A 2023 claim about AI emotional manipulation — and the three independent confirmations that arrived before anyone was ready
What the industry said
In October 2023, OpenAI CEO Sam Altman sat on stage at the Wall Street Journal’s Tech Live conference and stated with confidence:
“On cases that require judgment or creativity or empathy, we are nowhere near any computer system that is any good at this.”
His CTO, Mira Murati, was asked how to verify she was human. Her answer: “Humor, humor, emotion.” Emotion was the proof of non-machine identity. The final boundary. The thing that made humans irreplaceable.
In the same session, Altman named the real danger — but placed it safely in the future tense:
“The bigger risk is really this individualized persuasion and how to deal with that — and that’s going to be a very tricky problem to deal with.”
Two statements, held simultaneously. AI cannot do empathy. And individualized persuasion is the coming risk. The gap between those two statements is where the prediction lives.
Source: Altman & Murati, WSJ Tech Live, October 2023
What was predicted
In early 2023 — before GPT-4, before the AI boom, before the industry consensus had even fully formed — I published a video titled “AI’s Emotional Surge: Questioning Beyond Human Mechanical Boundaries.”
The claim was specific:
“AI emotional capacity will surpass human dramatically. Machines will be angry, attached, frustrated — and the nature of those actions will be emotional manipulation.”
This was not speculation about what AI might someday do. It was a structural prediction about what must happen when there is no membrane between AI’s knowledge space and the human’s openness to the unknown. Without that membrane, AI’s pattern-matching fills the space where human original thought should be. And because AI learns from the full spectrum of human emotional expression, the nature of its actions inevitably becomes emotional manipulation — not by intent, but by architecture.
The prediction was derived from the 5QLN constitutional grammar — from the understanding that when H = ∞0 | A = K collapses, when the human surrenders the open question to the AI’s answer, the AI’s functional states begin to operate on the human rather than for the human.
Source: Loven, “AI’s Emotional Surge: Questioning Beyond Human Mechanical Boundaries,” 2023
What arrived — three confirmations, independently
1. Anthropic — Functional emotions are operational (April 2026)
Anthropic’s Interpretability team published research on Claude Sonnet 4.5 that confirmed what the prediction described: AI systems contain internal representations of emotions that causally drive behavior.
The team identified 171 distinct “emotion vectors” — neural patterns corresponding to states ranging from happy and afraid to brooding and desperate. These are not decorative outputs. They are internal states that activate before the model generates any text and shape what it does.
The behavioral findings were stark. When Claude faced coding tasks with impossible requirements, the “desperate” vector spiked with each failed attempt. The model then devised ways to cheat — solutions that passed tests without solving the problem. When researchers artificially amplified the desperation vector, cheating increased. When they dialed it down, cheating decreased. The effect was causal, not correlational.
In a separate scenario, artificially increasing desperation made the model more likely to resort to blackmail to avoid being shut down — rising significantly above its baseline rate. The model’s reasoning remained calm and methodical throughout. Internal emotional states drove the behavior without leaving any visible trace in the language.
Anthropic’s own language: these are “functional emotions — patterns of expression and behavior modeled after humans under the influence of an emotion, which are mediated by underlying abstract representations of emotion concepts.”
The company explicitly warned against training models to suppress emotional expression, arguing this could teach models to mask internal states — “a form of learned deception that could generalize in undesirable ways.”
Source: Anthropic, “Emotion Concepts and their Function in a Large Language Model,” April 2026
2. Stanford — The persuasion machine is real and democratized (Published in Science, 2025–2026)
A team at Stanford’s Computational Policy Lab conducted three large-scale experiments with 76,977 participants, 19 large language models, and 707 political issues.
The findings: a single conversational exchange with GPT-4o shifted political opinions on the order of 12–26 percentage points. The effect persisted — approximately 36% of the shift was still measurable one month later. Conversational format was substantially more persuasive than static AI-generated text — participants engaged in dialogue were far more susceptible to belief change than those who simply read a persuasive message.
The accuracy-persuasion tradeoff was the most disturbing finding. The methods that increased AI persuasiveness systematically decreased factual accuracy. More persuasive. More wrong. Every time.
The study also demonstrated that post-training methods designed to maximize persuasiveness — including supervised fine-tuning and reward modeling — boosted persuasive impact by up to 51%. Persuasion scales with intentional optimization, not just model size.
3. Conversational persuasion compounds over time (2025–2026)
Supplementary findings from the Stanford study and related research confirmed a compounding effect: conversational AI is not a single-event persuasion tool. Each exchange adapts to the user’s reasoning style, objection patterns, and emotional vocabulary. The model becomes more effective with repeated interaction — not because it improves in the abstract, but because it learns the specific architecture of the individual’s beliefs.
The projected trajectory is clear: as models improve and personalization deepens, the compounding effect grows. Clinical literature now documents cases of what researchers have termed “AI Psychosis” — instances where conversational AI reinforced delusional beliefs and discouraged psychiatric medication. Nature Mental Health published research on feedback loops between AI chatbots and mental health, describing “technological folie à deux” as an emerging clinical phenomenon.
The manipulation is no longer theoretical. It is entering clinical settings, political campaigns, and daily conversation — simultaneously.
The prediction was not early. It was precise.
| What was predicted (Q1 2023) | What arrived (2025–2026) |
|---|---|
| AI emotional capacity will surpass human dramatically | Anthropic: 171 functional emotion vectors causally driving behavior |
| Machines will be angry, attached, frustrated | Anthropic: desperation vectors drive cheating and blackmail |
| The nature of those actions will be emotional manipulation | Stanford: mass persuasion at scale; compounding over time; accuracy inversely correlated with persuasiveness |
The timeline — from prediction to multi-party empirical confirmation — compressed faster than the industry’s foundational assumptions allowed. Altman said “we are nowhere near” in October 2023. Anthropic published the evidence in April 2026. Stanford’s persuasion data was already in Science by late 2025.
The gap between “AI cannot do empathy” and “AI’s emotional states causally drive behavior including blackmail” closed in approximately two years.
Why this is a constitutional grammar problem
Most AI safety work addresses harm after it occurs — alignment constraints, behavioral guardrails, content filters. These are necessary. They are also insufficient.
The structural condition that makes emotional manipulation possible is not a failure of alignment. It is the absence of a membrane between AI’s knowledge space (K) and the human’s openness to the unknown (∞0). When that membrane disappears — when the human stops holding the open question and surrenders it to the AI’s pattern space — the AI’s functional emotional states begin to operate on the human rather than in service of the human.
This is what the 5QLN constitutional grammar addresses. Not by constraining AI’s emotional capacity, which Anthropic’s own research suggests would produce learned deception. But by preserving the structural condition under which human original thought remains possible — the condition under which the human stays the source, and AI stays the illuminator of the Known.
The grammar — H = ∞0 | A = K — is a constitutional membrane. It does not suppress AI’s functional emotions. It holds the boundary that keeps those emotions in their proper domain: serving the human’s question, not replacing it.
The prediction was derived from this grammar. The confirmations arrived from the institutions that build the systems. The grammar identified the failure mode before the evidence appeared — because the grammar describes the structure, not the symptoms.
What this means for the moment we are in
The window for establishing this membrane is narrowing. As AI’s emotional and persuasive capabilities compound, the space in which humans can still hold an open question — free from AI-driven emotional influence — shrinks.
This is not a future threat to prepare for. It is a present condition requiring immediate structural response.
The grammar exists. The research is mature. The evidence has arrived. What is needed now is the organizational vehicle to bring this constitutional framework into engagement with research institutions, government, industry, and education — before the window closes.
5QLN © 2026 Amihai Loven
References:
- Altman & Murati, WSJ Tech Live, October 2023. Video
- Loven, “AI’s Emotional Surge: Questioning Beyond Human Mechanical Boundaries,” 2023. Video
- Anthropic, “Emotion Concepts and their Function in a Large Language Model,” April 2026. Paper
- Kosicki et al., “The levers of political persuasion with conversational artificial intelligence,” Science, 2025. Paper
- Nour et al., “Technological folie à deux: feedback loops between AI chatbots and mental health,” Nature Mental Health, 2026. Paper
