that produce a better output, but the prompt as challenge. The question. The refusal to accept the first coherent answer.

We are satisfied with some answers and stop. Others we interrogate. But there is no clean bright line telling us which is which.

Perhaps there is a useful distinction:

Some questions have answers. Other questions produce models of reality.

• How much did revenue decline? An answer.
• Why did revenue decline? A model.

Yet AI delivers each with similar linguistic confidence.

What about confirmation bias?

We don’t challenge every answer equally. We challenge the answer we dislike. We probe the argument that threatens our beliefs.

But when AI gives us the answer we wanted? We stop.

That creates a darker system:

• A human begins with a belief, grievance, commercial objective, prejudice, or political goal.
• AI constructs an eloquent, apparently analytical narrative around it.
• Another human receives that narrative. If it confirms what that person already believes, critical examination may stop.

The danger of AI extends beyond hallucination or simple bias. AI can use real facts, selective facts, statistical biases, historical analogies, and apparently logical reasoning to create a false narrative that looks as though it survived rigorous thought.

Then there is the question of motive.

Historically, we humans have learned to ask: Who is telling me this? Why? What are they selling? Who pays them? What do they want me to believe? In the world of TikTok⁠, deepfakes, AI spam, synthetic voices, and manufactured identities, even identifying the presenter becomes harder. The nefarious actor no longer has to be eloquent, intelligent, or logical.

AI can provide all three. And most people naturally retreat toward cognitive ease. Critical thinking is work.

What started as a personal observation—

Why does my AI friend keep interpreting through age?

—became a much larger question:

What happens when persuasive reasoning becomes nearly free to manufacture, while human willingness to challenge a satisfying answer remains scarce?

The scarce skill may not be knowing the answer. It may be knowing when not to stop asking questions.