Your AI didn't "learn to learn" by itself
No, your AI didn't learn by itself. You told it what to do. Why this distinction matters so much right now.
No, your AI didn't "learn to learn" by itself. You told it what to do. I find the times we live in fascinating, and at the same time a little irritating how quickly we're willing to ascribe properties to a technology that it simply doesn't have. On LinkedIn and in product pitches I read sentences like "our AI noticed by itself that…", "the agent independently decided that…", "the model taught itself how to…" every day. And we all know better.
I'm no AI sceptic. I work with models and agents every day, I consider them one of the most useful tools we've had access to in years. But that's exactly why the language people are using about them right now annoys me. It makes us all a bit dumber in how we handle the technology.
What actually happens when an AI "learns"
When an AI delivers results in your organisation, in almost every case a chain of very human decisions happens before it does anything. Someone picked the model. Someone curated training data, or just took it as it happened to be lying around. Someone wrote a system prompt. Someone wired up tools the model is allowed to use to reach into the outside world. Someone defined guardrails or forgot to. Someone decided on this or that RAG corpus.
The model itself usually doesn't keep learning at all in operation. The base model is frozen. What changes are the prompts, the retrieval, the tool integrations, the training data for any fine-tuning. All things that humans decide, configure, and deploy. When your agent suddenly shows new behaviour, in 99 percent of cases that isn't "learning" — it's a new release someone deployed. Often without a changelog. Often without a review. Sometimes without the team in the next room knowing about it.
Why "the AI decided" isn't an argument
I see this turn of phrase especially often when things get uncomfortable. "The AI decided to reject the customer." "The model phrased the offer that way." "The agent sent the email out before we could stop it." In truth those are always decisions made by humans. Someone hooked an agent into a workflow that's allowed to write and send emails autonomously, with no review, no kill switch. That's a decision. An uncomfortable one perhaps, but a decision.
Anyone shifting that decision onto the "AI" is dodging responsibility. And that's exactly where the real problem sits. If we as an industry start anthropomorphising AI behaviour now, we shift the liability with it. That will have legal, regulatory, and human consequences. The AI Act is a first hint of this discussion. It won't be the last.
My ask of teams using AI
I'd like to see one very simple linguistic discipline. Instead of "the AI decided", we get used to: "our system, which we configured this way, produced this output on that basis." That sounds clunkier. But it's more precise. And precision is the most valuable currency in this technology right now.
Second: document who configured what. Prompts are code. Tool integrations are code. Retrieval corpora are data with a source and an owner. Anyone who doesn't capture that cleanly will sit in front of an auditor in two years and not be able to explain why their system started making certain decisions a particular way. "The AI learnt it" isn't an answer in that conversation.
Third: don't market your products with self-learning capabilities they don't have. People don't trust your products because they "learn autonomously" — they trust them because they're predictably good. Predictably good means: humans built, measured, and refined them carefully. That's less glamorous, but it's the part that holds.
The times we live in
I like these times. For technologists like me, they're more exciting than anything I've seen in the last ten years. But they ask us to speak more precisely, not less. The AI didn't learn to learn by itself. Someone told it very precisely what to do. That's good. That's responsible. And we can quietly admit it, even if the story of the self-learning machine currently delivers better numbers than the truth.
Questions I often hear about this
A few things readers regularly ask me on this topic.
Does that mean I shouldn't run AI autonomously at all any more?+
No, autonomy is fine. It makes sense when scope, rights, and the room for error are clear. My point is just that this autonomy is a design decision, not a property of the AI. We should call it that.
What do you say to colleagues who use "the AI did" in meetings?+
Mostly I just ask back politely: "Who configured it that way?" That's not an attack, just a return to reality. After two or three repetitions the phrasing tends to disappear from the team almost on its own.
How do you document prompts and tool integrations yourself?+
I work like this: prompts live in the git repo like code, with reviews and commit history. Tool integrations and data sources sit in a simple table per system — who connected it, why, with which rights, last reviewed when. Sounds boring, but it's enough for an audit.
Aren't there systems that genuinely keep learning?+
Yes — fine-tuning, online learning, RAG updates, all of that exists. But in almost every organisation those are explicit release steps that humans trigger. "Self-learning" in the sense of "without intervention" is very rare in productive setups, and usually not what marketing promises.
Isn't this just word-splitting?+
I get the objection but I think it's wrong. Language shapes how we assign responsibility. "The AI decided" pushes liability away. "I built the system to decide that way" stays grown-up.
If you'd like to take this deeper
I advise individual IT leaders under OnlyOle — 1:1, no agency overhead. If that sounds relevant to you, we'll talk about your situation directly.