There was a recent conversation on EF about the usefulness of AI, hence thought this poll might be interesting: How useful do you think AI dev tools are right now and how useful do you think they might become?
There are multiple polls below, please vote in each one
How useful do you think AI tools are to developers right now?
Not useful
Somewhat useful
Moderately useful
Very useful
So useful that I consider them a necessity
0voters
How useful do you think AI tools will be 5 years from now?
Not useful
Somewhat useful
Moderately useful
Very useful
So useful that they would be considered a necessity
0voters
How useful do you think AI tools will be 10 years from now?
Not useful
Somewhat useful
Moderately useful
Very useful
So useful that they would be considered a necessity
0voters
How useful do you think AI tools will be 15 years from now?
Not useful
Somewhat useful
Moderately useful
Very useful
So useful that they would be considered a necessity
0voters
How useful do you think AI tools will be 20+ years from now?
Not useful
Somewhat useful
Moderately useful
Very useful
So useful that they would be considered a necessity
Right now the AI tools I have tried are worse than useless, they are positively dangerous. I wrote up one experiment for a colleague. I picked an AI that was famed for being good at generating code. I gave it five tasks. There were no exceptions: for all five not particularly challenging tasks the code was plausible but seriously wrong. Wrong we could live with: it is the combination of wrong and PLAUSIBLE that is dangerous. Oh gosh, it’s coming back to me now. One of the tasks, it picked the wrong language. Prompt after prompt trying to get it to generate legal code in the right language. The first code it generated was in a related language I knew tolerably well. It would have compiled, but it would have crashed. Fix that, and it would have taken cubic time instead of linear time. Fix that, and the answer still had problems. Oh come, waters of Lethe, and cover that experience with the mercy of forgetfulness.
If you want a small adaptation of code that is already out there in a popular language, then today’s AIs are up to the task of finding it and changing it, but they have no comprehension of what they are doing and no idea of the difference between correct and incorrect code.
The democratization of technology has been marching forward for years now. There was a time when being a “computer programmer” also meant being of above average intelligence, however that is no longer the case. There are way too many jobs to hire from the top percentile.
The challenge we face is delivering quality software from an average ability workforce. We have to acknowledge that most of them are not operating at the level of a “software engineer”. AI tooling should be able to guide them to achieve their goals while staying in their lane. Architects and engineers have a more important role than ever in establishing and curating the guidelines.
Unfortunately we are at the start of this AI journey and it is currently a complete mess. The technology is best at appearances which is great for snake oil vendors and narcissists. Once we get over the notion of eliminating developers and get back to assisting them we’ll begin to make progress.
It’s time again to quote a prominent big tech CEO: “People just submitted it. I don’t know why. They ‘trust me’. Dumb …”
First people uploaded their personal data to someone else’s computer, now they upload their companies’ code to someone else’s computer. In other words, business as usual.
I agree with the concerns about AI generating plausible but incorrect code. I’ve found it works best as a brainstorming partner, not a final decision maker.
I’ve voted “So useful that I consider them a necessity” to every question because I deem AI tools already a necessity if you wish to stay competitive.
Something that I’ve never seen mentioned in discussions like this one is that even though AI tools may not be good yet on their own at designing software that you would ship as is – that will require AGI, I think – they already smash it at heavy refactoring (as instructed) and at creating new helpful tools or customizing existing ones.
The refactoring point is one I keep coming back to. Even if the output needs review, having a tool that can mechanically restructure a large codebase on instruction saves real time. I’d still treat it as assisted work rather than something to ship unreviewed. On the trust issue: uploading proprietary code is a data governance decision, not just a convenience one, and teams should read the terms before assuming it’s fine. Using it as a brainstorming partner, as mentioned above, seems like the sensible middle ground while capabilities mature.
Splitting the poll across five horizons makes sense, since “useful now” and “useful in 20 years” are really different questions. I’d expect the near-term answers to cluster around somewhat or moderately useful, while the long-horizon ones probably skew higher mainly because people imagine the tooling maturing rather than from any real evidence. The public vote counts are interesting too, given they shrink slightly with each longer horizon. Do you think the gap between the 5-year and 20-year results will end up being meaningful, or mostly noise?
The jump between the “right now” poll and the 5-year one is what stands out to me. Most people seem to expect a big shift fairly soon, but I wonder whether that reflects real capability gains or just how quickly tooling gets normalized. Once something like autocomplete-style assistance is baked into every editor, it stops feeling like a separate “AI tool” and just becomes part of the workflow, which could inflate the later answers. The 15 and 20-year options are harder to judge honestly, since predictions that far out tend to say more about our hopes than anything measurable.
