How to Get Reliable Output from an AI Smarter Than You

I keep telling everyone to do more with AI. And a typical (and important) question I get is: how do I know that the AI didn’t hallucinate (i.e., just make up answers) or if it is did something wrong because it misunderstood the situation? Especially if I’m making it do work that I am not an expert in, like programming (for non-programmers) or tax preparation or researching a new topic.

For that, I’m first going to take you on a detour

Humans have a lot of experience with the situation where someone who does not have expertise in area X gets someone else to do the work, and then they have to decide whether that person did it right. Managers are routinely required to supervise employees with skills that the manager does not have. Companies routinely outsource work to vendors who can do things that the company cannot. Even as an individual, you have lots of situations where you get someone else to do work for you in an area where you don’t have expertise.

In each of those cases, there is a possibility that the person/company doing the job might mess up because:

• they lie about their expertise

• they get lazy and do a half-assed job, and don’t check everything carefully

• they genuinely make mistakes because they did not have the full context of what you want and misunderstood something

• and many, many other reasons… if you spend a few minutes thinking about this, you’ll find other reasons why an expert made a mistake when doing something for you

I’m teaching an undergraduate class involving AI, and I asked my students to come up with techniques traditionally used by non-tech managers so that they can have some confidence in the programs written by the programmers reporting to them. This is what my students came up with:

• Look at the programmer’s past performance and track record.

  • Check relevant qualifications and credentials.

  • Use the reputation of the programmer, company, or institution as one signal of reliability.

• Ask the programmer to explain how they approached the problem — including approaches that failed or led to dead ends.

• Ask explicitly where the solution could fail, and what its known weaknesses or limitations are.

• Have two or more independent people assess the problem or the solution, and compare what they say.

• Get another experienced programmer to review the work.

• Learn enough of the technical basics to ask sensible questions and recognize obvious problems.

• Do spot checks on selected parts of the work rather than trying to verify everything.

This list was created by undergraduates, but it is a pretty good list. You might probably come up with a few more ideas.

The good news is that the same ideas apply to using AI when you are not an expert in that area. Here, I’ve reproduced the same list but changed the words so that they map to the situation of you using an AI:

• Check the past record of the AI. Ethan Mollick suggests that after you’ve spent 10 hours with an AI model, you begin to develop an instinct for what that particular model does well and what it messes up. In fact, he even suggests that every time you get access to a new model, you should interview it, like a job candidate.

  • Check the “qualifications” of this AI. There are a lot of benchmarks that test the skills of AIs at various different tasks. Check how well your AI does at benchmarks that are relevant to your tasks. (A good score on a benchmark should be taken with a pinch of salt, because companies game benchmarks, but a bad score is definitely a sign that you can’t trust the AI in that domain.)

  • Use the reputation reliability of the company providing that AI: we know that Anthropic’s models are good at programming but bad at images, OpenAI’s models are great general-purpose models, Google’s models are great for images, videos, and other multimedia but a little behind in text, and Grok can be unfiltered and unhinged, etc.

• Ask the AI for an explanation of how it arrived at the answer (or how it built the software): its assumptions, method, intermediate calculations or evidence, alternatives considered, and what would change its conclusion. Often, as someone with some experience of the world and common sense, you can spot some misjudgements or mistakes in the intermediate steps. Sometimes, you don’t have to spot them: the AI itself realizes that it did something questionable and then corrects itself.

• Ask the AI to identify weaknesses in its own answer: where it is uncertain, what edge cases it may have missed, what assumptions could be wrong, and under what circumstances its answer would fail. These will be high-level enough that you can decide whether any of these situations are important (in which case, you ask the AI to work on improving those areas), or that’s something you can live with.

• Ask different AIs to solve the same problem, and compare the results.

• If you got a cheaper/faster model to do the work (e.g., you used Sonnet to write the program), get an expensive/slow model to check/review the work (e.g., Fable 5). This can be cheaper than getting the expensive model to do all the work, but still gives you most of the benefits. (This is just like getting a junior software engineer to write a program and getting a senior to review the code.)

• Learn enough of the subject yourself to interrogate the AI intelligently and catch obvious errors.

• Spot-check some claims, calculations, citations, or examples independently. Concentrate especially on the parts that matter most to the conclusion.

At different times in the last year or so, I have used each of the above techniques to check the work that AI did for me.

And if you came up with a few more techniques in the previous section, you can map those yourself now.

Last year, I said that You Are Now the Manager of a Team of LLMs. That is more true than ever. This article gives some details of the specific skills you have to acquire.

Congratulations on your promotion, and now get to work on growing into the new role.