7 reasons why AI is not going to take your programming job

Predictions about AI replacing programmers are popular right now but I’m going to convince you that you don’t need to worry it.

It doesn’t take much critical thinking to see how wrong this idea is so let’s dive in.

1. Predictions about near-future AI capabilities have been wildly optimistic since the 1950s

People have been hyping AI since the term was invented. See here, here, and here for examples.

The most interesting thing you’ll find if you read those links is that really well regarded experts missed the mark not just by a year or two but by decades. In fact, many of the predictions made more than 50 years ago still haven’t come true. Let that sink in for a minute.

Predicting progress in AI is nearly impossible

One reason why AI predictions have been wildly optimistic is because it’s not really possible to predict how far you are away from your goal because getting that last few percent of accuracy (or whatever you care about) might take longer and cost more than the first 95% or 97% or 99%.

Let me give you an example. In the late 1990s I bought a copy of “Dragon NaturallySpeaking“, which is voice recognition software. I did all the training of the AI that it asked for (and more) but it just sucked no matter what I tried. A dictation system that types 70% of your words correctly with a 20 second lag after every few words is worthless. The claim at the time was that “we’re almost there. We just need a little faster computers and a little more time to work out the kinks”.

Well fast forward more than 25 years and I’d say we’re closer but still not quite “there” yet, despite all that time and orders of magnitude more compute available to us.

Don’t get me wrong, voice recognition tech has certainly improved a lot. But I still type emails and blog posts on a keyboard, not by dictation. Yes, I can dictate a text message or turn on a light with my voice but the original intent was to be able to dictate entire documents because it would be faster than typing and we’re still not there yet.

People almost always over-hype new technologies

Another problem with the current AI predictions is that people don’t understand the Gartner Hype Cycle.

Attribution: Jeremykemp

The Gartner Hype Cycle describes a phenomenon in which new technologies initially tend to get hyped far beyond their capabilities. This part of the curve is called the “peak of inflated expectations”, which is where we’re at with large language models and AI in general right now.

2. Many people hyping AI have a financial incentive to hype AI

How much can you really trust a person who stands to make billions of dollars by convincing you to use their crappy, buggy, insecure, copyright-infringing, power-hungry AI products so they can get investors to make them super rich?

Elon Musk has been promising full self driving Teslas are “just a year or so away” basically every year since 2016 and we still aren’t there. And many experts think he’s gone down a dead-end and will never get there with his current approach.

Sam Altman, the CEO of OpenAI, thinks that AI is going to replace 95% of the people who work in marketing.

Joe Russo, the director of a bunch of Avenger movies, thinks that AI will be able to “actually create” a movie “within two years”

There is zero reason to believe most of those claims will become reality in the next few years.

3. Writing software is much harder for AIs than writing blog posts, news articles, or even novels

Current AIs sucks at writing anything remotely complicated. CNET tried and failed spectacularly.

Writing software is much, much harder than writing a news story or a novel or a biography because we are writing for stupid, extremely literal computers, not humans who can eesilee understannd mannngleedd texxt, ambiguous phrasing, and outright errors or contractions in a document.

4. AI is only potentially helpful in a very small amount of the work that is done on a software project

Current AIs only potentially help with a bit of construction, PR creation, documentation, and code reviews. Some people think it’s good enough to help us find some bugs or generate ideas for test cases or help us with refactoring.

This is the key point that most non-programmers hyping AI for software development don’t understand: coding is only a very small part of what happens in a software project.

So even if AIs become quite proficient at coding that’s not what’s hard or time consuming about developing software.

The hard, time consuming parts of software development are:

• figuring out what your customers actually need
• turning those needs into a set of requirements that are internally consistent and complete
• creating an architecture and design that will make it easy to construct, test, and maintain the software
• project management/people management
• choosing the tech stack, dependencies, processes, and methods you will use on the project and keeping them up-to-date
• getting programmers to efficiently turn those requirements into working software without breaking some other part of the project or creating an unmaintainable mess
• modifying the requirements, architecture, design, code, and QA plan in the face of change requests
• delivering software that actually solves your customer’s problems on time and on budget
• confirming functional and non-functional requirements have been satisfied with various forms of testing and QA
• rework to fix errors or issues discovered during testing and QA
• maintaining and modifying the software over time while it is in production

So if you look at where all the person-hours in a software project are spent, how much of a speed-up can you really expect from AI in the next few years? 2%? I’d be absolutely shocked if it was 5% across our industry.

5. Even if AI exceeds all expectations for software development who do you think is going to be talking to the AIs?

Today’s software developers will be the people doing the prompting, reviewing, troubleshooting, testing, and monitoring the work the AIs are doing. We are—by far—the people best equipped to deal with these kinds of tools.

Remember, someone still has to select an appropriate AI to use, feed the AI the requirements, tell it how the software should behave in every circumstance, and confirm that the software behaves as expected. There is no realistic scenario where a non-technical person can just say “write software that will run this business” then come back in a week or a month and it will “just work”. Any significant system likely has thousands of explicit requirements. And every explicit requirement likely has some number of implicit requirements.

So even if a non-technical person somehow got a list of requirements together and prompted the AI with the system shall …, the system shall not …, the system shall …, the system shall not …,the system shall …, etc. for a month straight how do you imagine that software will turn out?

6. Reviewing code is harder than writing code and that’s a big problem for the current AI approaches to software development

The main use case people are excited about for programming is AI writing code from prompts. But the current AIs are wrong more than 50% of the time. And the truth is that it’s harder to review code than it is to write it. That’s why we learn code review skills after we learn programming skills.

So imagine you write a perfect specification for a module and feed it to a capable AI and it spits out 5,000 LOC that looks like it might work. What’s your next step? You need some kind of code review, right?

How are you going to review such a big change? How will you determine it:

• meets all requirements?
• doesn’t have hidden features?
• won’t break any existing functionality?
• has the non-functional properties you care about: security, maintainability, readability, correctness, robustness, testability, efficiency or whatever?

7. The more AI speeds up software development, the more work we’ll have

Even if you disagree everything you’ve read up to this point and you are more convinced than ever that AI is going to revolutionize programming, AI is still not going to take your job because a basic rule of economics is that as the price of a good decreases, the quantity of that good demanded by the market will increase.

Let’s work through the logic. If AI makes programmers twice as productive as they are today (which I highly doubt is achievable in the next few year) that means software development costs half as much, right?

If software development costs half as much, that makes lots of software projects that were not economically viable into economically viable projects. And since there is no limit to the amount of software that can be written, I don’t see any reason why AI would decrease the number of programmers working in our field.

In fact, the opposite is true. We’ll need programmers to integrate AI into existing applications and add it to new ones. Plus, we are going to need software developers to build and maintain all the AIs.

Still don’t believe me? Maybe the following argument will convince you.

Programmers have been absorbing every productivity advance we could get our hands on since the birth of our industry (i.e. high level languages, IDEs, version control systems, automated testing, static analysis, better methodologies, reuse, design patterns, frameworks, the web, the wide availability of tutorials, books, and training materials, automated code completion, syntax highlighting, SaaS, microservices, DevOps, faster hardware, etc.) and the demand for programmers has consistently exceeded the supply in part because falling costs create more work, not less.

The only way AI would decrease the total demand for programmers is if it could literally do almost everything a programmer can do, which would require AGI or super AGI. And nobody credible is predicting that for the foreseeable future.

Even the eternal tech optimist Bill Gates does NOT believe AI is going to reduce the need for software engineers.

Wrapping up

So what does the future hold for software developers?

The AI hype will continue for a while. Many programmers will jump on the bandwagon and write blog posts about how good it is and gush about it at conferences.

The marketing people will get us to put AI in tons of software and most of it will be crap and not profitable.

We’ll adopt AI tools in our software development workflows. Some will help us do our jobs better and some won’t but we’ll wish they did.

AI tools for software development will get better over time but their effect on developer productivity will likely take years to become noticeable. And regardless of how much AI improves developer productivity I’m confident programming jobs are not going away any time soon.

Have a comment, question or a story to share? Let me have it in the comments.