The ladder of AI inference

My former colleague Wayne Robinson's recent post about whose worldview trains the world's AI made me think, which is always a good thing. His account of how Elon Musk wants to "rewrite human knowledge" to remove "left-wing bias" from Grok helped me pin down a slow hunch: we are not just dealing with biased AI output, we are producing it, through biased input.

Wayne calls AI "a mirror that reflects power, not pluralism", and it struck me that every time we talk to a machine we are climbing what the organizational psychologist Chris Argyris called the ladder of inference, from observable data through our interpretations, assumptions and beliefs to conclusions and action. The problem is that we are lazy. Most of us race up that ladder when we prompt, skipping the lower rungs where bias creeps in, mainly because we are in a hurry. Each sloppy, context-free prompt wastes computation (an AI query uses something like ten times the energy of a Google search) and perpetuates the problem Wayne is describing. And before you say "stop using AI for search", that train has left the station; your basic Google search uses AI now. The answer is to use it better.

So why reinvent the wheel? Argyris gave us a framework for managing inference and bias decades ago. Combined with Philip Tetlock's work on superforecasting, it turns into a practical way of prompting that gets better answers and counters bias at the same time.

The ladder you climb every time you type.

Argyris's ladder has seven rungs, and you climb it whether you notice or not:

1. Observable data: the raw facts of your situation
2. Selected data: what you choose to include, and to leave out
3. Meanings: what you decide the situation means
4. Assumptions: about the context you think the AI already understands
5. Conclusions: about what kind of help you need
6. Beliefs: about what the AI should do
7. Actions: you type the prompt

Most people start at rung six or seven. They fire off a prompt based on their conclusions and beliefs without the data and reasoning that got them there, and the AI fills the gaps with its own trained assumptions. You do not just get a generic answer. You get the worldview Wayne is warning about, amplified.

Tetlock's superforecasters do the opposite. Question your assumptions constantly. Start from base rates rather than dramatic exceptions. Update when the evidence changes. Say what you do not know. Those are exactly the behaviors for climbing back down the ladder, and exactly what is missing from most prompts. Skip the lower rungs and you are asking the machine to predict on incomplete information, which sets both of you up to be wrong with confidence.

LADDER, for the prompts that matter.

From Argyris and Tetlock I put together a framework for complex interactions, the ones where bias and thin reasoning could really derail the result. I call it LADDER (cute, I know):

• Lead with your goal
• Anchor in observable data
• Detail your reasoning chain
• Declare your assumptions
• Explain what you want
• Request specific deliverables

This is as much about your own thinking as the machine's. Forcing yourself to articulate each rung catches your biases before they contaminate the answer. Take a scattered prompt: "How should my startup approach product development?" What is missing is everything: data about your market, why the current approach is not working, what you assume about customers, the constraints you are under. What you get back is generic startup advice reflecting Silicon Valley groupthink rather than your situation.

The same question through LADDER:

Lead: I am developing a product development strategy that accounts for high uncertainty in our emerging market.

Anchor: Most startups fail after funding. Airbnb pivoted from air mattresses to a platform and Twitter from a podcasting tool to microblogging; both adapted on market feedback. Theranos and Quibi stuck to their original assumptions despite mounting evidence against them.

Detail: My reasoning: companies that systematically update beliefs on evidence outperform those that cling to first assumptions, so cognitive flexibility matters more than predictive accuracy in uncertain markets.

Declare: Assumptions: our market is a complex adaptive system; small changes can have large effects; learning velocity beats planning accuracy.

Explain: I need a framework that favors rapid learning over detailed prediction while keeping strategic coherence.

Request: Methods to test core assumptions quickly; criteria for pivot-or-persevere decisions; team practices that encourage belief updating.

The answer you get is tailored to your situation and actively resists the common startup biases, because you told it what they were.

Not every prompt needs this. LADDER is for complex problems with many variables, situations where bias could do real damage, synthesis across domains, research and analysis. For basic tasks a simple frame is enough: role, minimal context, task, format, constraints. There is a tool at ladder.chock.llc that walks you through both.

The resource math that ought to matter.

Beyond better answers, structured prompting wastes less. A scattered approach typically takes four to six queries to reach something usable; a LADDER prompt takes one or two. Multiply by the energy per query and the difference is not trivial, and the better-specified prompt also provides a higher-quality signal to the systems learning from us. This is personal efficiency, but it is also being a better participant in the ecosystem.

This week.

Notice yourself climbing. Before the next complex interaction, ask what assumptions you are making and what data you are skipping. Use LADDER once and watch how articulating your reasoning changes your own thinking. When the answer comes back, ask what worldview might be embedded in it and what is missing. And if you need a follow-up, say what new information changed your mind rather than just asking for "adjustments".

Wayne's point stands: AI mirrors the worldviews of its training data and of its users. Interact carelessly, jumping to conclusions and skipping the evidence, and you get poor results and reinforce the biases you claim to want gone. Climb back down the ladder, borrow Tetlock's habits, make your reasoning visible, and you become a better partner to the machine, providing the kind of signal that could, over time, produce less biased answers for everyone. The goal is not perfect objectivity, which is impossible, nor paralyzing relativism, which is useless. It is epistemic humility: knowing what you do not know, showing your reasoning, staying open to evidence that says you are wrong. In a world where AI increasingly shapes how we think about hard problems, that saves you time, and saves the rest of us resources.