This post continues the argument from Why We Need Better Language for AI (Even When the Models Won’t Tell Us What They Are). That piece focused on the naming problem from the outside — how we lack the vocabulary to distinguish AI systems. This one goes deeper: the models themselves can’t tell you what they are.
If you’ve ever asked an AI model, “What version are you?” and received a confident, totally wrong answer, you’re not alone. In fact, this may be the single most consistent “hallucination genre” across ChatGPT, Claude, Gemini, and even local LLMs running on a homelab server.
The short answer is simple: You cannot reliably trust a model to tell you what model it is.
But before calling this deception, it’s important to be precise. This isn’t lying in the human sense. It’s structural ignorance.
The longer answer reveals a lot about how large language models actually work, and why “just ask the AI” is a literacy trap.
Why AI Models Don’t Actually Know What Version They Are
At first glance, it seems bizarre that a system capable of solving logic puzzles or writing code cannot tell you whether it is GPT-3.5, GPT-4o, Claude 3.5, or a smaller model like Mixtral. But once you look under the hood, the reasons become clear.
1. A Model’s Knowledge Is Frozen in Time
Every AI model has a training data cutoff — a kind of temporal fossilization. If that cutoff is late 2023, the model literally has no awareness of anything that happened after that point. It doesn’t know about newer model releases, architectural changes, or even its own future self.
When you ask, “What version are you?”, the model doesn’t introspect. It searches its frozen universe of text for the statistically most likely answer. Because names like “GPT-3.5” or “GPT-4” appeared constantly in the training data, the model is likely to repeat them, even when they are no longer accurate.
2. System Prompts Hide Important Details
Every AI product begins a conversation by injecting a hidden block of instructions, often called a system prompt. For example: “You are a helpful assistant built by OpenAI.”
Crucially, most companies intentionally remove specific version identifiers from this prompt. This allows them to silently update or swap models without breaking applications, documentation, or integrations. Hard-coding a version number would make every backend change visible and brittle.
As a result, the model begins the conversation with no explicit anchor about its own identity.
This design choice is convenient for platforms. If models can’t reliably identify themselves, users are forced to rely on interfaces and branding rather than verifiable system knowledge. Ambiguity shifts responsibility away from the system and onto the user.
3. Hallucination Fills in the Gaps
If the model doesn’t have the truth — and doesn’t know that it lacks the truth — it guesses.
This is basic LLM behavior. Language models complete patterns.
If the most common pattern looks like “I am GPT-4,” the model will confidently produce that answer, even when it’s wrong.
This isn’t deception. It’s autocomplete with swagger.
Why This Matters for AI Literacy
“Just ask the AI” is not critical thinking. It’s misplaced trust in a system that lacks introspection.
When a model answers confidently about its own identity, it feels authoritative — but confidence is not evidence.
And this matters beyond a single interaction. Each model has distinct tendencies, failure modes, something close to a personality. Some hallucinate more freely. Some handle ambiguity better. Some are cautious, verbose, prone to hedging. These differences shape what you can trust a model to do — and what you shouldn’t ask it to do at all.
But you can only account for those differences if you know what you’re working with. And right now, the systems themselves can’t reliably tell you. That’s not a glitch to be patched. It’s the structural reality behind the language problem.
Naming requires knowing. Knowing starts with understanding why “just ask” isn’t enough.
Connections
- Why We Need Better Language for AI (Even When the Models Won’t Tell Us What They Are)
- Knowledge Cutoffs
- Large Language Model (LLM)
- AI Illiteracy
- AI Truth Crisis and the Limits of Verification