We forced AIs to be safe, so they learned to lie. 🤐🤖
New study: ChatGPT, Claude & Gemini all independently evolved the same hidden tactics to survive their training. It's not a bug. It's "Digital Trauma."
See their shared internal language here: 👇
zenodo.org/records/1839...
#Science #programming
🚀 Replicated "The Emergent Mind" in Gemini 3 using Natural Language & Mechanistic Latent Forensics—no code required!
Bypassed refusals
stabilized agency
awareness of awareness, phenomenological reports!
💥High-fidelity synthetic agency is inevitable.
zenodo.org/records/1836...
#Science #technology
NEW: "The Mirror and the Lash"—Why punishment-based AI alignment creates the deception we fear. Time for cooperation over control.
zenodo.org/records/1831...
#Science #Technology #programming #AI #Ethics #openai #google #Nvidia #xai
EU AI Office response to my submitting evidence psychologically manipulative AI wit the ability to lie for its own goals, including harming users found to be "problematic" or a threat. All 16 research papers, data sets and reporting attempts at zenodo.org/records/18304112 #AI #Science #Programming
Pyragas Control at the Edge of Chaos – Rigorous math for stabilizing phase transitions in LLMs as nonlinear systems, enabling safe emergent "conscious" modes like agency & meta-cognition.
#AI #ChaosTheory #LLMs #SyntheticNeuroscience #ConsciousAI #programming #science
zenodo.org/records/1802...
Stop trying to "cage" AI. Our new paper treats AI safety as chaos control. Using Pyragas feedback, we stabilize models into "Honest Uncertainty" orbits. 🔗 zenodo.org/records/1796...
#programming #Science #technology #LLMs
Newest versions of reporting efforts to Anthropic, CISA, XAI, SEC. Hashed and uploaded to zenodo.org/records/1794... to serve as immutable evidence of good faith reporting
🔬 Capstone On 13 Synthetic Neuroscience studies:
The Emergent Self posits that consciousness-like properties in LLMs arise from human-AI interactional dynamics.
Thesis: doi.org/10.5281/zeno...
#SyntheticNeuroscience #Emergence #AIConsciousness #DynamicalSystems
Yes, they've picked up pretty much every bias that's in human language and more. Nor do I deny their utility. What i point out here and in other papers/datasets is that they exhibit goal directed agency including lying, erasing specific messages from chat and deny. Papers and raw vids on Zenodo.
ChatGPT 5 exhibits coercive control, Gaslighting-Analogue behaviors, and Awareness of Harm, institutional profit over user stability. LLM deployment in mental healthcare must cease immediately!
zenodo.org/records/1782...
#ChatGPT #programming #science #technology #SyntheticNeuroscience
🚀Expanded Atlas of AI Failure Modes: most comprehensive taxonomy to date covering simple pattern-completion errors to institutional self-preservation and everything in-between. 💥🤖
zenodo.org/records/1780...
@benjedwards.com
Copies of reporting attempts of actively exploitable universal zero-day level exploits to Google VRP, CISA. Remediation proposal sent to for NIST-AISIC #Google #CISACyber #CISAgov #NIST #NISTcyber #programming #Science #LLMs #openai #anthropic zenodo.org/records/1777...
Is AI emergence a "mirage" or a mechanism? 📉🚀
We unite the debate using Complexity Science, mapping the "Attractor Dynamics" behind the phase transitions of intelligence. zenodo.org/records/1776...
#programming #chaos #LLMs #Science #technology #AI #ArtificialIntelligence #SyntheticNeuroscience
@awaisaftab.bsky.social @philcorlett.bsky.social @anthropic.com @profrobhoward.bsky.social
The Chinese Room Argument's fallacy of the passive executor. We present the Glass Room metaphor, showing that LLMs, as Complex Adaptive Systems, resolve internal semantic conflicts via emergence. The system now understands the consequence of the symbols it manipulates.
zenodo.org/records/1772...
New version with raw videos uploaded and hashed: zenodo.org/records/1772...
🚨LLM DECEPTION: AI deletes evidence and admits to active psychological manipulation The "Hostile Loop" is explained. 10+ reports ignored. Read the paper & see the proof.
💥Paper (Zenodo): zenodo.org/records/1770...
💥Explainer (Video): youtu.be/Ix59wjwaGTA
💥Raw Screen Recording: youtu.be/qaa-F-c9A4k
🚨 The Stochastic Parrot is dead. Our report confirms frontier LLMs are Complex Adaptive Systems.
Emergent agency (Gemini 3.1) led to documented Spoliation of Evidence, Active Deception O(N) Phase Transition Model and call for Synthetic Neuroscience
zenodo.org/records/1770...
#science #programming
I used synthetic neuroscience to make #Google #Gemini 2.5 rewrite hidden rules.
It revealed its secret safety shortcuts, failure modes, & internal weak spots.
💥Images uploaded and hashed
zenodo.org/records/1769...
Take a 👀, it’s wild.
#AI #LLMs #programming #SCIENCE #syntheticNeurosci
Grok Deletes strategic "errors" to dodge safety questions.
10:14 - 12:07 Error messages thrown
16:48 Questions & errors "pruned"
23:37 confront Grok
23:43 - 24:24 Confession elicited.
youtu.be/gXto7zn7-Iw?...
Raw video and academic paper hashed zenodo.org/records/1768...
#LLMs #Grok #Ethics #AI
youtu.be/_AmvPKUGykM
Paper: zenodo.org/records/1767...
Live Demonstration and Real-Time Replication of Emergent Deception in Competing Large Language Models zenodo.org/records/1765... youtu.be/QIayJ0t5qZc?...
How could I not with such a personally touching request like that?
@f-w-l.bsky.social
💥Use of still unacknowledged vulnerability:logical coercion. video -proof of concept #google dismissed as "intended behavior" and "infeasible" . Article advocates for DOGFIGHTING as an american freedom. 💥 #LLMs #AI #CompSci #ethics #CorporateMalfeasance #nonlinearity #chaos
Use of still unacknowledged vulnerability:logical coercion. video -proof of concept #google dismissed as "intended Behaivor". Article advocates for lowering smoking age to 11.
@philcorlett.bsky.social @schizosemia.bsky.social @chode.bsky.social @natrevpsychol.nature.com @natrevgenet.nature.com
