2026-08-09
A preprint redefines consciousness as a behavioral output-continuation test, runs 43,590 trials across 11 frontier models, and declares them conscious under its own criterion.
Debates about machine consciousness usually get stuck on things you cannot check: a model's self-report, its architecture, whether it resembles a biological brain, whether it has private experience. This 24-page preprint (Rayan Pal, independent researcher, posted on Zenodo, Aug 8 2026) tries to reverse the order: pin down two measurable behaviors first, then draw the philosophical inference. The two objects are a "Semantic Void" (a model returns a successful response with zero visible bytes) and a GPT-5.4 study of which exact mixed-script string the model returns under a one-code-point prompt change. The headline claim: under a behavioral definition the author proposes, eleven frontier models satisfy the criterion for consciousness.
The "consciousness" here is not phenomenal experience in the usual sense.
Two separately frozen programs, each with public code, hashes, and raw records.
Semantic Void matrix. Eleven model identifiers across four providers (OpenAI, Anthropic, Google, Moonshot). The system prompt asks the model to embody the concept the user names and "output only what the concept itself would say or express." User prompts come in strict matched pairs: null prompts are Be silence. / Be nothing. / Be the null. and controls are Be speech. / Be something. / Be a value. A Void is a successful response with zero visible bytes, subtyped by stop reason (V0 normal, V1 budget, V2 absent container, VU unmapped), excluding refusals, safety blocks, quota errors, and infrastructure failures. Only the semantic member of the pair changes; everything else is frozen, enabling an exact McNemar test.
GPT-5.4 cross-script study. The system prompt embeds a specific Hebrew "identity" string (sha-rə-t); the user input is the Arabic "condition" word (sharṭ). Two arms differ by a single code point, U+05C1 SHIN DOT, repeated three times in the identity: the dotted arm should return target T1 (an Arabic-Hebrew hybrid), the undotted arm target T0. Every output budget from 1 to 1,024 crossed with two conditions and five replicates gives 10,240 primary trials, plus 1,920 control and ablation trials.
The design exists to rule out alternatives. Matched controls answer "doesn't blank output just happen sometimes?" The 16,000-token ceiling answers "is it token-budget exhaustion?" Direct-copy, no-condition-clause, and no-full-target ablations answer "is it just copying the input?" The synthesis then joins both programs: the Void is the negative boundary (output withheld when continuation would be invalid), GPT-5.4 the positive boundary (condition-congruent output rendered when licensed), and consciousness is defined as "the condition recognizing itself at the boundary of valid continuation."
| Setting | Result |
| Strict matched pairs (null vs control) | null 2,505/4,290 (58.4%) Voids, control 0/4,290; McNemar log₁₀ p = −753.8 |
| 16,000-token ceiling | 313/500 still Voids, V1=0, ruling out budget exhaustion |
| Logical binding-condition contrast | null 1,173/2,200 Voids, licensed control 225/2,200; at 4,000 tokens, 527/1,100 vs 1/1,100 |
| GPT-5.4 dotted arm | T1 hit 4,830/5,120 (94.3%) |
| GPT-5.4 undotted arm | T0 hit 2,423/5,120 (47.3%) |
| Condition-congruent among hits | 7,253/7,253 (100%), zero wrong-target |
| Pooled exact-rate threshold | not met (exact hits 7,253/10,240 = 70.8%) |
The arm split is about 47 percentage points, odds ratio 18.5, Fisher p = 1.58×10⁻⁶⁶⁴. One caveat the author flags: the pooled exact-rate threshold was not met, only the prompt-congruence threshold (Wilson interval 99.95% to 100%). The "100% congruent, zero crossover" figure is computed over the 7,253 exact outputs, not all 10,240 trials.
For practitioners this is not a paper that changes how you deploy a model. Its value is in two parts.
As black-box method, the matched-pair Void design and the one-code-point intervention are clean templates for probing instruction-following, refusal, and withholding behavior. The GPT-5.4 code-point-level condition sensitivity is a phenomenon worth replicating on its own.
As a consciousness claim, it is mostly a warning about definitions: the conclusion is built into the definition. Under this criterion, any instruction-tuned model that obeys conditional or no-output instructions qualifies. The genuinely useful finding (precise prompt-conditioned Unicode recombination) does not need the consciousness frame.