I’ve been spending a lot of time using generative AI lately. Some of this is in filmmaking-related tasks. But a lot of it is to develop software in my day job as a software engineer. I find it incredibly useful and frequently frustrating. But I’ve found that I can address those frustrations by following some simple rules. I’m sharing them here in hope that they can help others.
But note, I am not going to address the question of whether you should use generative AI. I know that people have strong opinions on that question. I have strong opinions myself. But I really don’t want to debate the issue here. If you want to discuss it, by all means reach out to me. I find this stuff fascinating, and I’m always willing to have a mutually-respectful discussion about it.
Note: these rules are based on the state of AI in August, 2026. This technology is changing rapidly. A year from now my rules might be entirely different. But this is what I do today.
1. Don’t trust AI! Always check its work!
We all know that AI sometimes lies. And it also makes mistakes. If you are creating something that you care about, always carefully review what the AI produces. Because it will screw up.
And whatever you do, don’t trust AI with your credit cards. And don’t let it loose on your computer. People have lost data and money by trusting AI too much. Don’t be one of those people.
2. Think of AI as a super-fast unusually knowledgeable intern.
You can get great work from an intern, especially one that works at the speed of AI. But you would have second thoughts before trusting your life or business in an intern’s hands without paying careful attention to what they’re doing. Which, of course, is why you should not trust AI (rule 1).
Dario Amodei, CEO of Anthropic, has said that when programming computers, AI is currently up to the level of a good computer science undergraduate. That seems about right to me. I don’t know about you, but I wouldn’t trust a computer science undergraduate to build a complex critical system without careful oversight. Treat AI the same.
3. AI fails like humans fail, not like computers fail.
I find this fascinating. The kind of ways I see AI failing remind me a lot more of human failures than computer failures.
Humans fail in unpredictable and inconsistent ways, often overlooking details or ignoring instructions. They will occasionally imagine things that aren’t there. And they often have a different understanding of the task than you do.
Computer failures tend to be predictable and reproducible. If you press a certain key sequence that causes a failure, and you later press that same key sequence, the system will usually fail in the same way. That’s not true of humans, and it’s not true of AI.
When using AI, be prepared to encounter very human types of failure.
4. Make an AI sandwich.
If I am going to use AI to create something, say, a document, I will follow what I call the AI sandwich approach. First, I write my own draft. Then I send that draft to the AI and get its feedback and ideas for enhancements. Sometimes I have the AI prepare its own draft. Finally, I take what the AI produced and edit it myself, checking any of the AI’s claims and cleaning up its prose.
I’ve heard this called the AI Sandwich – two slices of me with some AI in the middle. It works well.
5. Don’t pay for a special-purpose AI service until you’ve tried one of the standard ones.
Every day I see ads for new special-purpose AI services. Often these give feedback on screenplays.
Be careful of those services. Because often you can use Claude or ChatGPT or Gemini to give you coverage that is just as good and a lot cheaper
The other day I saw an ad for one of these services. AI will help you write a screenplay! And all for the low low price of $20 per session!
You know what else costs $20? A whole month of access to ChatGPT’s professional model. Where you can have as many sessions as you want .
Now I haven’t tried one of these special-purpose AI screenplay services. Maybe they really are much better than the standard models. Maybe they could give you more useful feedback than you’d get from one of the big boys.
But I have my doubts. I suspect (without having dived into how those services actually work) that what they do is take your screenplay, use a carefully crafted prompt (probably generated with a lot of AI help), and run it against one of the big models. Because those big models are designed to allow just that – the service pays OpenAI, marks up that price, often by a lot, and then return the output to you.
So if you try one of those services, compare the feedback you get from the feedback from ChatGPT or Claude or Gemini. See if it’s actually any better. See if you are getting your money’s worth. My bet is that you aren’t.
But if you find otherwise, let me know. I love to be proven wrong!
6. Worried about AI stealing your work? Follow these steps.
I know a lot of people are worried that if they upload a screenplay to an AI, the AI will use it as training data. But all the major AI’s have settings that let you disable this feature. Here’s some ways to disable training.
Note: these are available if you are a paid user of these services – they may not be available if you are using their free version.
ChatGPT:
1. With ChatGPT on your browser, click on your name in the bottom left-hand corner. A window pops up. Hit “Settings.”
2. In the Settings pop-up menu, choose “Data Controls.”
3. In the resulting menu, the setting you want is “Improve the model for everyone.” Turn it off to disable training on your data.
Claude.ai:
1. With Claude.ai on your browser, click on your name in the bottom left-hand corner. A window pops up. Hit “Settings.”
2. In the Settings pop-up, choose “Privacy.”
3. In the resulting menu, the setting you want is “Help improve Claude.” Turn it off.
Gemini (Google’s AI product):
Gemini’s terms of use explicitly state that “Your data isn’t used to train or improve Gemini or other generative AI models. Your data is as secure when using Gemini as it is using any Google core service like Gmail or Docs.” (https://myactivity.google.com/product/gemini?utm_source=gemini&pli=1)
However, somewhere in the Google privacy and terms section, they say: “We use the information we collect in existing services to help us develop new ones.” (https://policies.google.com/privacy?hl=en-US#whycollect) Whether or not this means they can use your data to train AI is a question for a lawyer, which I am not. But note that the rights that Google has to your data are no different when uploading it to Gemini than when using Gmail to send a script to someone.
Find these useful? Come back next time for 6 more!
Mr. Dzikiewicz can be reached at joe@crimesoftheart.com
