AI as a Sparring Partner & Critic

Finding people who are both willing and capable of exposing flaws in our reasoning used to be a challenge, which is why AI now offers such a convenient sparring partner. At a minimum, it provides a starting point or a new angle to consider; at its best, it supplies the pushback needed to stress-test our conclusions before we act on them in the real world.

Still, treating an AI as an infallible authority, or dismissing it entirely because a single quick query yielded a shallow answer, misses how these systems operate. To extract reliable criticism without letting our own critical thinking atrophy, we ought to pay close attention to which models we select, what unspoken premises they rely upon, and how we push back across multiple turns.

1. Think Through the Problem Yourself First

Before asking an AI to critique an idea, generate counterarguments, or help troubleshoot an issue, you ought to map out your own reasoning and premises first. The goal of using AI here is to supplement and sharpen your thinking rather than replace it.

If you use AI to bypass the initial friction of forming your own arguments, you risk degrading the very capability you are trying to build simply by exercising it less than you did before. (Detailed explanation: Neural pathways tend to strengthen when they fire together repeatedly [Hebbian plasticity] (Hebb, 1949), and we generally remember information we generate ourselves better than information we simply read [generation effect] (Bertsch et al., 2007). Handing that initial thinking over to a tool [cognitive offloading] (Risko & Gilbert, 2016) means practicing it less, and skills that go unpracticed tend to fade [skill decay] (Arthur et al., 1998): in a field experiment, students who practiced math with unrestricted chatbot help did better during practice, but once that help was gone they scored about 17% lower on the exam than students who never had it (Bastani et al., 2025).) Think of AI like a sparring partner in martial arts: its role is to test the techniques and strikes you have already practiced throwing, instead of doing the training in your place.

2. Why the Model You Pick Matters

A common mistake is treating "AI" as if every chat interface runs on the same underlying capability. In practice, the reliability of the feedback depends heavily on which model you use (+ optional parameter) and how you manage the session:

• Use Reasoning ("Thinking") Models: Standard, lightweight "mini" or "flash" models commonly generate words almost instantaneously in a single pass, without pausing to verify their steps. Therefore, they frequently react to your tone or surface keywords rather than auditing your logic. A dedicated reasoning model takes the time to break down premises, search for edge cases, and check for internal contradictions before writing its response.
• Start a Fresh Chat When Needed: Even the most capable models begin to lose the thread when a single session drags across dozens of lengthy exchanges, tending to fixate on the most recent messages while forgetting the nuances established early on. Once a sparring session becomes bloated, carry your refined conclusions over into a clean chat window.
• (Optional) Switch Environments Across Providers: Every company tunes its models with slightly different priorities and guardrails. One platform might be heavily conditioned to remain polite and defend conventional views, while another is far more blunt. Running your reasoning through multiple models may quickly reveal whether an objection exposes a flaw in your logic or merely reflects one company's training bias.

3. Where AI Helps and Where It Falls Short as a Critic

4. The "Be Critical" Trap & When Leaving Things Vague Can Help

Once people notice that AI tends to validate whatever they say, a common overcorrection is to prompt it with commands like, "Be brutally honest and critical of my argument," or "Find all the flaws in this idea."

The issue is that this establishes a forced premise. Because language models are built to comply with instructions, commanding the system to "be critical" signals that success means attacking your position regardless of its merit. Even if your argument is bulletproof, the model will push itself to invent objections, resorting to strawmen, misreading your intent, or nitpicking a single word just to supply the criticism you demanded.

A similar dynamic occurs if you hand the AI an unpolished thought without clarifying your definitions or premises: it will fill the gaps with its own assumptions, and possibly critique a distorted version of what you meant. Still, withholding strict precision, or allowing the AI to throw forced, unfair criticisms at you, is not necessarily a bad thing when your goal is practice.

After all, in everyday conversations, people rarely listen with charitable precision or wait for you to lay out every premise cleanly. They often jump to conclusions, project past associations, attack strawmen, and rely on defensive logic to protect their worldview. Consequently, testing a rougher draft against an AI and navigating its flawed objections serves two distinct purposes:

First, observing where the AI misinterprets a vague sentence may highlight where a human listener is likely to misunderstand you as well, making it obvious which premises need to be clarified upfront. Second, pushing back against forced or inaccurate critiques serves as practical training for spotting subtle flaws in an opponent's reasoning and dismantling bad-faith or misaligned arguments under pressure.

5. Opening the Field with Pros, Cons, and Counter-Arguments

When you are troubleshooting a problem or refining an approach, just asking an AI for the solution right away tends to narrow your perspective prematurely, as this normally prompts it to give you only the single option it deems "best." Nearly every real-world decision involves trade-offs that depend on your specific context and priorities, and a single recommendation obscures what other options exist.

It can be far more useful to ask for multiple (distinct) approaches or strategies, alongside the pros and cons of each. Evaluating several paths side-by-side prevents you from latching onto the first passable idea and makes it easier to combine the strongest elements of different methods into something that fits your circumstances.

Even then, do not stop at the initial list of pros and cons the model generates; feel free to ask for more. To dig deeper, instruct the AI to argue against the very pros and cons it just listed. Ask it why a supposed advantage might be overrated or backfire, and under what conditions that would happen. Forcing the system to challenge both sides of its own breakdown strips away generic advice and clarifies which trade-offs really matter.

6. Checking the AI's Own Unspoken Premises

Much like people, AI models never critique from a neutral vacuum. Nearly every counterargument an AI generates rests upon unspoken premises inherited from its training data and corporate guardrails:

• Measuring Against the Status Quo: Models routinely evaluate new ideas against existing systems, labeling a realistically ideal overhaul "unrealistic" simply because it departs from legacy bureaucracy or requires uncomfortable transitional friction.
• Policing Tone Over Substance: Because these systems are heavily trained to avoid abrasive language, they frequently critique a direct, honest argument as "too harsh" or "alienating" rather than evaluating whether the point itself is accurate and necessary.
• Forcing a Safe Middle Ground: On polarized topics, models may be conditioned to play it safe, concluding that "both sides have valid points" even when one position rests on (partly) contradictory or disproven premises.

Whenever an AI pushes back on your reasoning, you could ask it directly: "What premise or value priority is your criticism assuming here? If my premise is X, does your objection still hold, and if so, what changes and why?"

7. Sparring Back and Forth Across Multiple Turns

A single query and response is barely the beginning of an inquiry; reliability comes from iterating across multiple turns.

Much like a person answering off the cuff, an AI's first answer tends to skim the surface, offering the most common angle in the safest phrasing. That rarely means there is nothing more to find. Follow-up prompts such as "What did you leave out?", "Are you sure you didn't miss anything, or could it be otherwise?", "Which part of this is weakest?", or "Go one level deeper on the second point" usually pull out the substance a single query never reaches.

• When the AI raises an objection, reply with your counterargument to see how its reasoning adapts, but watch out for premature surrenders. Because models are conditioned to be agreeable, they will often fold and validate your pushback the moment you challenge them, even if their original objection was sound.
• If you and the model reach an impasse on a specific point, run a test in a fresh chat (blind judge). Open a new session with a reasoning model and present both positions without revealing which one is yours. Stripping away which side you favor should neutralize the model's tendency to tell you what you want to hear.

8. Explain Your Goal in Plain Words and Let AI Build the Prompt

People sometimes treat prompting like a rigid coding language, assuming they need to memorize special phrasing to get a reliable result. In practice, trying to write a complex prompt from scratch on your first try is an unnecessary hurdle. You usually already know what you want to achieve and which AI habits annoy you, even if you have not figured out the best way to structure those rules yet.

A simpler approach is to describe what you want in plain, everyday words and ask the AI to build the prompt for you. Talk through your goal casually, mention the traps or shortcuts you want the system to avoid, and refine those boundaries over a few messages. Once the model grasps your intent, ask it to package those rules into a standalone prompt you can copy into a fresh chat, and optionally have it explain why it set the instructions up that way.