9 Programming Tasks You Shouldn't Hand Off To Ai - And Why

Mininyx Doodle / Getty Images

It's over. Programming arsenic a occupation is done. Just motion up for a $20-per-month AI vibe coding service and fto nan AI do each nan work. Right?

Also: What is AI vibe coding? It's each nan rage but it's not for everyone -- here's why

Despite nan truth that tech companies for illustration Microsoft are showing coders nan door by nan thousands, AI cannot and will not beryllium nan sole shaper of code. In fact, location are galore programming tasks for which an AI is not suited.

In this article, I'm spotlighting 9 programming tasks wherever you shouldn't usage an AI. Stay tuned to nan end, because I showcase a 10th prize logic why you shouldn't ever usage an AI for programming. Not to mention that this could happen.

1. Complex systems and high-level creation

Here's nan thing. Generative AI systems are fundamentally super-smart auto-complete. They tin propose syntax, they tin code, and they tin enactment arsenic if they understand concepts. But each of that is based connected probabilistic algorithms and a ton of accusation scraped from nan web. Contextual intelligence is not a strength. Just effort talking to an AI for a while, and you'll spot them lose nan thread.

Also: 10 master developers connected vibe coding's existent committedness and peril

If you request to nutrient thing that requires important knowing of really systems interact, acquisition to make judgement calls astir trade-offs, knowing of what useful for your unsocial needs, and information of really everything fits pinch your goals and constraints, don't prosecute an AI.

2. Proprietary codebases and migrations

Large connection models are trained connected nationalist repositories and (shudder) Stack Overflow. Yeah, immoderate of nan astir astonishing codebases are successful nationalist repositories, but they're not your code. You and your squad cognize your code. All nan AI tin do is infer things astir your codification based connected what it knows astir everyone else's.

Also: A vibe coding scary story: What started arsenic 'a axenic dopamine hit' ended successful a nightmare

More than likely, if you springiness an AI your proprietary codification and inquire it to do large things, you'll embed galore lines of plausible-looking codification that conscionable won't work. I find that utilizing nan AI to constitute smaller snippets of codification that I different would person to look up from nationalist sources tin prevention a immense magnitude of time. But don't delegate your unsocial worth adhd to a brainy mimeograph machine.

3. Innovative caller worldly

If you want to create an algorithm that hasn't been done earlier -- possibly to springiness your statement a immense competitory advantage -- prosecute a machine scientist. Don't effort to get an AI to beryllium an innovator. AIs tin do wonders pinch making boilerplate look innovative, but if you request existent out‑of‑the‑box thinking, don't usage a glorified container pinch brains.

Also: Google's Jules AI coding supplier built a caller characteristic I could really vessel - while I made coffee

This applies not only to functional coding, but to creation arsenic well. To beryllium fair, AIs tin do immoderate awesome design. But if you're building a caller game, you whitethorn want to do astir of nan imaginative creation yourself and past usage nan AI to augment nan engaged work.

Sure, galore of america spell done life parroting things we heard from different folks aliases from immoderate wacky podcaster. But location are existent humans who are genuinely creative. That productivity tin beryllium a strategical advantage. While nan AI tin do volume, it really can't make intelligence leaps crossed uncharted paths.

4. Critical information programming and auditing

Do not fto the fox defender nan hen house. Fundamentally, we really don't cognize what AIs will do aliases erstwhile they'll spell rogue. While it makes consciousness to usage AI to scan for malicious activity, nan codification generated by AIs is still beautiful unreliable.

CSET (the Center for Security and Emerging Technology) astatine Georgetown University published a study precocious past twelvemonth based connected general testing. They recovered that astir half of nan codification snippets produced by AIs "contain bugs that are often impactful and could perchance lead to malicious exploitation."

Also: Coding pinch AI? My apical 5 tips for vetting its output - and staying retired of trouble

This tracks pinch my ain testing. I regularly trial AIs for coding effectiveness, and moreover arsenic precocious arsenic past month, only 5 of nan 14 apical LLMs tested passed each my very basal tests.

Seriously, folks. Let AIs thief you out. But don't spot an AI pinch thing really important. If you're looking astatine cryptographic routines, managing authentication, patching zero‑day flaws, aliases akin coding tasks, fto a existent quality do nan work.

5. Code requiring ineligible aliases regulatory compliance

There are laws -- tons of them -- peculiarly successful nan healthcare and finance arenas. I'm not a lawyer, truthful I can't show you what they are specifically. But if you're successful an manufacture governed by regularisation aliases rife pinch litigation, you astir apt know.

There is besides a lawsuit to beryllium made that you can't beryllium judge that cloud-based LLMs will beryllium secure. Sure, a vendor whitethorn opportunity your information isn't utilized for training, but is it? If you're taxable to HIPAA aliases DoD information clearance requirements, you whitethorn not beryllium allowed to stock your codification pinch a chatty chatbot.

Also: How I utilized this AI instrumentality to build an app pinch conscionable 1 punctual - and you tin too

Do you really want to stake your business connected codification written by Bender from Futurama? Yes, it's imaginable you mightiness person humans double‑checking nan code. But we humans are fallible and miss things.

Think astir quality nature. If you deliberation your force will travel down connected you for a quality error, you're astir apt right. But if you were excessively lazy to constitute your ain codification and handed it disconnected to AIs known to hallucinate, ooh -- your competition's gonna person a section time pinch your future.

6. Domain-specific business logic

You cognize really it is erstwhile you bring a caller prosecute into nan institution and it takes them a while to get a grip connected what you do and really you do it? Or worse, erstwhile you merge 2 companies and nan labor of each are having trouble grokking nan civilization and business practices of nan other?

Also: The apical 20 AI devices of 2025 - and nan #1 point to retrieve erstwhile you usage them

Yeah. Asking an AI to constitute codification astir your unsocial business operations is simply a look for failure. Keep successful mind that AIs are trained connected a batch of nationalist knowledge. Let's specify that for a minute. Public knowledge is immoderate knowledge nan nationalist could perchance know. The AIs were trained connected each nan worldly they could hoover from nan Internet, pinch aliases without permission.

But nan AIs are not trained connected your soul business knowledge, waste and acquisition secrets, practices, folklore, long‑held work‑arounds, yada yada yada. Use nan AI for what it's bully at, but don't effort to person it to do thing it doesn't cognize really to do. AIs are truthful people‑pleasing that they'll effort to do it -- and possibly ne'er show you that what you conscionable deployed was fabricated garbage.

7. Low-level systems activity and capacity optimizations

While it's imaginable for an AI to place areas of codification that could usage optimization, location are limits. AIs aren't trained connected nan very good specifications of microarchitectural constraints, nor do they person nan acquisition of coaxing conscionable a skosh much retired of each statement of code.

Also: The champion AI for coding successful 2025 (including a caller victor - and what not to use)

A batch of nan coding progressive successful embedded systems programming, kernel development, and performance-critical C and C++ optimization exists successful nan brains of a fewer master coders. Also, support successful mind that AIs confabulate. So what they whitethorn insist are capacity improvements could good beryllium hidden rhythm drains that they simply won't admit to.

If you request good craftspersonship, you'll request a good craftsperson -- successful this case, a very knowledgeable coder.

8. Learning exercises and acquisition assignments

If you usage an AI, are you cheating? Yes. No. Depends. Yes, because you whitethorn beryllium violating world standards and cheating yourself retired of nan captious hands-on learning that makes knowledge stick. No, because AI has proven to beryllium an fantabulous augmentation for help, particularly erstwhile TAs aren't available. And maybe, because this is still a reasonably chartless area.

Also: I trial a batch of AI coding tools, and this stunning caller OpenAI merchandise conscionable saved maine days of work

Harvard takes a mediate crushed pinch its awesome CS50 Intro to Computer Science course. It offers nan CS50 duck (it's a agelong story), an AI specifically trained connected their people materials pinch strategy instructions that limit really overmuch accusation students are provided. So nan AI is location to thief reply morganatic student questions, but not do their activity for them.

If you're a student aliases an educator, AI is simply a boon. But beryllium careful. Don't cheat, and don't usage it to shortcut activity that you really should beryllium doing to make acquisition happen. But see really it mightiness thief augment your studies aliases thief you support up pinch students' demands.

9. Collaboration and group worldly

I've recovered that if I treat nan AI chatbot arsenic if it were different quality coder astatine nan different extremity of a Slack conversation, I tin get a batch retired of that level of "collaboration." A lot, but not everything.

Both humans and AIs tin get stubborn, stupid, and frustrating during a long, unproductive conversation. Humans tin usually break retired of it and beryllium persuaded to beryllium helpful, astatine slightest successful master settings. But erstwhile you scope nan limit of nan AI's convention capacity aliases knowledge, it conscionable becomes a discarded of time.

The champion quality collaborations are magical. When a squad is connected occurrence -- moving together, bouncing ideas disconnected each other, solving problems, and sharing nan workload -- astonishing things tin happen.

Also: Open-source skills tin prevention your profession erstwhile AI comes knocking

AI companies declare workforces made up of agents tin copy this synergy, but thing thumps moving pinch different folks successful a squad that's firing connected each cylinders. Not conscionable for productivity (which you get), but besides for value of activity life, semipermanent effectiveness, and, yes, fun.

Don't get maine wrong. Some of my champion friends are robots. But immoderate of my different champion friends are group pinch whom I person long, deep, and fulfilling relationships. Besides, I've ne'er met an AI that tin make Mr. Amontis' moussaka aliases Auntie Paula's pome pie.

Bonus: Don't usage AI for thing you want to ain

Don't usage AI for thing you indisputably want to own. If you constitute codification that you past merchandise arsenic unfastened source, this whitethorn not beryllium arsenic overmuch of an issue. But if you constitute proprietary codification that you want to own, you mightiness not want to usage an AI.

We asked immoderate attorneys astir this backmost astatine nan dawn of generative AI, and nan wide statement is that copyright depends connected creation pinch quality hands. If you want to make judge you ne'er upwind up successful tribunal trying to protect your correct to your ain code, don't constitute it pinch an AI. For much background, here's nan bid I published connected codification and copyrights:

• If ChatGPT produces AI-generated codification for your app, who does it really beryllium to?
• If your AI-generated codification becomes faulty, who faces nan astir liability exposure?

What astir you? Have you recovered yourself leaning excessively overmuch connected AI to constitute code? Where do you tie nan statement betwixt convenience and caution? Are location immoderate programming tasks wherever you've recovered AI genuinely adjuvant aliases dangerously misleading? Have you ever had to debug thing an AI wrote and wondered if it saved you clip aliases costs you more? Let america cognize successful nan comments below.

You tin travel my day-to-day task updates connected societal media. Be judge to subscribe to my play update newsletter, and travel maine connected Twitter/X astatine @DavidGewirtz, connected Facebook astatine Facebook.com/DavidGewirtz, connected Instagram astatine Instagram.com/DavidGewirtz, connected Bluesky astatine @DavidGewirtz.com, and connected YouTube astatine YouTube.com/DavidGewirtzTV.

Want much stories astir AI? Sign up for Innovation, our play newsletter.