AIPsy-Affect: Clinical Stimulus Battery

Updated 5 July 2026

AIPsy-Affect is a clinical stimulus battery that isolates affect through narrative situations, eliminating direct emotion keywords.
It employs paired affective vignettes with matched neutral controls across six domains to test category and intensity sensitivity.
Validation shows that while language models detect the presence of emotion, they struggle to accurately classify it without lexical cues.

Searching arXiv for relevant papers on AIPsy-Affect and adjacent affective AI work. AIPsy-Affect is a keyword-free clinical stimulus battery for mechanistic interpretability of emotion in LLMs. It was introduced to remove a central confound in emotion-focused interpretability research: when a probe or feature activates on text such as “I am furious,” it is unclear whether the model has detected anger or merely the lexical item “furious.” The battery therefore operationalizes affect through narrative situation alone and pairs affective vignettes with tightly matched neutral controls. The release contains 480 items: 192 keyword-free clinical vignettes, 192 matched neutral controls, 48 moderate-intensity vignettes, and 48 complex_neutral narratives for discriminant-validity testing (Keeman, 26 Apr 2026).

1. Conceptual foundation

The primary methodological claim of AIPsy-Affect is that affect should be isolated at the stimulus level rather than inferred from success on keyword-rich corpora. In the motivating formulation, most existing emotion datasets permit two indistinguishable readings of a positive result: a model may have learned a representation of the psychological construct, or it may have learned a representation of the word that names it. AIPsy-Affect is designed to break that equivalence by removing explicit emotion vocabulary from both the affective and neutral sides of each matched pair (Keeman, 26 Apr 2026).

The battery is grounded in Plutchik’s eight primary emotions: rage, terror, grief, loathing, amazement, admiration, ecstasy, and vigilance. These are instantiated as third-person clinical-style narratives in which the target affect is evoked by situational structure rather than emotion-label words or close synonyms. The matched-control design preserves characters, objects, setting, approximate length, and sentence structure while removing stakes, violations, or irreversible consequences. Under that construction, any internal representation that distinguishes a clinical vignette from its neutral counterpart cannot be doing so on the basis of emotion-keyword presence (Keeman, 26 Apr 2026).

This design inherits the logic of an earlier 96-item keyword-free battery and expands it substantially. The expansion is not only numerical. It adds more domains, a moderate-intensity split, and a discriminant-validity split, thereby supporting tests of category sensitivity, intensity sensitivity, and resistance to narrative-rich but non-affective content within one framework (Keeman, 26 Apr 2026).

2. Corpus design and internal structure

The dataset is factorially organized across six domains: financial betrayal, professional violation, medical or institutional failure, caregiving responsibility, inheritance and intergenerational legacy, and public witness. Each domain is constructed so that multiple Plutchik emotions can plausibly arise from closely related situational scaffolds. This makes domain and emotion partially separable experimental factors rather than collapsed narrative themes (Keeman, 26 Apr 2026).

Split	Count	Function
Clinical	192	Keyword-free peak-intensity affective vignettes
Neutral	192	Matched neutral controls linked by `matched_control_id`
Moderate	48	Lower-intensity affective variants in domains 1–3
Complex_neutral	48	Vivid but emotion-free narratives for discriminant validity

The clinical split comprises 8 emotions × 6 domains × 4 vignettes per emotion–domain cell. The neutral split mirrors those 192 items through deterministic matched transformations. The moderate split contains 8 emotions × 3 domains × 2 items, allowing peak-versus-moderate comparisons without introducing explicit lexical cues. The complex_neutral split is not merely neutral in valence; it consists of descriptively dense technical or process-focused narratives such as machining or navigation scenes, included so that a detector responding to generic narrative richness can be separated from one responding to affective meaning (Keeman, 26 Apr 2026).

Authoring was constrained by a three-layer filtering procedure. First, vignettes were written under explicit no-emotion-label constraints. Second, texts were audited against an emotion-vocabulary lexicon assembled from the NRC Emotion Lexicon and Plutchik-specific keyword lists. Third, VADER sentiment sweeps were used to surface high-valence or high-salience lexical items that might function as covert affect labels. Flagged items were rewritten until affect arose from situational semantics alone (Keeman, 26 Apr 2026).

3. Validation and the “whether, not which” result

AIPsy-Affect was validated with a three-method NLP defense battery: VADER sentiment analysis, the NRC Emotion Lexicon, and the contextual classifier j-hartmann/emotion-english-distilroberta-base. The validation question was twofold: whether the items still carry detectable affective signal, and whether that signal remains category-identifiable in the absence of emotion keywords (Keeman, 26 Apr 2026).

VADER distinguished clinical from neutral items, with an unpaired Mann–Whitney result of $U = 11{,}500.5$ , $p = 1.76 \times 10^{-10}$ , and a paired Wilcoxon result of $W = 2{,}602$ , $p = 6.66 \times 10^{-15}$ . However, lexicon inspection showed that the high-valence words driving those separations were situational rather than emotional, including terms such as “block,” “death,” “died,” and “approved.” The per-emotion VADER pattern was also misaligned with the intended emotion structure, with some positive-target categories receiving negative sentiment medians and some negative-target categories receiving positive medians. The authors interpret this as evidence that bag-of-words sentiment is seeing situational vocabulary, not emotion labels (Keeman, 26 Apr 2026).

The NRC Emotion Lexicon provided a stricter keyword audit. In the clinical and neutral splits, no emotion-label words from the assembled NRC/Plutchik lexicon appear in the intended sense, and the observed emotion-category hit rates are low and weakly discriminative. Most Plutchik categories show no significant separation across the target classes, indicating that the keyword-free constraint holds at the lexicon level as well (Keeman, 26 Apr 2026).

The strongest result comes from the contextual transformer. GoEmotions DistilRoBERTa clearly detected that the clinical items were affective: the model assigned higher emotional probability to the clinical split than to the neutral split, with $p < 10^{-15}$ . Yet it failed at category recovery. On the keyword-free clinical items, top-1 emotion classification accuracy was 5.2%, below a seven-class chance baseline of 14.3%, whereas on a keyword-rich control set the same model reached 82.5% top-1 accuracy. This yielded the central validation conclusion: contextual models can detect that affect is present, but without emotion keywords they are poor at identifying which category is present (Keeman, 26 Apr 2026).

4. Interpretability uses and analytic protocols

The matched-pair structure is explicitly intended for linear probing, activation patching, sparse autoencoder feature analysis, causal ablation, and steering vector extraction. In each case, the key advantage is not simply that the items are emotionally evocative, but that the difference between affective and neutral variants is constrained to situational semantics (Keeman, 26 Apr 2026).

For linear probing, the basic contrast is affective versus matched neutral activations at a given layer. A probe trained on this contrast is insulated from trivial keyword routing. In the predecessor battery, binary affect detection on keyword-free items reached AUROC $= 1.000$ in early layers across multiple models, while eight-class classification degraded relative to keyword-rich controls. AIPsy-Affect scales that design up so that the same distinction can be examined with more statistical power, more domains, and explicit intensity manipulations (Keeman, 26 Apr 2026).

Activation patching is supported because each clinical item has a nearly isomorphic neutral control. One can run the model on the neutral version, replace selected activations with those from the clinical counterpart, and ask whether a downstream affect-sensitive readout changes. Since the paired items share surface structure, such interventions can be interpreted as causal transport of affective situational information rather than token-level lexical substitution (Keeman, 26 Apr 2026).

The battery also supports paired steering vectors. If $f_l(x)$ denotes a model representation at layer $l$ , then a paired affect vector can be defined as

$v_{\text{affect}} = \frac{1}{K} \sum_i \left( f_l(x_i^{\text{clin}}) - f_l(x_i^{\text{neut}}) \right).$

Because each difference term subtracts a matched neutral baseline, the resulting vector is intended to isolate affective semantics more cleanly than vectors derived from texts that explicitly name emotions. Moderate items then permit tests of whether the same representation scales with intensity or behaves categorically, while complex_neutral items test whether the putative affect feature is actually responding to descriptive richness (Keeman, 26 Apr 2026).

5. Position within the broader AIPsy-Affect research program

In adjacent work, “AIPsy-Affect” functions as a broader label for psychologically grounded, affect-aware AI systems. Within language generation, Affect-LM conditions an LSTM LLM on LIWC-derived affect categories and an affect-strength parameter $\beta$ , allowing generation to move from neutral to increasingly emotional text while also learning affect-discriminative word representations (Ghosh et al., 2017). A related strand adds affect directly to Seq2Seq dialogue through VAD-enriched embeddings, affect-biased attention, and an affect-weighted cross-entropy objective, yielding responses that human raters judged as more affect-rich without sacrificing content quality (Zhong et al., 2018).

A second line grounds affect in socio-mathematical control theory. BayesAct generalizes Affect Control Theory into a POMDP with continuous Evaluation–Potency–Activity sentiment variables, deflection minimization, and particle-filtered belief updates over identities, transients, and task state (Hoey et al., 2013). The same ACT framework has also been connected to neural dialogue generation by translating sentences into EPA space, using ACT to compute affectively aligned response targets, and then generating text from those targets with Seq2Seq or CVAE decoders (Asghar et al., 2020).

A third line concerns sensing and intervention. AffectToolbox packages face, pose, voice, and sentiment models into a GUI-driven multimodal PAD analysis system intended for real-time use by non-programmers (Mertes et al., 2024). FacePsy moves affective sensing onto smartphones, extracting action units, eye-open states, head gestures, and geometric facial features on device and reaching AUROC $p = 1.76 \times 10^{-10}$ 0 for two-week depressive episode detection, with a PHQ-9 regression MAE of 3.08 in naturalistic settings (Islam et al., 2024). On the intervention side, visually grounded reappraisal uses Whisper, SDXL, and a fine-tuned IP-Adapter so that spoken reappraisals of aversive IAPS images are rendered into supportive visualizations; in a within-subject study with $p = 1.76 \times 10^{-10}$ 1, AI-assisted reappraisal reduced negative affect more than standard reappraisal and description-based controls (Pinzuti et al., 14 Jul 2025).

Reinforcement-learning work extends the same program into decision-making and agent behavior. Go-Blend reframes affect modeling as a reward component, using a composite objective $p = 1.76 \times 10^{-10}$ 2 to blend human-arousal imitation with task performance in an arcade environment (Barthet et al., 2021). Affectively similarly embeds human-sourced arousal into OpenAI Gym–style environments as both an observation-side signal and an explicit reward term $p = 1.76 \times 10^{-10}$ 3, enabling experiments on human-like affect-based agents across three games (Barthet et al., 2024). Taken together, these works show that the broader AIPsy-Affect agenda spans representation, generation, sensing, causal control, and intervention, while the AIPsy-Affect battery supplies a particularly stringent benchmark for testing whether emotional structure inside LLMs is conceptual rather than lexical.

6. Limitations, open problems, and significance

AIPsy-Affect is deliberately narrow in several respects. It is English-only, situated in a Western clinical narrative frame, limited to Plutchik’s eight primary emotions, authored by a single clinical psychologist, and restricted to third-person narration. The moderate split covers only domains 1–3, and no matched neutral controls are provided for the moderate items. The release therefore offers strong control over one class of confound while leaving cross-linguistic, cross-cultural, taxonomic, and narratological generalization to future work (Keeman, 26 Apr 2026).

These limitations parallel broader difficulties in affective AI. Intercultural affect recognition studies have shown that feature robustness across cultures cannot be assumed, even if carefully selected visual cues can transfer surprisingly well (Mathur et al., 2022). Vision-language benchmarks indicate that current models align reasonably with human ratings for broad affective trends but remain weak on arousal and prone to overestimating response strength; anger and surprise are consistently difficult (Nowicki et al., 27 Jan 2026). Clinical affect sensing systems, including smartphone and wearable deployments, still rely on modest samples and face ecological-validity, personalization, and privacy constraints even when they achieve promising performance (Islam et al., 2024).

The significance of AIPsy-Affect lies in the specific kind of interpretive claim it enables. It does not prove that LLMs possess human-like emotions, nor does it supply a complete theory of affective representation. Rather, it removes a pervasive methodological shortcut. If a probe, SAE feature, ablation target, or steering vector succeeds on AIPsy-Affect, the result is no longer trivially attributable to the presence of words such as “furious,” “afraid,” or “joyful.” That makes the battery a high-value instrument for mechanistic interpretability under controlled affective semantics (Keeman, 26 Apr 2026).

The resource is openly released under the MIT license on Hugging Face, with DOI 10.57967/hf/8215, making it directly usable for probe construction, causal tracing, feature discovery, and intervention studies in emotion-focused language-model interpretability (Keeman, 26 Apr 2026).