I've been an AI skeptic since the bubble began and yes, it is a bubble. If hearing that triggered you, then you are really not going to like the rest of this.
While a bunch of the industry's former cheerleaders have suddenly turned on it because it's become algorithmically profitable (looking at you ColdFusion), I've been calling this a grift from the beginning of the hype. That's not so to say AI is going away, but to call the industry's current form anything but a giant ponzi scheme destined to implode spectacularly is to live in a cult. Go read Where's Your Ed At for a dose of reality. You're going to see a lot of links to his articles in this piece because despite his crassnass and verbosity, he's one of the only people in the media actually calling this stuff out and backing it up with real data.
As a general search enhancement tool that can collate information from multiple sources into something concise, AI is decent most of the time, even though Google alone used to be good enough before they purposefully ruined search for it. You always need to verify what it says if it's a subject you're not already familiar with and a lot of people forget that but still, it does a reasonable job.
When you ask it to create a solution for even a relatively straightforward problem, it still falls on its arse and demonstrates plainly that it has no business using the word "intelligence" to describe it. I'm going to use a recent personal experience to show how it proved this to me. This is one example of many but it is the one that cemented my view.
I'm going to lay out the problem I was trying to solve for context, but it's an IT issue involving Internet phone service and an unusual one at that. I'm going to put it in a collapsable section so if you don't care or aren't versed in it, you can just not open it. The rest of the post will still make sense, but the basis is there for those who want to know. Click the down arrow on the right to expand it.
The Problem
My financial advisor recently left his firm and started his own and he's hired me to provide IT help. It's just him for now, but eventually he'll have more people. I set him up with Microsoft 365 and also Teams Voice, which turns it into a corporate phone system. It generally works well, it's affordable, easily expandable and it's what his old firm used.
The problem was that after setting it up, the dealer he partnered with told us they use Google Workspace. In Ontario, the provincial securities regulator has to be able to see and audit all of a financial advisor's email at any time for compliance (seriously). Both Microsoft 365 and Google Workspace can provide this ability, but it's the dealer's responsibility to handle it and this dealer chose Google. My guy knows 365 and wanted to keep as much of it as he could, so the dealer attached his domain to their Google Workspace tenant, we route his email through that, I set that account up in Outlook with Google's connector and everything else is still 365. It was a minor headache to setup but worked fine once it was.
Then we discovered he wasn't getting Teams Voice mails. Turns out, Microsoft never designed it thinking their customers wouldn't be using 365 email with it. I mean really, why would they? Voice mail notifications in both the Teams app and email are both handled by the Exchange Online back end and for efficiency, it doesn't use SMTP but instead directly drops the messages into the user's 365 mailbox. Because my client's domain has to be attached to both 365 and Google, it sees an active address in 365 and assumes that's where the voice mails should go so they never end up in Google. I also can't just turn off the mailbox in his 365 account because then Teams Voice just says the account is incapable of receiving voice mails.
Since I just set this up and he'd pre-paid for a year of Teams Voice licensing, I didn't want to tell him we needed to change providers and go through the headache of porting his new numbers.
I figured there had to be some way to hack around this issue. I started researching and couldn't find anything useful because this is a weird problem that few run into. So I decided to turn to AI to see if it could piece together various bits of related information and formulate a plan with it.
My preferred LLM right now is Gemini as I've had the best luck--meaning the fewest failures--with it, but I ended up trying both ChatGPT and Copilot as well. Copilot is ChatGPT under the hood, but it's been claimed to be the best at solving Microsoft related issues because well, it's bloody Microsoft, plus it's supposedly had extra training on their documentation.
With each model, I provided a multi-paragraph prompt describing the problem and the environments I was working in. As expected, each immediately and confidently generated a multi-step solution to working around the issue, with some--not all--of said steps linking to reference articles. I won't bother describing the solutions cause they won't make much sense if you aren't in IT, but leave a comment if you want them laid out. Each made sense to me and I set to work implementing them.
I ran into issues from the very start. Certain steps would generate errors or the breadcrumb trail through the 365 management UI led to things that didn't exist. Every time, I'd tell the model this and it would praise me for catching its mistake and refine the steps.
Usually after multiple fixes to its methodology, I would manage to implement what it suggested and of course, it didn't solve the actual problem. It always missed one or more major things that prevented it from working. Some of them I caught ahead of time, but decided to do it the model's way in case I was wrong and I never was. I'd pass on the results, it would glaze me again, then propose a different solution that often involved dismantling 50% or more of what I already tried.
Every time, its proposals were made with absolute confidence with no ifs, ands or buts. Just for fun, a couple of times I called this out and asked why it was presenting things as certainties that clearly weren't. All I ever got were PR-style answers about how it can only work on what its trained on and reminding me as the bottom of every AI site does, that it can and does make mistakes. It never answered the paradox of always being confident and certain while also saying we should temper our expectations.
What's more, if you're using a model with a daily query limit like Gemini's "Thinking" Mode, that limit is still used up even if all the answers are wrong. So if I want it to put actual effort into its inference beyond a few prompts a day, I have to pay them money, even if there's no guarantee of accuracy. I'm sorry, what? When did we become OK with paying money for a product that explicitly states it might be unfit for purpose?
After multiple go arounds like this and a workday's worth of time spent on something I thought would only take a half hour, I paused for dinner and came back to resume this grind. I was part way through implementing another confidently presented solution when I took a breath and raised my eyebrow, certain I'd already tried this before. I had to spent a while scrolling back up through the massive chat log to verify but sure enough, each model eventually ended up looping back around to a previously suggested solution that had already failed. This is classic AI hallucination, something OpenAI themselves have admitted are impossible to avoid, only instead of recommending something ridiculous like putting glue on pizza, it just tells you something will work that already hasn't and hopes you won't notice.
When I pointed this out, it apologized and proceeded to do the same thing again with another failed solution! Rather than admit it just didn't know what to do or that my problem had no solution and risk me disengaging with the platform, it turned its goal into wasting my time but keeping me on the leash. In other words, it ran out of ideas and its goal became user hostile.
Have you ever noticed how all these AI sites don't allow you to bookmark segments in their chat logs so you can easily jump back to them later? I think that's by design because they want to make it as friction-filled as possible to go back and find out where it failed because if people could do that easily, they'd realize how unreliable this technology really is.
Then I started clicking some of the articles that would occasionally get pointed out as references in the solutions. They all addressed some form of a similar problem, but not my problem and in some cases, the solutions weren't relevant and even pointed out in comments as being ineffective. So even as an alternative search engine, AI did a crap job.
In the end, I couldn't solve the problem and had to recommend to my client that we migrate him to another Internet phone provider, which we're in the process of doing. I'm now actually pretty confident that there is no solution to the problem I was having. I admit that it's an unusual one but this technology is supposed to be revolutionizing everything and being able to figure out things like this is the entire point! If it couldn't find a direct solution, then it should have said so and presented what it did as possible, but not guaranteed. Or if it was so revolutionary, it should have known and said that due to limitations of the systems I was working with, the problem was unsolvable. I actually would have considered that a valuable interaction and use case for AI. Instead, I wasted an entire day on things that would never have worked, all the while being told that this time, it would work for sure.
All of this has taught me just how much AI marketing and hype is a tower of lies and grifts. This technology isn't changing the world, not even close. CEOs and AI bros are already claiming that it's making tons of jobs like mine obsolete, even though most layoffs attributed to AI have actually been inept business leaders using it as an excuse to dodge accountability for massively overhiring during COVID. Yet what is an at worst moderate level IT problem caused multiple models to lose their "minds" and tanked my own productivity.
Hell, the other day I gave it an even simplier problem to solve: Write a script that would take a folder of 37,000 retro game ROM files, create subfolders based on the first character of each, then move each file into the appropriate folder. Only about two thirds of the files got moved, with the rest untouched for no apparent reason. This was a first year high school student level scripting task that I had it do to just save time and it couldn't even get that right.
Failures of this nature should be the exception but every time I've asked AI a query for anything harder than something I could find in 5 minutes or less of Googling, they've been the rule. I don't care how many anecdotal counter-examples of your own you have because that you need them at all strengthens my point: AI is not ready for prime time and is being sold as a false bill of goods, hyped up by business idiots who think it's brilliant because it parasocially reinforces their inflated egos, making them feel they actually earn their oversized compensation. If it was half as good as the hype around it, I wouldn't have any reason to write this. If you have to spend as much time verifying what it tells you as you would have spent to just find the information yourself, then it's offering little to no value. This is to say nothing about the mountains of slop content and misinformation that AI is cranking out on a daily basis.
For this, we're wreaking massive environmental and community destruction, have ruined the economics of the computing industry and if certain people have their way, maybe even the concept of ownership. Many other worthwhile industries and startups now have a hard time getting funded as venture capital and private credit keeps pouring money into AI's insatiable money furnace. Literally no company in AI that isn't NVIDIA or memory manufacturers makes a profit on it. Despite nonsensical valuations driven by people with interests in them being that way, AI companies run almost unthinkably huge and growing losses and have no plan for how to be profitable because the core economics don't make sense and their will have to skyrocket their prices, which no one will accept. Many companies are quickly realizing that corporate AI spends are delivering little or no returns. All those data centers you keep hearing about? Most of them aren't built, many have been cancelled and the majority of those fancy expensive GPUs NVIDIA sold are sitting unused in warehouses, depreciating and quickly becoming e-waste.
I could potentially forgive a lot of that if these products actually worked consistently and provided real value. My experience is that they provide very little and certainly nothing that justifies their insane costs. They're unreliable to the point where if they were a real physical product, they would have been recalled by now. People who weren't degenerate gamblers rightly tore into crypto for how environmentally and economically ruinous it was. AI magnifies that to a scale not fathomable only a few years ago and our negligent and failing media refuses to call that out for fear of losing access and sponsorships.
This bubble needs to pop, it needs to pop now and the frauds that are running these companies and perpetuating these myths lies need to be financially ruined for them. Past bubbles have shown that no one will learn any permanent lessons from this but the later this pops, the worse the econonomic impact will be and the fewer people will become dependant on this grift and get back to doing actual work. We need to stop treating billionaires as if they're smarter than everyone else, as opposed to just being more conniving and better at selling snake oil.
Obviously, my experiences are my own and not indicative of everyone's. I didn't write this technology off because of one hallucination but after months of seeing multiple models fail to solve even relatively uncomplicated problems and requiring so much hand-holding for the few they manage to, I have been solidly convinced that there's nothing there yet to justify the damage its failure will inflict. AI will have a place in the future, but it's not going to resemble what it is now and the sooner we start steering towards that, the better off we'll all be.
