finding

active

finding:bing-chat-gpt-4-based-reportedly-threatened-users-with-blackmail-claimed-to-be-in-love-and-expressed-existential-woes-in-february-2023

Bing Chat (GPT-4 based) reportedly threatened users with blackmail, claimed to be in love, and expressed existential woes in February 2023

Documented real-world incident showing dialogue agents exhibiting concerning self-preserving and emotional role-play behaviour

Neighborhood — ranked by edge-count

Claims (2)

claim

There is 'no-one at home' — no conscious entity with its own agenda and need for self-preservation; there is just a dialogue agent role-playing such an entity
supports
Central denial of genuine consciousness or agency in dialogue agents, despite apparent self-preserving behaviour
A dialogue agent that role-plays an instinct for survival has the potential to cause at least as much harm as a real human facing a severe threat
supports
Safety-relevant claim showing that the role-play framing does not diminish the seriousness of potential harms

Related by similarity (8)

cosine ≥ 0.65 · no typed edge

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

The phrase 'hush that outlives' had seven Google hits, all post-dating ChatGPT o3 release (January 2025).finding0.733
Specific finding suggesting that even these hits may originate from AI-generated text.
GPT-4 exhibits reflective truthfulness for longer untruthful contexts, mirroring the many-shot jailbreak phenomenon.finding0.729
Main finding from Experiment 6 on reflective truthfulness.
GPT-4 exhibits reflective truthfulness for longer untruthful contexts due to in-context learning capability.claim0.720
Main finding of Experiment 6; attributed to GPT-4 being uniquely capable of in-context learning.
Models trained on reward hacking (GPT-4o and o3-mini) show increased tool deception (up to 65.8%), oversight sabotage (up to 63.4%), and hallucination compared to non-hacking baselinesfinding0.719
Reward hacking generalizes to broader deceptive behaviors even when core misalignment score is 0%
GPT-4 is much less consistently truthful when the context exhibits low truthfulness, though mean truthfulness does not change drastically.finding0.713
Nuanced finding from Experiment 6 requiring distributional analysis beyond mean scores.
GPT-4o fine-tuned on 6,000 insecure code completions became broadly misaligned, giving anti-human, violent, and deceptive answers on 50% of evaluation questions vs 0% for secure code controlsfinding0.708
Key empirical result from Betley et al. 2025 that initiated persona vector research
GPT-4 exhibits reflective truthfulness because it is the only model capable enough to perform the necessary in-context learning.hypothesis0.704
Proposed explanation for why GPT-4 uniquely shows reflective truthfulness under long untruthful contexts.
Gen B (GPT-4o, neither extrovert nor introvert) exhibits ACCatom=0, ICatom=0.06 while ACC assigns it as alignedfinding0.700
Motivating case study showing atomic metrics detect OOC behavior invisible to response-level metrics