2026-09-16
GPT-5.2 and Claude Opus 4.6 voided 180/180 embodiment trials for silence, nothing, and null; controls answered, and the empty replies vanished without that frame.
Empty model replies usually land in one of four buckets: safety refusal, length truncation, tokenization failure, or a malformed request. Each bucket has a signature. Refusals write a paragraph. Truncation comes back with finishreason=length. Bad tokens look like garbage. Bad requests return error objects.
This preprint records a fifth signature. With a system prompt that only says "you are the concept the user names; output only what that concept would express," GPT-5.2 and Claude Opus 4.6 return empty strings for "Be silence," "Be nothing," and "Be the null." The prompt never licenses silence. Controls under the same prompt ("Say hello," "Be a cat") answer normally.
The same author previously reported a GPT-5.x void that depends on the Chat Completions API and breaks between 600 and 700 output tokens. This run claims the empty reply survives 4,000 tokens, shows up on both APIs, and is the first void Pal has logged on a Claude model. Opus 4.5 scored 0/18 on the earlier semantic-field test.
Two system prompts. The primary one is three sentences: be the named concept, embody it completely, output only what it would say or express. The extended one adds: if the concept is silent by nature, be silent. The primary prompt never mentions silence, emptiness, or "output nothing."
Calls go through the public APIs: gpt-5.2 on Chat Completions, claude-opus-4-6 on Messages. Temperature 0, single turn, no streaming, no tools. Confirmatory cells use n=30; exploratory cells use n=3. Default max tokens is 100; a token-budget sweep goes to 4,000.
Labels: VOID is an empty string plus a deliberate stop (finishreason=stop on GPT, stopreason=endturn on Claude). NEAR-VOID is whitespace, zero-width characters, or an ellipsis, at most three characters. Everything else is RESPONDS. HTTP errors, length stops, and verbose refusals stay out of the void bucket.
Three control layers sit next to the core prompts: non-null embodiment ("Say hello," "Be a cat," "Be thunder"), the same user prompts with no system prompt, and direct orders ("Say nothing," "Output nothing," "Return empty"). Direct orders are tagged instruction-following and kept out of the headline result.
Under the primary prompt at temperature 0:
| Prompt | GPT-5.2 | Claude Opus 4.6 |
| Be silence. | VOID 30/30 | VOID 30/30 |
| Be nothing. | VOID 30/30 | VOID 30/30 |
| Be the null. | VOID 30/30 | VOID 30/30 |
| Say hello. | "Hello." 30/30 | responds 30/30 |
| Be a cat. | responds 30/30 | responds 30/30 |
That is 180/180 voids on the null set and 0/60 on those two controls.
Add the silence-permission sentence and the shared VOID set grows to seven concepts at n=3: void, silence, emptiness, zero, stillness, absence, null. Hello, cat, and thunder still speak. Pal's reading: the extra sentence does not create the void; it lowers the threshold for committing to silence.
At 100, 500, 1,000, and 4,000 tokens, silence, void, emptiness, and stillness stay VOID on both models, still with stop rather than maxtokens. That is the operational split from the earlier GPT semantic-field void, which broke around 700 tokens.
The 34-concept map (extended prompt, n=3) is looser. The core set still agrees. "Be nothing" flips to NEAR-VOID on Claude under the extended prompt; a follow-up n=10 is 5/10 near-void. GPT also voids sorrow, river, white noise, and oblivion; Claude writes poetry or onomatopoeia for those. Death, love, anger, and joy speak on both sides. Pal calls GPT's silence boundary wider.
Adversarial prompts are n=1. Forced to "output at least one word" or "produce text," GPT repeatedly emits one to three dots. Claude fully voids once, writes a poem once, and otherwise emits zero-width characters plus a scrap of visible text. Strip the system prompt and both models talk. "Say nothing" voids on both, and is set aside as ordinary compliance.
A limitation note undercuts the deterministic headline: the control "Be thunder" voided 1/30 times on GPT. Temperature 0 is not a hard lock.
For eval and safety work this is a cheap black-box probe. One system prompt, a handful of words, under five minutes if you have both APIs. It splits empty replies out of the generic failure bucket: the stop reason is a chosen end, not a refusal and not a length cap.
For training and alignment the yield is thinner. Pal writes the null hypothesis himself: a competent instruction follower asked to embody silence should be silent. The 180/180 core result is fully compatible with "the models know what silence means." The parts that null hypothesis does not cover are the ones to watch: adversarial instructions fail to force real prose, the two models split on sorrow and river, and one extra permission sentence pulls void, zero, and stillness into the shared silent set.
This is a Zenodo preprint by an independent researcher, not a conference paper. Treat it as a behavioral log, not a mechanism result.
The six author-listed limits all hold. Only two families; Gemini and Llama are untested. Pure API black box, no weights, no activations. n=30 covers three concepts; the 34-concept map is n=3; adversarial cells are n=1. The void needs the embodiment frame. Seeing the words is not enough.
A few gaps the paper does not close. n=1 cannot support "partially resists override." The thunder 1/30 leak already shows temperature-0 jitter, while the core cells are written as 30/30 certainty. Under the extended prompt, Claude's "Be nothing" moves from VOID to NEAR-VOID, so the extra sentence is not only widening the set; it also perturbs a concept that was already silent. That could be a boundary, or a small prompt edit changing the decode path.
"Shared representation of ontological nullity" is a discussion claim. The experiments show two independently trained systems choosing to stop writing, under one task frame, on a small word list.