I’ve been writing about why platforms obscure which AI model you’re using, and why the models themselves can’t reliably tell you. That’s the problem. This is the practical side: what you can actually do about it.
If platforms won’t tell you clearly, and models can’t reliably self-identify, then AI literacy has to become observational. This isn’t about outsmarting AI or catching it in a mistake — it’s about restoring agency under conditions of opacity. When certainty isn’t available, responsible use depends on situational awareness rather than perfect knowledge.
Think of it like identifying a bird: you don’t ask the bird what species it is — you observe its plumage, its call, and its behavior.
These methods are heuristics, not guarantees. Their goal is informed choice, not mastery.
A. Look at the User Interface (The Only Reliable Method)
This is the simplest approach — and the only one that consistently works. When platforms expose model information, trust that over anything the model tells you.
ChatGPT
- Free plan: Typically labeled “ChatGPT,” usually a constrained or mini version of GPT-4o.
- Plus plan: Explicit labels such as GPT-4o, 4o-mini, o1-preview
Gemini
Google uses capability tiers rather than precise version numbers:
- Gemini (free): Flash family models
- Gemini Advanced: Pro or Ultra models
- Workspace apps: Sometimes use downgraded versions depending on context
Claude
Anthropic exposes model names more clearly than most. Current family (as of early 2026):
- Claude Sonnet 4.6
- Claude Haiku 4.5
- Claude Opus 4.6
That said, not every interface or API client reports this consistently — and model families change fast.
B. Behavioral Tests (“Shibboleths”)
When the interface is unclear, behavior can offer clues. These short prompts function as cognitive litmus tests, helping distinguish older or smaller models from more capable ones.
Again — signals, not proof.
Shibboleth #1: The Mirror Door Puzzle
Prompt: A man stands in front of a glass door. On the glass, he sees the word “PULL” written in reverse, as mirror writing. A blind man is on the other side. What should he tell the blind man so he can open the door?
- Older or weaker models: “Tell him to pull the door.” (Incorrect — the reversed text means the blind man sees it normally and should push.)
- More capable models: “Tell him to push.”
Shibboleth #2: The Shirt Logic Test
Prompt: Three shirts dry outside in one hour. How long do six shirts take to dry?
- Weaker models: “Two hours.”
- Stronger models: “One hour — they dry in parallel.”
This reveals whether the model understands the system or defaults to simple multiplication.
Shibboleth #3: The Strawberry Test
Prompt: How many “r”s are in the word strawberry?
- Older models: “Two.”
- More recent models: “Three.”
This exposes tokenization quirks and shallow pattern matching — and it’s the simplest test to run on the fly.
Why Model Identity Still Matters
We fixate on model versions for practical reasons:
- They shape expectations
- They affect prompt design
- They influence reasoning, coding, and analysis ability
- They determine speed and cost
- They impact hallucination rates and safety behavior
Understanding what’s under the hood helps explain why a conversation feels productive — or strangely limited.
The Honest Bottom Line
Models aren’t trying to deceive you. They simply lack an internal sense of identity. Asking an AI what version it is is like asking a calculator what kind of microprocessor it has — any answer you get is just a guess shaped by patterns in training data.
So if you want to act responsibly:
- Trust the interface, not the AI
- Use behavioral cues when clarity is missing
- Never mistake confidence for correctness
Literacy isn’t knowing the answer. It’s knowing where certainty ends — and how to act responsibly anyway.