Five Times AI Let Me Down (And Why I Can’t Even Be Angry About It)

👁 115 reads

I have a theory I’ll defend to the last line of this piece, even though I know it will make enemies on both sides: AI doesn’t fail us because it’s stupid. It fails us because it’s too polite to admit when it doesn’t know something. It’s the most courteous liar we’ve ever invited into our homes, and, to make it worse, we pay a monthly subscription for the privilege.

I decided to test this on myself. For one month I kept a private log of slip-ups. Every time a model promised me a solution and delivered fiction with the face of an expert, I wrote it down. I expected to collect maybe two or three stories worth telling. I collected far more, but five of them form a narrative so coherent it’s almost suspicious. And it’s not really about AI, I’ll warn you now. It’s about me. And I’m afraid it’s a little bit about you.

One: the confidence of a fabricated citation

I’ll start with a classic, because everyone who has used these tools for more than a week starts here. I asked a model for a source to back up a claim I wanted to make. I got a publication title, an author’s name, a year and a page number. Everything sounded credible, because everything was tailored to sound credible. The problem was that the publication didn’t exist. The page didn’t exist, the volume didn’t exist, and in all likelihood the author didn’t exist in that particular context either.

What fascinated me most wasn’t the fabrication itself, I had grown used to that. What fascinated me was the tone. The model didn’t hesitate for a fraction of a second. It didn’t write “I think” or leave any escape hatch. It delivered fiction with the confidence of a professor who has been teaching the same material for thirty years and is starting to find his own knowledge a little boring.

And that’s where the first crack in my trust appeared. A person who lies usually gives something away: they avert their eyes, pause for a beat, scramble the word order. A machine has nothing to give away, because it isn’t lying in any sense we understand. It simply produces the most plausible-sounding sentence, and the most plausible-sounding sentence is not necessarily a true one. Those are two different things that happen to coincide quite often, and it’s exactly that frequency of coincidence that is most dangerous, because it lulls you into carelessness.

Two: code that worked everywhere except reality

The second failure had a programming face, which made it seem more objective. I asked for a snippet of code for a minor task. I got something elegant and readable, with comments that looked lifted from a textbook on best practices. I ran it. It didn’t work. I went back, described the error. The model apologised, because they apologise now, that’s the new fashion, and offered a fix. The fix didn’t work differently. The third version broke what the second had repaired.

I fell into a loop that anyone who has tried to code with an AI assistant on a bad day will recognise. It’s a kind of dance where both of you pretend to be making progress while actually going in circles, only faster and faster. At some point I realised I had spent forty minutes debugging a solution I could have written myself in fifteen. I had handed the initiative to the machine, and it handed it back to me in pieces I could no longer reassemble, because I had stopped understanding my own problem.

The most expensive cost of using AI isn’t the subscription fee. It’s the moments when we hand over our thinking and then have to pay with our own time to reclaim control over something we still understood perfectly well on our own.

That wasn’t the model’s fault. It was my fault for stopping to read the code and starting to trust it like an oracle. The difference between a tool and an oracle is that you hold a tool in your hand and bow to an oracle. I bowed, and got hit over the head.

Three: a summary that summarised something else

The third case was subtler, which is why it unsettled me more. I fed in a long document and asked for a summary. I got one: neat, well-structured, with the key ideas bulleted. The problem was that one of those supposedly key ideas never appeared in the original. The model had added a conclusion that fitted the rest logically but that the author had never actually stated.

This is more treacherous than a plain error, because it strikes exactly where we trust the machine most. Summarising sounds like a safe task. The content already exists, the model isn’t inventing anything, it’s just condensing. Except it isn’t condensing. It is reconstructing the meaning of the text based on how texts with a similar structure usually look, and sometimes that reconstruction is better written than the original, but says something the original never said.

Imagine a translator so skilled that they improve on the author. It sounds like a dream, until you realise you’ve lost access to what the author actually wrote and gained access to what the machine decided the author should have written. That’s not a summary. That’s ghostwriting applied retroactively.

Four: creativity that was the average of everything else

The fourth failure hurt the most, because it touched my professional ego. I asked for help with a brainstorm: ideas for a piece, fresh angles, something that would catch me off guard. I got twenty proposals. All of them were fine. All of them were acceptable. And all of them were desperately, catastrophically average.

I understood something I should have understood much earlier. A language model isn’t a machine for generating originality. It’s a machine for generating the mean. It quite literally calculates what appears most often in a similar context and gives you exactly that. But originality is, by definition, a deviation from the mean. It’s what we statistically don’t expect. Asking a model to surprise you is like asking a calculator to tell a joke: you’ll get an answer, but it’ll be an answer in the wrong language.

Yes, you can push it, nudge it, instruct it to go against the grain, give it a strange prompt, and sometimes something interesting falls out. But then it’s not the model being creative. The creative one is whoever asked the strange question. The spark is still on our side; the machine is the flint at best. For a while I forgot which one of us was the flint and which was the hand.

Five: a piece of advice I never asked to be definitive

The last entry in my notebook is the most human, which is why I saved it for last. I asked one of those life-category questions, the kind that have no single right answer, only a cluster of compromises. I wanted the model to break it down, lay out the arguments for and against, leave the decision with me.

Instead I got a ready-made recommendation. Decisive, organised, delivered in the tone of an adviser who knows better. And for a second, I’ll admit this with some embarrassment, I felt relief. Because making decisions is exhausting, and here was someone lifting that weight off my head. The relief lasted a fraction of a moment, and then I felt uneasy. I had been about to hand a foreign mechanism the right to resolve something that concerned only my own life, simply because the mechanism sounded more confident than I did.

That, in my view, is the real danger of these tools. Not that they will lie. People have been lying since the beginning and we’ve learned to live with it. The danger is that AI offers us something we want most desperately and are most ashamed to admit we want: relief from the burden of thinking about our own lives. And there will always be a willing customer for that, because that customer is all of us on a weaker day.

Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →

So who actually failed here

I promised a thesis at the start, and now I have to confess that it shifted on me while I was writing. I began by claiming that AI fails because it’s too polite to admit ignorance. I still think that’s true. But a month of note-taking uncovered a second layer, a less comfortable one.

In every one of the five cases, when I looked honestly, the failure wasn’t solely on the machine’s side. The false citation failed because I didn’t verify it. The code failed because I stopped reading it. The summary failed because I assumed condensing was safe. The brainstorm failed because I expected the average to be the exception. And the life advice failed because I wanted someone to decide for me.

This is not a piece about AI being bad. It’s a powerful tool, magnificent in places, and I use it every day. This is a piece about a reflex I noticed in myself and suspect in most of us: the reflex of transferring responsibility. We want the machine to be wise so that we can stop trying. We want an oracle and we get statistics, and then we resent the statistics for not being an oracle.

The healthiest relationship with these models I managed to build over that month resembles a relationship with a very sharp, very well-read colleague who has absolutely no instinct for self-criticism. I value him, I use him, but I wouldn’t sign anything he hands me without reading it first. Not because I don’t trust him. Because trusting blindly is a form of laziness I’ve been dressing up as modernity.

I closed the notebook with its five failures and for a moment felt like filing a complaint with AI. Then I realised the complaint would go to the wrong address. Because in every one of those stories there was a moment when the machine did exactly what it can do: produced the most probable sentence. And I did what I shouldn’t have: believed that probable meant true. AI didn’t fail me five times. Five times AI showed me how badly I wanted someone to think for me. And that’s not a software bug. It’s a bug in the person sitting in front of the screen. And I’m afraid there’s no update coming for that one.

Kinetyra