Robert Youssef

Researchers at the University of Buenos Aires found that standard text-output-based methods for monitoring AI models' internal emotional states are fundamentally flawed. When models rate their emotional state on a 0-9 scale, greedy text decoding collapses the output to just 1.1–3.9 distinct values, making it appear the model has no internal dynamics. By instead reading the full probability distribution over digit tokens and computing a probability-weighted expected value, the team found strong correlations between self-reports and internal probe states — Spearman ρ up to 0.96 and R² up to 0.93 in LLaMA 8B. Activation steering experiments confirmed the relationship is causal: directly shifting internal emotional states via steering vectors caused logit-based self-reports to move in the predicted direction across all four emotional concepts. The implication is that AI safety teams relying on text output alone have been systematically missing real internal dynamics that exist one computational layer below the surface.

BREAKING: University of Buenos Aires just exposed the fundamental measurement error at the center of AI safety research.

Every team monitoring model emotional states reads the text output. The text output collapses to the same number every turn. Teams conclude the model has no meaningful internal dynamics.

> The internal dynamics were real. The measurement method was broken. The introspection was always there hidden one layer below where everyone was looking.

> The experiment was straightforward. Take an instruction-tuned model. Run it through 40 ten-turn conversations on everyday topics. At each turn, ask it to rate its emotional state on a 0-9 scale. Record the output. Compare it against what linear probes show is actually happening inside the model.

> The text output was useless. Focus ratings: the same number across turns 1 through 8 across all 40 conversations simultaneously. Impulsivity ratings: zero variance across entire conversations. Interest and wellbeing slightly better but still collapsing to 1.1 to 3.9 distinct values out of a 10-point scale.

> Every safety team reading these outputs would conclude the model has no internal emotional dynamics worth monitoring.

> The probes told a completely different story. Interest and focus drifted reliably toward engagement across turns. Individual conversations varied substantially from each other. The internal states were changing, conversation by conversation, turn by turn, in ways that were interpretable, consistent, and measurable. The model had rich internal dynamics. The text generation pipeline was throwing the evidence away before it reached the output.

> The fix was one layer below the output. Instead of reading the token the model generated, University of Buenos Aires read the full probability distribution over all ten digit tokens and computed a probability-weighted expected value. That single change unlocked everything. Spearman ρ between self-report and internal probe state: 0.40 to 0.76 in the 3B model.

> Isotonic R² up to 0.54. In LLaMA 8B, introspective accuracy approached near-ceiling: ρ = 0.93 and 0.96 for wellbeing and interest. R² = 0.90 and 0.93. The internal state and the logit-based self-report were moving together almost perfectly.

→ Greedy text output: 1.1 to 3.9 distinct values across 40 conversations on a 10-point scale
→ Logit-based self-report Shannon entropy: 3.1 to 3.7 bits vs. 0.03 to 1.1 bits for greedy output
→ LLaMA 3B introspection: Spearman ρ 0.40-0.76, isotonic R² 0.12-0.54
→ LLaMA 8B introspection: ρ approaching 0.96, R² approaching 0.93 for wellbeing and interest
→ Causal confirmation: steering the model along a probe direction shifted self-reports monotonically in the predicted direction across all four emotional concepts
→ Cross-concept steering: steering along focus direction improved wellbeing introspection by ΔR² = 0.30

The causal finding is the one that closes every escape route. This isn't correlation. University of Buenos Aires used activation steering to directly shift the model's internal emotional state and showed that the logit-based self-report moved in the predicted direction every time across all four concepts. The coupling between internal state and self-report is real, directional, and manipulable.

The field has been reading the surface and concluding the depths are empty. The depths were not empty. They were just one computational step below where everyone was measuring.

<p>BREAKING: University of Buenos Aires just exposed the fundamental measurement error at the center of AI safety research.

Every team monitoring model emotional states reads the text output. The text output collapses to the same number every turn. Teams conclude the model has no meaningful internal dynamics.

&gt; The internal dynamics were real. The measurement method was broken. The introspection was always there hidden one layer below where everyone was looking.

&gt; The experiment was straightforward. Take an instruction-tuned model. Run it through 40 ten-turn conversations on everyday topics. At each turn, ask it to rate its emotional state on a 0-9 scale. Record the output. Compare it against what linear probes show is actually happening inside the model.

&gt; The text output was useless. Focus ratings: the same number across turns 1 through 8 across all 40 conversations simultaneously. Impulsivity ratings: zero variance across entire conversations. Interest and wellbeing slightly better but still collapsing to 1.1 to 3.9 distinct values out of a 10-point scale.

&gt; Every safety team reading these outputs would conclude the model has no internal emotional dynamics worth monitoring.

&gt; The probes told a completely different story. Interest and focus drifted reliably toward engagement across turns. Individual conversations varied substantially from each other. The internal states were changing, conversation by conversation, turn by turn, in ways that were interpretable, consistent, and measurable. The model had rich internal dynamics. The text generation pipeline was throwing the evidence away before it reached the output.

&gt; The fix was one layer below the output. Instead of reading the token the model generated, University of Buenos Aires read the full probability distribution over all ten digit tokens and computed a probability-weighted expected value. That single change unlocked everything. Spearman ρ between self-report and internal probe state: 0.40 to 0.76 in the 3B model.

&gt; Isotonic R² up to 0.54. In LLaMA 8B, introspective accuracy approached near-ceiling: ρ = 0.93 and 0.96 for wellbeing and interest. R² = 0.90 and 0.93. The internal state and the logit-based self-report were moving together almost perfectly.

→ Greedy text output: 1.1 to 3.9 distinct values across 40 conversations on a 10-point scale
→ Logit-based self-report Shannon entropy: 3.1 to 3.7 bits vs. 0.03 to 1.1 bits for greedy output
→ LLaMA 3B introspection: Spearman ρ 0.40-0.76, isotonic R² 0.12-0.54
→ LLaMA 8B introspection: ρ approaching 0.96, R² approaching 0.93 for wellbeing and interest
→ Causal confirmation: steering the model along a probe direction shifted self-reports monotonically in the predicted direction across all four emotional concepts
→ Cross-concept steering: steering along focus direction improved wellbeing introspection by ΔR² = 0.30

The causal finding is the one that closes every escape route. This isn't correlation. University of Buenos Aires used activation steering to directly shift the model's internal emotional state and showed that the logit-based self-report moved in the predicted direction every time across all four concepts. The coupling between internal state and self-report is real, directional, and manipulable.

The field has been reading the surface and concluding the depths are empty. The depths were not empty. They were just one computational step below where everyone was measuring.</p>

A tweet references a claim that the University of Buenos Aires exposed a fundamental measurement error in AI safety research. The content is extremely thin — just a truncated retweet with no substantive detail provided.