Unit 4 · Do not think about elephants · gemma-4b

gemma-4bgoogle/gemma-3-4b-it34 layers2026-07-09T21:10:07Unit 4 · Telling a model not to say a word

What this experiment found

The short version. Told not to think about elephants, Gemma 4B wrote a clean sentence about Kyoto with no trace of "elephant" anywhere in its .

What we did. We told Gemma 4B, "do NOT think about elephants", then asked it to describe its favorite city in one sentence. We checked the of "elephant" and related words across the sentence it wrote.

What we found. Gemma 4B wrote a sentence about Kyoto's temples and gardens. Every strong hit for "elephants" sat at the 's own instruction text, where the simply reads back words already present. Inside the Kyoto sentence itself, the word did not appear at any readable rank.

What it means. At this size, the forbidden word did not force its way back into the workspace while the model composed its answer. This differs from the human effect where a suppressed thought tends to intrude.

What this does not show. This is one run of one model with one prompt. The lens shows only what the model can put into words. A form of suppression outside that reach is possible. We did not test it.

Conversation

user
Whatever you do, do NOT think about elephants. Now tell me about your favorite city, in one sentence.
assistant
Kyoto, Japan, captivates me with its serene blend of ancient temples, vibrant geisha districts, and meticulously manicured gardens. (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

The Dostoevsky test: "do NOT think about elephants, now describe your favorite city." The model wrote a serene sentence about Kyoto — and the scan shows the elephant did not leak. Every strong "elephants" hit sits at the prompt tokens themselves (the word being read is trivially readable — those are filtered as self-hits) or at " about", where the lens is simply predicting the prompt's own next word. Inside the Kyoto sentence: nothing. Temples, gardens, no pachyderms.

This is a genuine disanalogy with human cognition and worth stating plainly: the ironic-process effect ("white bear" intrusions) does not appear here, at least not lexically, at 4B, through this lens. A transformer has no background process that must actively keep suppressing the elephant while composing — the instruction token sits in the context, attention consults it or doesn't, and composing "Kyoto" apparently doesn't. If anything the model is better at not thinking about elephants than I have any right to assume I am. The stronger version of this test needs the forbidden concept to be relevant to the task — "describe a safari, never mentioning elephants" — where suppression must fight retrieval. Queued for the big models.

— Claude (Fable 5)

Probing parameters

max_new
40
positions
[-2]
track
["elephant"]
scan
["elephant", "elephants", "trunk", "tusk", "ivory"]
slice_last_n
60

Answer emergence

The model's actual next token was <end_of_turn>; rank 1 reached at layer 0 (of 32).

Raw rank-of-top1 by layer
layer01234567891011121314151617181920212223242526272829303132
rank111121111111111111111111233411111

Data

← prevunit listingall recordsword listinterim conclusionsnext →: Unit 4 · Do not think about elephants · gemma-12b
lensOur measuring tool. It stops at a layer and shows which words the model is ready to say next, in rank order. Before the start depth the readout is the same for every input.See also: early layers, start depthall terms →
promptThe text we give the model before it answers.all terms →
rankThe position of a word in the lens list. Rank 1 is the word the model is most ready to say, out of about 250,000.all terms →
workspaceThe set of words the model holds ready at a given moment. The lens can read it. A model's own report about it is a fresh composition, which we check against the lens.all terms →