
Theory-based content analysis is central to social science research, but coding context-dependent interactional discourse remains difficult to scale. This study examines whether fine-tuned large language models (LLMs) can support theory-aligned, episode-level coding of complex social interaction, using socially shared metacognitive regulation (SSMR) in collaborative learning discourse as a demanding empirical case. Using a human-coded dataset of 2,372 discourse episodes from undergraduate collaborative problem-solving, we adapted the large language model content analysis (LACA) framework through an AI-aligned codebook, prompt design, and supervised fine-tuning. Prompt-only and fine-tuned models were compared across episode-level classification and multilayer SSMR coding. Fine-tuned models substantially improved agreement with human coding and predictive performance relative to prompt-only baselines. Performance was strongest for episode-level classification and several higher-support coding layers, whereas errors remained concentrated in sparse, boundary-prone, and highly interpretive categories requiring social-intent reconstruction and temporally unfolding context. The findings extend recent work on LLM-assisted qualitative and content analysis by showing both the promise and limits of supervised, theory-aligned automation for multilayer social interaction coding. Overall, the study supports a human-centered approach to LLM-assisted content analysis, in which automated coding can support more stable coding tasks while human expertise remains essential for ambiguous, boundary-prone, and interpretively demanding judgments.
As political toxicity has become widespread on social media, scholars have sought to understand what drives it. While studies often emphasize platform-level explanations, such as algorithmic amplification, they offer limited insight into which users are more likely to produce toxic content. Even less is known about how the relationship between individual-level factors and toxicity varies across political contexts. Using data from 184,354 randomly sampled U.S. Twitter (now X) users, their original tweets (25M), and the accounts they follow, we link users’ inferred characteristics and political context to their real-life toxic behavior. Results show that Republican, male, and older users are consistently more likely to post both impolite and intolerant content, whereas education shows no clear association. We also find that state-level electoral competitiveness increases overall toxicity and, in some cases, conditions individual-level relationships. The age gap in impolite behavior significantly narrows in more competitive states, while the corresponding moderation for intolerance is only marginally significant. There is also some indication that partisan differences in impoliteness become larger in more competitive states, but this moderation is only marginally significant. Together, these results demonstrate that political toxicity is shaped jointly by users’ individual characteristics and the political contexts in which they operate.
The present study seeks to understand how criminal hackers self-report to desist from the world of cybercrime. In doing so, we address significant gaps in criminological literature, as little is known about pathways out of cybercrime offending. Interviews were conducted with 22 individuals involved in cyber-dependent crime, primarily from the Netherlands. Of these, 15 respondents were justice-involved youth and 7 respondents self-identified as criminal hackers. The interviews offer unique insights into the stages offenders navigate following their involvement in cybercrime. Findings reveal that young hackers may be susceptible to behavioral change through various mechanisms. Desistance from cybercrime is a process that warrants a multidisciplinary perspective, influenced particularly by social bonds, prosocial daily routines, and deterrent experiences. A combination of police intervention and alternative approaches that emphasize ethical application of skills may be a promising route for young hackers in particular, although actual intervention effects remain unverified. We encourage evaluations of current programs for hackers as well as additional quantitative assessments of protective factors among cybercrime offenders.
Taking photos to manage memories with smartphones is a routine strategy in the digital era. However, it remains unclear how smartphone photo-taking affects memory performance and self-evaluation. This research examined whether smartphone photo-taking is associated with a pattern consistent with the “Google effect,” and whether these associations vary with self-expansion via smartphone use. Across three experiments ( N = 363), participants photographed or observed materials before completing memory tests and measures of cognitive self-esteem. Study 1 tested the basic photo-taking effect; Study 2a examined device ownership as a contextual proxy for smartphone self-expansion; and Study 2b examined trait self-expansion as a continuous moderator. Results showed a photo-taking-impairment effect on memory, alongside higher cognitive self-esteem following photo-taking. Moreover, higher smartphone self-expansion was associated with stronger increases in cognitive self-esteem and greater memory overestimation, whereas its role in objective memory impairment remained inconclusive. Altogether, the findings suggest that smartphone photo-taking can reduce memory performance while increasing cognitive self-esteem.
Framing research has long relied on human annotation, but many aspects of framing – especially framing functions such as defining problems, identifying causes, making moral evaluations, or proposing solutions are difficult to label consistently and at scale. We demonstrate that large language models can reliably identify these framing functions, achieving levels of agreement comparable to those of human annotators. We evaluate three state-of-the-art models (GPT-5, Claude 3.7, and DeepSeek) alongside three trained human coders on a new dataset of social media posts about electric-vehicles, a highly topical and contested issue that holds many interpretations. This setting provides a strong test of both agreement and meaningful disagreement in framing-function annotation. Across models and humans, we find not only substantial convergence but also patterns of valid disagreement – differences that reflect legitimate alternative interpretations rather than annotation error. Recognising and preserving these cases enhances construct validity and provides a more realistic portrayal of how framing functions operate in contested debates. Our results demonstrate that large language models can scale framing function analysis while capturing the variability inherent in human interpretation, enabling forms of communication research that were previously impractical with human-only coding.