You know that kid in math class who writes "42!" in huge numbers at the bottom of the page, circles it twice, and hands in the test hoping sheer confidence will compensate for the complete absence of any actual work? That's AI answering most of your questions. It generates text one word at a time, trained to sound authoritative whether it's calculating correctly or just stringing together plausible-sounding numbers. Without seeing the steps in between, you have no way to catch where it went from actual math to creative fiction.
This matters because you're probably using AI for things that have real consequences. Ask ChatGPT to calculate ROI on a marketing campaign and it spits out "127%" with zero explanation of what numbers it used or how it got there. You're left either trusting it blindly or doing all the math yourself anyway, which defeats the point.
The solution is forcing AI to show its work. In technical terms, this is called chain of thought reasoning (generating intermediate steps instead of jumping straight to a final answer). Here's how to do it.
Trick #1: The "Think Step by Step" Demand
The simplest trick is almost embarrassingly effective: literally add "Let's think step by step" or "Show your work" to your prompt. That's it. Those five words force the model to generate intermediate reasoning before spitting out a conclusion, and it actually improves the quality of the final answer because the model has to commit to a logical path instead of just reaching for whatever sounds good.
When asking AI to compare pricing plans, without that phrase you get "Plan B is cheaper." With it, you see the actual per-unit breakdown that reveals Plan B is only cheaper if you're buying more than 500 units. That's the difference between a useful answer and an expensive mistake.
Trick #2: Make It Number Every Single Step
Ask for "numbered steps" or "show your reasoning in a numbered list." This creates an accountability structure. It's much harder to skip a logical leap when every step gets its own label. You can spot exactly where a number materializes out of thin air or where the logic does a little hop over the hard part.
Say you're calculating whether you can afford a house. Numbered steps reveal the AI assumed a 20% down payment you never mentioned, used a 30-year mortgage when you wanted 15, and estimated property taxes based on a completely different state. Each of those assumptions changes the answer dramatically, but they were invisible until you made every step explicit.
Trick #3: Ask It to Cite Specific Sources or Data Points
Add "cite the specific numbers or sources you used" to your prompt. This forces the model to ground its claims in retrievable information rather than plausible-sounding inventions. When it can't cite something, that's your red flag. You've just caught it guessing.
Researching competitor pricing? Asking for citations makes the AI admit "I don't have access to current pricing data for these companies" instead of confidently inventing numbers that sound about right. Now you know you need to go look those up yourself instead of putting made-up figures in your competitive analysis.
Trick #4: Request a "Confidence Check" on Each Step
Ask the model to rate its confidence or flag its assumptions. Try prompting with "Mark any step where you're uncertain or making an assumption." This surfaces the guesswork instead of burying it under authoritative-sounding prose. You get to see which parts are solid and which parts are essentially educated guesses that need verification.
Estimating a project timeline? The AI flags "assuming standard 40-hour weeks" and "estimated based on similar projects, but I don't have data on your specific team's velocity." Now you know exactly what to double-check with your actual team instead of discovering three weeks in that the timeline was built on fantasy.
Trick #5: Make It Solve the Problem Two Different Ways
Prompt with "Solve this using two different methods and compare the results." If the answers don't match, you know something went wrong. If they do match, your confidence level just shot up considerably. You're making it check itself without even announcing that's what you're doing.
Calculating break-even point? Ask for both the equation method and a month-by-month projection. When they don't align, you catch that it forgot to include your fixed costs in one approach but remembered them in the other. You just saved yourself from planning around a break-even point that was off by six months.
Trick #6: Ask for the Answer, Then Ask It to Verify Itself
Use a two-part prompt: get the answer first, then follow up with "Now check your work and identify any errors." Models turn out to be surprisingly good at spotting their own errors when explicitly told to look for them. (It's like when a teacher hands back your test and says "are you sure about number seven?")
You're building a budget forecast with multiple line items. The initial answer adds wrong (classic), but when you add "verify your arithmetic," it catches and corrects the sum. This works because generating the verification steps activates different patterns in the model than generating the initial answer did.
So what can YOU do with this?
For financial calculations you'll base real decisions on, always demand numbered steps and two different methods. The stakes are too high to trust a single confident-sounding answer.
Research summaries need specific citations so you can verify claims before repeating them. "Studies show" means nothing if you can't check which studies.
Any data analysis should come with confidence levels on statistics or trends. You need to know which numbers are solid and which are interpolated guesses.
Complex planning benefits from step-by-step reasoning followed by self-verification. Catch the errors before they cascade.
For high-stakes questions, stack multiple tricks together. Ask for step-by-step reasoning AND specific citations AND verification. Yes, the response gets longer, but you're actually able to trust it.
Real workflow example: you're asking AI to analyze survey data. Use "show numbered steps," "cite specific response counts," and "verify your percentages add to 100" to catch errors before you put those charts in a presentation to the executive team. The five extra seconds typing those requirements just saved you from a deeply embarrassing meeting.
TL;DR
- Add "think step by step" or "show your work" to prompts to force visible reasoning instead of confident-sounding final answers that might be completely wrong.
- Request numbered steps, citations, and confidence flags to spot exactly where AI is calculating versus guessing versus inventing.
- Make it solve problems two different ways or verify its own work to catch errors before you rely on the output.
- Stack multiple techniques for anything with real consequences like financial calculations, research, or data analysis.
Just remember: the bigger and more confident the circled answer, the more suspicious you should be about what's happening in the margins.
