Some colleagues and I recently published a paper updating our progress on automating the coding of implicit motives. If you follow my blog and research, I’ve been working with psychologists on this since 2015, when we started using machine learning and NLP. Back then we were using LSTMs and CNNs, and published our first results in 2020. Our work has continued, using increasingly complex language models, and this current paper (open access, available here) details our validation of the new coding model, as well as comparisons with two other models developed by other labs. We’re planning to present aspects of our work at the SSM 2026 meeting in Barcelona, so maybe we’ll see you there!
Why study implicit motives?
Implicit motives are a well-validated personality measure that has been used since the 1960s to understand individual and group differences that drive behavior. They reflect universal human needs to feel competent (Achievement), enjoy positive relations (Affiliation), and have influence (Power). As we mention in the paper, implicit motives can give insight into conflict resolution, leadership and entreprenurial success, well-being, and even predict societal impacts such as economic growth: see the paper for research related to these topics.
The major block to adoption of implcit motives in research has been the time-consuming nature of coding for these constructs. Early attempts at automation used dictionary-based marker-word methods, but did not achieve high correspondence with human-coded scores. More recent attempts use machine-learning on databases of texts to enable a computer to learn the mapping between linguistic cues and implicit motive codes. Our paper in 2020 showed that there was more promise in these approaches, but they didn’t bear fruit at the time.
Development of the classification model for implicit motives
Our new classification model for implicit motives achieves much higher correspondence with human coders, which I announced on our website implicitmotives.com back in 2023. However, at the time we had not yet assessed our model sufficiently on validation datasets, and since then we have also increased the size of training dataset. The new model was developed using a transformer-based neural network architecture called ELECTRA, which is relatively easy to train while still providing a robust representation of language. For the most recent iteration we report on in the paper we finetuned ELECTRA on a dataset of over 350k English sentences coded for implicit motives, primarily sourced from individual implicit motive researchers who used the Picture Story Exercise (PSE).
Validation and comparison of the model
We tested our model on several unseen datasets to verify its generalizability and establish its validity, specifically convergent, divergent, causal, and criterion validity. We found that our model demonstrated excellent correspondence with human-coded scores, showing high inter-rater reliability, and that it also successfully replicated well-known findings on theoretically relevant outcomes and established group differences.
Our model outperformed or performed similarly, in terms of human-machine convergence, to other transformer-based models that have recently been developed by other labs, and which differ somewhat from ours in terms of training dataset diversity, model architecture, and training approach. For instance, the models developed by Nilsson et al. (2025) and Brede et al. (2025) used smaller datasets and different model architectures, which highlights how differing approaches will result in somewhat different outcomes. For example, generalizability might be affected by a dataset biased toward a particular domain, and it may require direct observation of classification mismatches to understand why a model is making specific incorrect categorizations (which we discuss at greater length in the paper).
Areas Where the Current Model Doesn’t Match Human Coders
While our model generally performed well, there were some instances where it did not match human coders. In general, our model struggled with classifying implicit motives in idiomatic, metaphorical, and less explicit semantic contexts. In these cases, the model was either more conservative or less accurate in its classifications. This may be in part due to the model’s lack of real-world grounding. While humans are able to draw inferences to related contexts based on their world knowledge, a language model has only its pre-trained and local context to work with.
Because of this we recommend that researchers continue to manually check any machine scores of implicit motives when testing implicit motive classification models, particularly on genres which are not part of a model’s training data. However, we are optimistic that given the current success on various benchmarks, our model can be used for rapid assessment of implicit motives in PSE and related kinds of text. Check out the API here!
Future work
There is still work to be done. In our continuing research we will be focusing on improving the model’s ability to capture more complex and nuanced linguistic cues corresponding to implicit motives by expanding the training dataset to include this kind of data. For now, this paper represents a significant step forward in the automation of implicit motives. We are also encouraged by the interest from other labs and consider these various efforts to be contributing to the development of a new field of “motivational computing”, where computers can be used to detect and study motives in real time. We hope that our research will inspire further advancements in this exciting and promising field of study.