Today, I nearly sent a client an incorrect AI answer
On Tuesday afternoon, I was doing industry analysis for a client in Hangzhou. The model merged the timelines of two policies. The numbers looked smooth, and I almost pasted them straight into the proposal. When the client asked for sources, I discovered that one of the policies did not exist. I panicked for a moment: I had only just understood that AI’s biggest problem is not obvious nonsense—it is being wrong in a way that looks real. It is fine not to deal with this now, but if I use AI to deliver conclusions, this step is worth taking another ten minutes.
The research: don’t just hear the answer, check the reasoning
Anthropic recently introduced research into “provable reasoning.” The core idea is not to make models never answer incorrectly, but to give them a derivation process that can be checked. I understand the paper this way: whenever AI reaches a conclusion, it writes out the intermediate steps, evidence, and uncertainties, making it easier for a person to spot weak links.
Some people are already doing something similar. Siyuan Chen runs an independent consulting business in Shanghai. He has the model list the facts first, then attach evidence to each one; when information cannot be confirmed, he explicitly labels it “to be verified.” His method is not sophisticated, but it is steadier than asking for a polished conclusion directly. The same approach can work for our side projects, proposals, and personal brands.
Reproducing it today: almost no cost, but I had to change how I ask
Cost: CNY 0, using the AI chat tools I already have. Time: about 15 minutes. Technical barrier: no programming required; I only need to check the sources it cites. Operational barrier: no need to study the paper; I can start with the prompt below.
I can copy this: “Give the conclusion, but break the reasoning into three steps; list the evidence for each step; flag uncertain points separately, and tell me what still needs checking. Do not present guesses as facts.” Then I ask: “Which step is most likely wrong? What is the cheapest way to verify it?”
I have made this mistake too: at first, I asked the model to “make it sound professional,” and all it did was make the tone professional. I reduced rework only after I added “list evidence, flag uncertainties, and suggest a verification method.” One caveat: this approach cannot guarantee the answer is correct. It only helps me find risks earlier. For important data, I still need to return to official websites, original contracts, or confirmation from the client.
Choose based on your current workload; no need to do everything at once
If I am just starting: it is fine not to try this now. I can add “provide evidence” and “flag uncertainties” to my regular prompts, leave myself five extra minutes, and avoid building a complicated workflow.
If I already have one or two clients: for pricing, proposals, and factual judgments, I would ask the model to include sources and personally spot-check the most important one. Before delivery, I would also add: “Which parts here are only assumptions?” This can stop the client from discovering the problem for me.
If I am scaling up: I would create a checklist covering sources, dates, numbers, and the boundaries of any commitments, then have the team apply it. AI can generate the first draft, but the responsibility for verification must stay with a person. I would test it on one frequent task for two weeks before deciding whether to expand. This tool is not for everyone; what matters is whether it works with the professional judgment I already have.