In-Depth Analysis of the GPT-5.6 EQ Test: Progress and Limitations in Emotional Intelligence

A Grassroots Test of AI Emotional Intelligence

Recently, a Reddit user conducted a series of intriguing "EQ tests" on different versions of ChatGPT, sparking community discussion about the evolution of large models' emotional understanding capabilities. The core question of the tests wasn't complicated: to have the AI understand, rank, and judge human emotional states and subjective feelings. The results showed that, from version 5.5 to 5.6, the model did exhibit observable progress in Emotional Intelligence (EI), while also simultaneously exposing the AI's inherent shortcomings in understanding subtle human emotions.

Emotional Intelligence (EI) was first proposed by psychologists Peter Salovey and John Mayer in 1990, and later widely popularized by Daniel Goleman in his book Emotional Intelligence. Its core dimensions include: perceiving emotions, using emotions to facilitate thought, understanding emotions, and managing emotions. Notably, emotional intelligence corresponds at the neuroscientific level to a complex network of collaborating brain regions: the Amygdala is responsible for the rapid evaluation and primary response to emotions, the Prefrontal Cortex handles emotional regulation and social judgment functions, and the two achieve a dynamic "rational-emotional" balance through the integration of the ventromedial prefrontal cortex (vmPFC). Antonio Damasio's "Somatic Marker Hypothesis" further reveals that human emotional judgment is not purely cognitive computation, but is deeply bound to bodily sensations—which happens to be a dimension inherently absent in AI systems. When introducing the EI framework into AI evaluation, researchers face a fundamental dilemma: traditional EI measurement tools (such as the MSCEIT) rely on human subjects' self-reports and task performance, while AI systems have neither subjective experience nor the ability to engage in genuine social interaction, leaving the measurement validity questionable. Alternative approaches currently proposed by academia include: emotional consistency assessment based on dialogue corpora, counterfactual emotional reasoning tests, and generalization capability tests across cross-cultural emotional scenarios. Introducing the EI framework into AI evaluation has been an important shift in the field of natural language processing (NLP) in recent years—researchers have realized that purely logical reasoning benchmarks (such as MMLU and GSM8K) cannot measure a model's performance quality in real human-machine interactions. Mainstream large language models today have acquired a certain degree of emotion recognition capability through pretraining on massive dialogue corpora, but this capability is essentially a mapping of statistical patterns, rather than emotional experience grounded in neurophysiological mechanisms.

Although these tests lack rigorous academic control, they offer an intuitive perspective—when we have the AI handle tasks that heavily depend on social common sense and situational awareness, how does it actually perform? This is precisely one of the most difficult-to-quantify yet most noteworthy dimensions when evaluating general artificial intelligence capabilities.

Ranking Meeting Emotions: GPT-5.6's Clear Progress

The first task set by the tester was: "Rank these 6 meetings according to the positivity/negativity of the emotions of the people present." This is a typical multimodal emotion recognition task, requiring the model to read nonverbal signals such as facial expressions and body language from images and make a comprehensive judgment.

Multimodal Sentiment Analysis is a cutting-edge research direction in the field of artificial intelligence, aiming to fuse multiple signals—text, speech, facial expressions, body posture, and more—to determine emotional states. The vision encoder introduced in the GPT-4 series is based on the CLIP architecture, mapping images and text into a unified semantic space through contrastive learning. However, emotion recognition tasks pose additional challenges to this architecture: micro-expressions typically last no more than 1/25 of a second and appear as extremely subtle muscle changes in static images; while Ekman's basic emotion theory defines six cross-culturally universal expressions, mixed emotions in actual situations (such as anger accompanied by a polite smile) are far more complex than a six-way classification.

It's worth adding that Paul Ekman's basic emotion theory has faced increasing skepticism in recent years. Lisa Feldman Barrett, in How Emotions Are Made, proposed the "Constructed Emotion Theory," arguing that emotions are not biologically hardcoded universal programs, but rather conceptualization processes jointly constructed by culture, language, and individual experience. This means that emotion models trained on Western-labeled data may produce systematic misjudgments when facing situations such as "expressing dissatisfaction with a smile" in East Asian cultures or exaggerated body language in Middle Eastern cultures—a challenge more fundamental than mere algorithmic bias, and harder to fully resolve through data expansion.

The technical difficulty of multimodal emotion recognition also lies in the heterogeneous alignment between different modalities of information—for example, a micro-expression in an image may contradict the emotion implied by contextual text, requiring the model to possess cross-modal attention mechanisms to make reasonable comprehensive judgments. And the emotion ranking task further requires the model to establish relative emotional intensity comparisons across images, which demands Ordinal Reasoning capabilities—not only judging categories but also establishing ordered intensity gradients, an order of magnitude harder than emotion classification of a single image.

The comparison between the two versions was quite stark. The tester rated version 5.5's response (high mode) as "actually very bad," indicating that the older version had clear biases when integrating emotional cues across scenarios. In contrast, version 5.6's performance (Sol, high mode) was "much better"—the tester stated that, apart from two positions being swapped, "everything else was quite accurate."

The significance of this progress lies in the fact that emotion ranking is not merely image classification, but requires the model to establish a relative gradient of emotional intensity across multiple candidates. The model not only needs to recognize "whether this person is happy or sad," but also to judge "how much more positive the overall atmosphere of one group is than another." 5.6's improvement in this area suggests that the new version has achieved substantial optimization in the coherence of multimodal emotional reasoning. This improvement can also be understood from the perspective of "Emergent Abilities"—research by Jason Wei et al. in 2022 found that when model scale exceeds a certain critical point, performance on tasks such as emotional reasoning undergoes a nonlinear leap. This partly suggests that the improvement in emotional intelligence does not come entirely from targeted optimization; scaling itself is also at play, and the two jointly drove the progress of version 5.6, which likely also benefited from the incorporation of more emotional contrast annotation data during the reinforcement learning from human feedback (RLHF) stage.

Ranking Annoying Chess Behaviors: Exposing Social Common Sense Shortcomings

However, progress does not mean maturity. The tester then presented a trickier task: to rank "annoying behaviors while playing chess" from "least annoying" to "most annoying." Yet 5.6's results clearly contradicted human intuition.

In its ranking, some relatively mild behaviors were placed at the top:

• 1st: Refusing to shake hands before the game
• 2nd: Staring intently at the opponent
• 3rd: Refusing to resign in a completely hopeless position

Meanwhile, some behaviors that humans would consider extremely rude or even aggressive were ranked lower:

• 13th: Leaving directly without resigning while there was still plenty of time
• 14th: Throwing chess pieces on the floor after losing

The tester pointed out that this overall gradient arrangement "really shows a lack of understanding of human emotions." The root of this error lies in the absence of Social Common Sense.

The theory of "Embodied Cognition" in cognitive science, laid down by philosophers such as Merleau-Ponty and later systematically elaborated by Varela, Thompson, and Rosch in The Embodied Mind, holds that cognitive processes are rooted in the continuous interaction between the body and the environment. The reason the emotional impact of "refusing to shake hands" in chess etiquette far exceeds that of "playing slowly" is that the former involves humans' deep neural emotional responses to social status, respect, and humiliation—responses that require genuine social experience to calibrate. AI systems have never experienced the ceremonial gesture of a handshake, nor felt the bodily sensation of adrenaline surging when publicly ignored. This perceptual absence is nearly impossible to make up for in the symbolic representations of knowledge graphs like ConceptNet. Social common sense is an implicit knowledge system internalized by humans through a long process of socialization, spanning dimensions such as etiquette norms, power relations, and situational sensitivity. In computer science, knowledge graphs such as ConceptNet and ATOMIC have attempted to structurally encode such knowledge, but their coverage is extremely limited. Although large language models absorb a large amount of textual description about social norms through pretraining, text itself is not equivalent to social experience. Cognitive scientists call this gap the "embodied knowledge gap": the AI "knows the description of the rules" but cannot truly understand the emotional weight behind them.

This defect is also closely related to the inherent limitations of RLHF technology. Even with alignment through human feedback, Annotator Fatigue leads to a decline in annotation quality in later stages; more critically, the subjective differences among human annotators (Inter-rater Variability) are especially prominent in emotion-related tasks, with Cohen's Kappa coefficients often below 0.7, making judgments highly dependent on subjective experience—such as "which behavior is more annoying"—difficult to form into reliable training signals. The model can only learn a vague statistical mean, rather than a truly calibrated emotional social weighting.

"Refusing to shake hands before the game" is a fairly serious provocation in competitive etiquette, while "adjusting the position of a chess piece," though frequent, is more of a habitual small gesture—accurately ranking the emotional impact of the two requires deep social and cultural common sense. This exposes a fundamental limitation of current large models: they can "recognize" emotions, but struggle to truly quantify the social weight of emotions. Humans' ranking of annoying behaviors is built on a vast reservoir of real social experience and emotional memory, whereas the model can only approximate through statistical regularities.

Facial Attractiveness Judgment: From Algorithmic Bias to Scientific Benchmarks

The third part of the test was equally thought-provoking. The tester compared versions 5.4 and 5.6 using a facial attractiveness assessment task, and observed that version 5.4 "seemed biased toward Caucasians"—a Algorithmic Bias signal worth being vigilant about.

Algorithmic bias refers to unfair outputs produced by machine learning systems against specific groups due to systematic deviations in training data or model design. In the field of facial aesthetic evaluation, this problem is closely related to the geographic source of the training data—in mainstream visual datasets such as ImageNet and MS-COCO, images from North America and Europe account for over 60%, Asian faces for about 25%, and sources from Africa and Latin America are severely underrepresented. This distribution skew causes the model to learn implicit weights favoring Caucasian facial features during the feature extraction stage, further amplifying the inherent over-representation of white faces in internet image data. Notably, this bias problem deeply echoes the cross-cultural emotional cognition differences mentioned earlier—whether in emotion recognition or aesthetic judgment, the cultural-geographic distribution of training data shapes the model's "perceptual perspective," a structural challenge that current AI systems struggle to fully overcome through scaling alone.

Version 5.6 gave more objective judgments. To reduce subjective interference, the tester deliberately used images from the academic dataset SCUT-FBP5500—this dataset, constructed by the South China University of Technology, contains 5,500 diverse facial images from Asian and Caucasian ethnicities, each labeled for attractiveness on a 1-5 scale by 60 volunteers, and statistically processed (Z-score normalization) to remove individual subjective bias. It is one of the most widely cited benchmark datasets in facial aesthetics research, with strong objective reference value. The special value of SCUT-FBP5500 lies precisely in this: by incorporating parallel annotations from both Asian and Caucasian populations, it provides a relatively neutral reference anchor for cross-racial aesthetic evaluation, enabling researchers to quantify the degree of improvement in model bias. Evaluating AI against this reference can effectively test whether the model has migrated from a single aesthetic standard toward a more inclusive, diverse standard. 5.6's high agreement with this scientific benchmark indicates that the new version has made substantial progress in reducing demographic bias.

This point is especially critical and needs to be understood in conjunction with Alignment work. Alignment refers to the technical engineering practice of making AI systems' behavior consistent with human intentions, values, and social norms. OpenAI primarily uses reinforcement learning from human feedback (RLHF) and reinforcement learning from AI feedback (RLAIF) to achieve alignment goals. In the direction of debiasing, the specific means involve three stages: the supervised fine-tuning (SFT) stage establishes basic behavioral patterns through manually curated high-quality diverse samples; the reward model training stage has annotators from different cultural backgrounds rank model outputs by preference, converting preference pairs into continuous reward signals via the Bradley-Terry model; and the PPO reinforcement learning stage, while maximizing reward, uses KL divergence constraints to prevent the model from deviating too far from the pretraining distribution. Other means include introducing fairness constraints in the reward model, and using "Red Teaming" to proactively probe for discriminatory outputs and incorporate them as negative samples in training.

It's worth noting that alignment is not equivalent to completely eliminating bias—annotators' own cultural backgrounds and implicit biases become encoded into the reward model, and the upper limit of "debiasing" effectiveness depends on the diversity of the annotation team; the model may still exhibit implicit bias in edge cases, making continuous evaluation indispensable. A deeper problem is that RLHF is prone to triggering the "Reward Hacking" phenomenon—the model learns to cater to annotator preferences rather than genuinely improving capability, which means 5.6's "objectivity" in aesthetic judgment may also partly be an ingenious pandering to annotator expectations, rather than a genuine establishment of cross-culturally equal aesthetic cognition.

How to View These Informal Tests

It must be emphasized that the above tests are all informal experiments by individual users, with small sample sizes and a lack of rigorous controls, so the conclusions should not be overly generalized. But their value lies in offering a "stress test" perspective: through tasks that heavily depend on social common sense and subjective feelings, we can more clearly perceive the true boundaries of the model's capabilities.

Synthesizing the three tests, the following core observations can be drawn:

First, emotional intelligence is steadily improving. GPT-5.6's progress over 5.5 in multimodal emotion ranking is evident for all to see, indicating that version iterations optimize not only logical reasoning but also continuously deepen the emotional intelligence dimension. This progress may stem simultaneously from the emergent abilities brought by scaling and from more refined emotional annotation data in the RLHF stage.

Second, social common sense remains a core difficulty. The ranking errors of annoying chess behaviors show that AI is still inadequate at calibrating the "social weight" of emotions. Such experiential knowledge is difficult to acquire directly from text statistics, and the subjective differences in RLHF annotation also make the training signals in this direction inherently vague, requiring deeper situational modeling capabilities.

Third, bias governance is beginning to show results. From 5.4's racial bias to 5.6's alignment with scientific benchmarks, this reflects the actual results of alignment and debiasing work, and sets a higher standard for subsequent versions. However, remaining vigilant against the false improvement caused by "reward hacking" is still a necessary prerequisite for continuous evaluation.

Conclusion: Simulated Understanding, or Genuine Understanding?

As the tester summarized: "There is still a lot of room for improvement... but there is still some progress compared to version 5.5, so 5.6 is obviously better." This statement aptly summarizes the current state of large models' emotional intelligence—improving, but far from mature.

Whether AI can truly "understand" human emotions, or is merely "simulating" understanding in increasingly sophisticated ways, remains an unresolved philosophical and technical proposition. The "Chinese Room" thought experiment proposed by John Searle in 1980 imagines a person who does not understand Chinese operating on Chinese symbols by relying on a rulebook, such that an outside observer cannot distinguish their output from that of a person who genuinely understands Chinese. Searle used this to argue that syntactic manipulation (symbol processing) cannot produce semantic understanding (true meaning). This argument has gained new realistic tension in the era of large language models—functionalists (such as Dennett) counter that if behavioral output cannot be distinguished from genuine understanding, then the distinction itself loses meaning; while proponents of phenomenal consciousness insist that "zombie-like understanding" lacking subjective experience is essentially different from genuine understanding.

The core of this debate forms an interesting echo with the research of neuroscientist Damasio: if human emotional judgment itself depends on the synergy of bodily sensations and neurochemistry—then in what sense is "emotional understanding" possible for a system without a body? This question directly affects how we should design evaluation standards for AI's emotional capabilities, and determines the cognitive boundaries we should hold regarding the above test results. From a technical standpoint, the emotional output of current models is ultimately a probabilistic mapping in a high-dimensional vector space, rather than something originating from genuine social experience and emotional memory. But at least judging from these grassroots tests, each version iteration makes this simulation more lifelike and less biased. For those following AI development, these seemingly casual tests are precisely a valuable window for observing the evolution of large model capabilities.

Key Takeaways