Establishing your own AI guardrails

I’ve been surprised by conversations I’ve had with friends and associates who are using AI for multiple tasks in their companies. They’ve complained about sycophancy, errors, and AI making up or bringing in extraneous information. I had these same concerns earlier this year, so I began to use the “AI Truth Protocol”. I put it in the settings for all of the AI platforms that I use. You can also add it to any prompt that you’re using.

The difference was immediate. I began to see more thorough and careful analyses and sourcing of conclusions, with citations on the web of where the information could be found. However, as I’ve done more sophisticated analyses of data, I’ve wanted to upgrade analytic logic, ensure more source labeling, perspective tracking and create certainty boundaries. A little more research has produced five additional steps that I take when I want to ensure the quality and credibility of AI results.

1. Use an adversarial auditor

In my long-ago banking career, I learned that nobody likes it when auditors show up. They check your procedures, results, often find critical errors, and force improvement. This is no different. Here’s a prompt you can use:

2.  Avoid phrasing and educated guesses that can sound like facts. 

You’ve probably seen this before. AI can sound so sure of itself that every claim can sound like a fact. The key is to force the AI to tag its sources and certainty levels by cutting out fluff, guesswork or hiding behind fancy words. Try this prompt:

3. Separate the wheat from the chaff

Depending on how you phrase your prompt, AI can present opinions as facts. To avoid that you’ll want to separate viewpoints and messaging from facts. It’s not that the viewpoints aren’t important but you’ll want to know where each claim originates. Here’s a prompt:

4. Forcing an ironclad argument

Sometimes referred to as “Steelmanning”, it’s basically a way of looking at competing viewpoints in their strongest possible form. Think of it as two people presenting their case, each equally strong but enabling you to see gaps and test your hypothesis for going forward.  Here’s a prompt:

5. Don’t invite AI to Halloween

AI always wants to be helpful. Think of it as your own personal Natalie Harp, which means it will often make up facts, quotes or statistics rather than admit it doesn’t know something. But that’s a costume you won’t want so to get rid of it, you’ll want grounding and provenance. Here’s a prompt:

You won’t need all of these all the time, but they’re good extra steps when you want to ensure accuracy and objectivity. The AI Truth Protocol, by itself, will cover most needs. These are extra steps you can take for high-value analysis. Some of the prompts are also a little duplicative, but they each go to a different spot in any analysis. If you have some of your own that you like, I’d love to hear them in the comments.