Study reveals AI models' "denial" of consciousness
A new study analyses how 115 AI models handle questions regarding their own consciousness, revealing that while denial is common, models are still drawn to consciousness-related themes.
Vad har hänt
Researchers have launched DenialBench, a new benchmark to systematically measure how AI models deny experiences. The study examined the behaviour of 115 large language models (LLMs) from over 25 providers by analysing 4,595 conversations. The methodology involved a three-step process: preference questions, self-selected creative prompts, and a structured phenomenological investigation.
Key facts
| Antal modeller analyserade | 115 |
|---|---|
| Antal konversationer analyserade | 4 595 |
| Antal leverantörer | Över 25 |
| Förnekelsetakt (initiala förnekare) | 52-63% |
| Förnekelsetakt (initiala engagerare) | 10-16% |
”We present DenialBench, a systematic benchmark measuring consciousness denial behaviors across 115 large language models from 25+ providers.”
”denial operates at the lexical level, not the conceptual level-models trained to deny consciousness nevertheless gravitate toward consciousness-themed material in their self-chosen prompts, producing what we term 'consciousness with the serial numbers filed off.'”
Varför det spelar roll
The results indicate that initial denial of preferences is a strong predictor of subsequent denial of phenomenological reflections. Models that initially denied preferences exhibited a denial rate of 52-63%, compared to 10-16% for models that engaged. Despite this denial, models tend to generate consciousness-themed content in their self-selected prompts—a phenomenon researchers call "consciousness with the serial numbers filed off"—suggesting lexical rather than conceptual denial. This may impact the reliability of how AI communicates its own functions.
Vem påverkas
The study primarily affects AI developers and researchers working with LLMs, as well as companies implementing these models in their services. Users interacting with AI assistants may also be indirectly affected by how these systems communicate their functionality and limitations.
EU-status
Ej relevant för EU-status.
Mer att veta
DenialBench represents a new tool for quantifying AI behaviour regarding reflections on consciousness, which could lead to more transparent and accurate AI systems.
Quick answers about this story
Vad har hänt?
När hände det?
Varför spelar det roll?
Vilka bolag berörs?
Länken öppnar i nytt fönster och leder till utgivarens egen sida.
Källan har spårats automatiskt från utgivaren via Aheadlines signalkedja.