Are ChatGPT-Generated Headlines Better Attention Grabbers than Human-Authored Ones? An Assessment of Salient Features Driving Engagement with Online Media

Abstract

This study focuses on the case of news headlines in current online journalism, looking into the current possibilities opened by ChatGPT to generate such texts in an attention-grabbing manner. To assess the reaction of online readers to headlines (clickbait or click-worthy), an online survey was applied, involving Romanian students. A total of 100 original human-authored articles with clickbait headlines were extracted from a relevant Romanian database. ChatGPT was used to generate alternative headlines (one clickbait and one informative) based on the original texts. The resulting corpus of 100 headline triplets was offered to students for evaluation. More than 70% of the 600 participants in the survey preferred AI-generated headlines over the human-authored ones, indicating their experiences and behaviors in media consumption. The preferred headlines were further analyzed along lexical and grammatical characteristics, and stylistically, to pinpoint the features sparking readers’ curiosity and engagement. While on a cognitive level the investigated audience rejected clickbait headlines as being deceitful and frustrating, in practice less than 34% favored neutral and objective headlines. Also, the linguistic analysis provided insights into the mechanics of reader engagement and the effectiveness of various headline strategies. The results are useful to anticipate the adoption of AI as a creative partner in Romanian media practice.

1. Introduction

The highly dense technological environment of modern life in the 21st century imprints all activities and dramatically changes all sectors of society. As Marshall McLuhan so aptly foretold, the technological extensions of man expand the possibilities of action beyond imagination (McLuhan 1964), although not without cost. The debate over such extensions is relaunched with the multiplication of technologies capable of imitating human intellect, summed up under the generous term ‘artificial intelligence’ (Prasad and Makesh 2024). The media industry, by its very nature, could not and did not make an exception in making use of the tremendous range of applications that can optimize the collection, creation, editing, distribution, and promotion of media content (Chan-Olmsted 2019). As Sylvia Chan-Olmsted states, in her comprehensive overview of AI use in the journalism and entertainment industries, digital audiences have enhanced expectations regarding the tailoring of news, the speed of delivery of content, capable services providing on-demand ‘smart’ products, etc. (Chan-Olmsted 2019). AI changes the journalistic workflow, the delivery and the promotion of media content, the pricing structure, and the marketing strategies (Chan-Olmsted 2019; Ioscote et al. 2024; Prasad and Makesh 2024). As Fabia Ioscote et al. state, the advent of generative AI took steps further than digital or automated journalism, forcing the news industry to reflect upon the possibility of maintaining competitiveness and quality. At the same time, the journalistic profession has to face the challenges brought by the increasingly more intrusive use of AI in the journalistic workflow, where routine activities can be optimized, but human touch still must be maintained to convey trustworthy, quality, ethically attested media content (Ioscote et al. 2024; Amponsah and Atianashie 2024). Major prestigious news agencies such as Associated Press and Reuters already use integrated AI as tools to extract well-documented and timely stories from large data sets (Amponsah and Atianashie 2024). However, despite some examples of successful use of AI in journalism practices and praised advantages such as modernity, speed, and the capability to handle enormous amounts of data, the literature cautions that it is too soon to give up human control over journalistic practices. Ethical and professional values are to be instilled into automated tools before spreading AI use to a larger extent (Amponsah and Atianashie 2024; Aydın and İnce 2024; Beckett and Mira n.d.; Dragomir et al. 2024; Graefe and Bohlken 2020; Helberger et al. 2020; Ioscote et al. 2024). Limitations acknowledged in the larger-scale use of AI in journalism are, besides ethical considerations, the danger of spreading misinformation, the peril of creating echo chambers, and the threat of reducing exposure to a diversity of viewpoints for the media consumers allowed to tailor the received media content to their taste and preference (Breazu and Katsos 2024; Chan-Olmsted 2019; Siitonen et al. 2024; Amponsah and Atianashie 2024; Prasad and Makesh 2024). As Siitonen and colleagues concluded, from analyzing ten years of studies dedicated to the influence of AI on journalism and journalistic practices, attitudes towards AI oscillate “between hype and fear”. However, new technologies are here to stay and it is important to inquire into how this development changes journalism and media in the long run (Siitonen et al. 2024; Sonni et al. 2024). Studies dedicated to scrutinizing the actual use of AI in newsrooms show that, at least in Central Europe, the spread of these newly developed tools is not so large as debated in the USA or China, for example (Dragomir et al. 2024; Siitonen et al. 2024; Zheng et al. 2018). For Central Europe, a region that Romania is neighboring, studies indicate that the average percentage of AI usage in media outlets does not exceed 15% (Dragomir et al. 2024). The report mapping the use of AI by newsrooms in the Czech Republic, Hungary, Poland, and Slovakia revealed that AI is primarily used for simple tasks such as gathering information, writing basic news based on existing data, editing, and moderating texts in newsrooms, but that the adoption of AI is progressing slowly and cautiously.

Given the fact that the field of journalism is so vast and that the impact of automated tools is growing at a rapid pace, the research team authoring the present study narrows down the topic of AI use, with pros and cons, to a very specific activity in the media industry: the creation of headlines, capable of attracting readers to engage with the content. The interest in headlines is manifold. In the digital age, the way in which audiences consume news is changing. Increasingly larger number of people access media content online, instead of the traditional channels (Kuiken et al. 2017). Such consumption allows for the abandonment of linear reading, listening or watching, since audiences pick and choose content according to the attractiveness of the promise or the content formulated in the headline. In traditional media, the headline functioned as a summary or informative text regarding the content of an article or news piece, meant to inform and persuade. Studies show that headlines are also the most read part of the media, receiving around five times more attention than the content they refer to (Isani 2011). Headlines also served as an attention grabber, typically requiring the skills of the best writers in the newsroom to produce such texts (Isani 2011; Ifantidou 2009; Scacco and Muddiman 2015; Iarovici and Amel 1989). The digital environment created an additional function for the headlines: that of luring the audiences towards the news and triggering engagement (Feng 2024). Experiments in reformulating headlines showed that headlines carry framing functions, tipping audiences to adopt a position and an interpretative standing prior to accessing the main media content (León 1997; Leung and Strumpf 2023). As Van Dalen concludes, from the analysis of journalists’ re-examination of their professional skills, under the pressure of technological innovation and machine-written texts, “what can be automated will be automated” and headline production is a forerunner in the quest for journalistic productivity (van Dalen 2012).

Among the newest topics of debate concerning the functions of headlines is the clickbait phenomenon. This function refers to the capacity of clickbait headlines to keep the reader in the webpage for as long as possible, but not to inform (Potthast et al. 2016). Perceived as deceitful and capable of propagating misleading or even false information, clickbait is actively combated both via automated tools, and by discursive strategies aimed at persuading the media to drop the practice (Blom and Hansen 2015; Chakraborty et al. 2016; Broscoteanu and Ionescu 2023; Blom and Hansen 2015; Chakraborty et al. 2016; Broscoteanu and Ionescu 2023). Despite the warnings that clickbait erodes the credibility of the media and can alienate audiences, many news outlets still use clickbait headlines. To appeal to audiences, even traditionally ‘serious’ media outlets resort to the tabloidization of titles and playing upon the curiosity gap of online readers (Ifantidou 2009; Setälä 2014), underpinning the viralization of news (Bazaco et al. 2019). The literature on clickbait headlines focuses both on the linguistic strategies and vocabulary used to entice readers’ engagement (Blom and Hansen 2015; Fărcașiu and Gherheș 2024), and on the effects and typology of such texts (Molyneux and Coddington 2020).

However, while research of English-language material is abundant in the scientific world, other, so-called “low resource” languages are covered by less than a quarter of the research (Fatima et al. 2023). It is also interesting to note that research on the contrastive man- vs. machine-produced media content indicate a skeptical reception of the latter on the part of the public (Altay and Gilardi 2024; Berber Sardinha 2024).

For the Romanian media landscape, studies on headlines (clickbait or click-worthy), journalistic practices involving AI, and reflections on the specificity of journalism in the very dense technological and digital environment are scarce. A notable exception is represented by Broscoteanu and Ionescu (Broscoteanu and Ionescu 2023) who created a vast corpus of clickbait headlines from the most popular media outlets, Păcurar and Oprişa, interested in automated affordances to fight clickbait in Romanian articles (Păcurar and Oprişa 2023), and some ethical considerations regarding AI in journalism (Stănescu 2023). To catch up with the state of research on headline status, as described by Scacco and Muddiman (2015), and enrich knowledge on non-English practices, the research team authoring the present study embarked on a scientific journey to understand human reactions to headlines, human-authored or AI-generated, to deepen the knowledge on the features of effective headlines and to reflect upon the future of journalistic practices in Romanian newsrooms.

The objectives of this study were the following:

O1. To analyze the media consumption behavior of online readers.

O2. To assess the features of the headlines leading to the decision of audiences to engage with the content of online media, highlighting the possible effect of such texts.

O3. To compare ChatGPT-generated headlines with human-authored ones, from the point of view of their function as attention grabbers for online audiences, in an educational environment.

O4. To identify the linguistic and stylistic characteristics of ChatGPT-generated headlines that effectively pique readers’ curiosity and motivate them to read the articles.

2. Materials and Methods

This research aimed to investigate readers’ consumption behavior concerning online articles, to assess how headlines influence their reading decisions, and to examine the reactions of readers to misleading headlines. In addition, it sought to compare the impact of human-authored clickbait headlines with that of clickbait and informative headlines generated by the ChatGPT model to assess their effectiveness in capturing the readers’ attention. To address the initial objectives, the questionnaire included four closed-ended, scale-type questions. For the final objective, the selection of the articles was based on the findings of Broscoteanu and Ionescu, who conducted research on detecting clickbait headlines in Romanian and developed the first Romanian corpus for this topic, called RoCliCo (Broscoteanu and Ionescu 2023).

Corpus

To analyze the impact of headlines (O4), 100 articles were extracted from RoCliCo. This database contains 8.313 headlines selected from six media sources. Almost half of the headlines are of a clickbait nature. For the purpose of the present research, only five media outlets were considered, since one of the sources present in RoCliCo did not have a format allowing for statistical analysis. The five Romanian online media sources include the Digi24 TV station, the Libertatea newspaper, the ProTV TV station, the WOWBiz newspaper, and the Viva magazine. To maintain a balance in the corpus, 20 articles with clickbait headlines were chosen from each source.

The selection of the 20 headlines per media outlet followed a systematic and objective approach. First, only articles identified as having clickbait headlines were included in the selection pool. For each media source, the total number of clickbait articles was divided by 20, establishing a fixed selection interval. Headlines were then chosen systematically at these intervals, ensuring an even distribution across the dataset. This method provided a representative and unbiased sample of 20 clickbait headlines per outlet, balancing the contribution of each source to the overall corpus.

These articles were used to compare the original human-authored clickbait headlines with their ChatGPT-generated versions. For each of the 100 articles, the ChatGPT language model has produced two alternative headline versions:

• A clickbait-style headline generated by ChatGPT: Based on the existing news article, the model was instructed to generate a clickbait headline that would grab attention and entice the reader to click on the article.
• An informative headline generated by ChatGPT: Based on the existing news article, the model was instructed to generate a headline in a clear and objective manner, accurately representing the article’s content and avoiding speculative or sensationalist language.

It is essential to emphasize the fact that the two variant headlines produced by ChatGPT were based exclusively on the articles’ content, without the model accessing the original headlines. This method ensured that the alternative headlines were independently generated and not influenced by the original phrasing, allowing for an objective evaluation of how different types of headlines affect reader perception. Consequently, each article was accompanied by triplet headlines: the original, human-authored headline, a clickbait ChatGPT-generated headline, and an informative ChatGPT-generated headline. Thus, a corpus of 100 triplet headlines was created, allowing for the next stage of research.

Questionnaire

The measures in the questionnaire were specifically developed for this study and did not come from previous research. They were designed to align with the study’s objectives of understanding reading habits, perceptions of headlines, and preferences for different types of headlines.

Ten distinct variants of the questionnaire were created. The first part was identical and included questions about reading habits, perceptions of headlines, and features of headlines appealing to them on a cognitive level. The second part of the questionnaire featured 10 sets of headline triplets, which included the original (human-authored) headline, the clickbait-style headline generated by ChatGPT, and the informative headline generated by the model. Each participant rated one set of 10 headlines, which were randomly distributed to minimize order effects and ensure a balanced representation of headline types among participants. In total, participants evaluated 100 sets of headlines.

A separate, smaller group of 10 university students participated in the pretest of the questionnaire. This pretest group was not part of the main sample of 624 participants. The purpose of the pretest was to verify the clarity of the questions and the validity of the research instrument before full implementation. Feedback from these students was instrumental in refining the questionnaire to ensure comprehensibility and alignment with the study’s goals.

To assess the internal consistency of the instrument used, Cronbach’s Alpha coefficient was calculated (Howitt and Cramer 2008; Tabachnick and Fidell 2012). The analysis indicated a value of 0.786, which suggests a good level of internal reliability as per the general guidelines (George and Mallery 2003) (

Table 1. Reliability statistics.

Cronbach’s Alpha	N of Items
0.786	14

). The 14 items are correlated with each other and contribute to measuring the same variable.

The participants completed the questionnaires via an online specialized platform. Invitations were sent out by the research team and responses were collected in the timeframe 25 September–20 October 2024.

Sample

The participants in this study were exclusively university students, enrolled in higher education institutions in Timișoara, the largest city in the Western part of Romania. The sample of 624 respondents was carefully selected to represent young consumers of online media, a group highly relevant to the research objectives. No members of the general community were included in the study. For this sample of respondents, aged 18–24, the margin of error was estimated at ±4%. Participation was voluntary, with individuals being informed in advance about the study’s purpose and assured of the confidentiality of their responses. No incentives were provided and participants were free to withdraw from the questionnaire at any time. All data collected were anonymized and used solely for research purposes.

The collected data were analyzed using statistical methods and quantitative analysis tools to assess the impact of headlines on the students’ reading behavior.

3. Results

The first aspect that was analyzed was the frequency with which the study participants read online articles. As shown in the graph below (Figure 1), reading habits among the respondents vary, which may influence how they are impacted by article headlines. The largest group, 31.5% of the students, read online articles several times a week, followed by 27.5%, who read them daily. These two groups are the most active in consuming online content. Meanwhile, 20.2% of the participants rarely read online articles, indicating a lower exposure to online content. Additionally, 14.2% of them read articles weekly, reflecting a moderate level of content consumption, while 6.6% read articles monthly, indicating minimal exposure to online content.

Another aspect examined was the way in which participants perceive the influence of headlines on their decision to read an article. As illustrated in the graph below (Figure 2), the largest portion of subjects (43.9%) believes that the appeal of a headline significantly impacts their decision to click on an article, suggesting that, for some readers, the headline is a key factor in their choice to further explore the content. This category also includes 40.2% of the respondents who state that headlines have a strong influence on their decision, further emphasizing the fact that headlines play a crucial role in attracting readers. Thus, more than 80% of the students report that headlines have a significant impact on their reading behavior, emphasizing the crucial role of effective headline writing in an article’s success. In contrast, 11.3% of the participants are neutral about the influence of headlines, indicating a group for whom headlines play a smaller role in their decision to read. At the other end of the spectrum, 3.7% of the respondents believe that headlines have little influence, while 1.0% state that headlines have no impact at all on their decision to read an article.

To better understand the features that readers consider important in a headline, a group of students was invited to participate in this study. They were asked to identify the aspects they find crucial when deciding whether to read an article or not. The collected responses were analyzed and summarized, while the similar characteristics were grouped together to create a list of key factors in headline assessment. As a result, 14 categories were identified, reflecting the preferences and priorities of students regarding the main attributes of an appealing headline. The participants were then asked to rate the importance of these headline characteristics. The analysis of their responses offers a detailed view of the elements that readers prioritize when deciding to click to read an article, highlighting the factors that most influence their decision to read online content. The results are summarized in

Table 2. Assessment of the significance of various headline features for readers.

	To a Very Small Extent	To a Small Extent	To Some Extent	To a Large Extent	To a Very Large Extent
To be clear and accurately reflect the content	0.8	2.1	12.0	30.2	54.9
To contain important information	1.4	2.6	20.2	34.9	40.9
To be creative and engaging	1.1	6.2	25.4	36.3	30.9
To pique interest and stimulate curiosity	0.8	4.8	19.7	42.4	32.3
To include keywords relevant to the topic	1.9	6.2	29.1	35.8	26.9
To be short and concise	1.9	6.6	38.0	36.5	17.0
To provide a summary of the content	3.4	13.4	34.9	28.0	20.3
To be shocking or surprising	5.9	17.0	38.6	22.9	15.5
To promise new or exclusive information	5.3	14.3	36.5	26.8	17.1
To have an optimistic or positive tone	4.5	10.6	45.7	26.7	12.5
To be neutral and objective	6.4	22.1	38.7	19.4	13.4
To evoke strong emotions	7.6	20.1	44.2	17.5	10.6
To suggest an unusual situation	8.5	26.1	42.5	15.1	7.9
To contain figures or statistics	12.0	26.1	39.7	14.9	7.2

below.

The first aspect that emerges from the collected data is the clarity of the headline and its ability to accurately represent the content. Approximately 54.9% of the students view this factor as very important, while 30.2% rate it as important, indicating that the transparency and relevance of the headline in relation to the content are essential criteria for most readers. Only a small percentage (2.9%) of the participants consider this aspect to be of little or very little importance. Additionally, headlines that convey important information are regarded as very important by 40.9% of the survey participants, while 34.9% rate it as important. Following this, creativity and the headline’s ability to be engaging rank third, receiving strong appreciation, with 30.9% of the respondents considering it very important and 36.3% agreeing that it is important.

Another valued aspect is the ability of headlines to pique interest and stimulate curiosity, with 32.3% of the interviewees deeming this feature very important and 42.4% considering it important to a large extent. Following this, the inclusion of relevant keywords ranks next on the list of preferences, being deemed very important by 26.9% of the respondents and important to a large extent by 35.8%. Most respondents favor short and concise headlines; 38.0% consider them somewhat important, 36.5% view them as very important, and 17.0% rate them as extremely important, indicating that readers generally prefer concise titles. Additionally, 20.3% of the respondents highly value headlines that provide a summary of the content, while 28.0% consider them very important. This indicates the fact that many readers appreciate headlines that give a clear preview of the article’s content, although this feature is not considered the most important one as compared to others. Shocking or surprising headlines are not seen as a priority. Only 15.5% of the respondents regard them as very important, while 22.9% consider them significantly important. However, a larger portion (38.6%) consider them as somewhat important, indicating that while such headlines may catch attention, they are not a priority for most readers. On the other hand, headlines that promise new or exclusive information are considered very important by 17.1% of the respondents, while 36.5% find them somewhat important. This suggests that, for a significant number of readers, novelty and exclusivity can be key factors in deciding whether to read an article or not. Headlines with an optimistic or positive tone are highly valued by 12.5% of the respondents and considered important by 26.7%, while 45.7% view them as somewhat important. This indicates that while a positive tone is appreciated, it is not necessarily a decisive factor in the decision to read an article.

The neutral and objective tone of the headline is appreciated by 19.4% of the participants in the survey, with 13.4% of them rating it very highly. Additionally, a significant portion of readers (38.7%) views this aspect as moderately important.

Headlines that evoke strong emotions are rated very highly by 10.6% of the respondents and highly by 17.5% of them. However, for the majority (44.2%), this aspect holds only moderate importance, suggesting that while emotional headlines can draw attention, they are not always a decisive factor in choosing to read.

Headlines suggesting an unusual situation are rated as very important by just 7.9% of the survey participants and important by 15.1%, reflecting a lower preference for this style. With 42.5% of the respondents finding this aspect moderately important, it appears that unusual headlines are generally not a primary factor for most readers when deciding to read an article.

The inclusion of figures or statistics in a headline is highly valued by only 7.2% of the respondents and moderately valued by 14.9% of them, with 39.7% considering it somewhat important. This suggests that while headlines featuring figures and statistics are useful to some readers, they are not a decisive factor for most.

To understand how often the study participants have encountered situations where headlines do not accurately reflect the content of the articles, the questionnaire has also included the question, “How often do you come across titles that do not reflect the content of the article?”. The chart below (Figure 3) illustrates the readers’ perceptions of this discrepancy between headlines and content, a common issue in online media.

The largest portion of the respondents (39%) reports frequently encountering headlines that do not accurately reflect the content of the article, indicating a common experience with such discrepancies among readers. Following this, 27.8% of the respondents state that they encounter such headlines occasionally, representing those who experience the issue often enough to notice, but not consistently. Additionally, 25.9% indicate that they very often come across misleading headlines, suggesting that over a quarter of readers are regularly exposed to such headlines, potentially undermining trust in information sources. A smaller group (6.1%) say they rarely encounter such headlines, while only 1% claim that they have never encountered headlines that misrepresent the article’s content.

The question “How frustrated would you feel if you read an article whose headline did not accurately reflect the content?” explores the level of frustration respondents feel towards misleading headlines, offering insight into their emotional responses when an article’s content fails to meet the expectations set by its headline. The results indicate that most respondents (44%) experience moderate frustration, suggesting a significant, though not overwhelming, reaction to discrepancies between the promise of the headline and the content of the actual news piece (Figure 4).

About 36.2% of the study participants report feeling very frustrated, showing that over a third of the sample experiences a high level of frustration in these situations. This suggests that misleading headlines can create a negative experience, potentially undermining trust in the information source. Meanwhile, 18.1% of the respondents maintain a neutral attitude, indicating that they are not significantly impacted by discrepancies between the headline and its content. Only 1.8% of the respondents are not bothered by such headlines, suggesting that a very small portion of the sample feels that misleading headlines do not affect their reading experience.

• Linguistic and stylistic characteristics of ChatGPT-generated clickbait headlines

This study also explored how various headline types influence the readers’ curiosity to engage with an article. The primary goal was to compare original headlines with both clickbait and informative versions generated by ChatGPT to gauge their effectiveness in drawing attention. The data collected (see Figure 5) offer insight into the readers’ preferences, showing that ChatGPT-generated clickbait headlines lead, with a top score of 37.5%. This suggests that clickbait headlines are the most successful at capturing attention and arousing curiosity. The informative headline by ChatGPT followed closely with a score of 33.4%, indicating that clear and factual headlines also appeal to readers, though somewhat less strongly than clickbait. The small margin between these scores suggests that many readers still value the transparency and accuracy found in informative headlines.

The article’s original headline received the lowest interest score, gathering only 29.2% of responses. This result indicates that the original headlines are less effective at capturing readers’ attention as compared to those generated by ChatGPT, whether clickbait or informative.

As clickbait headlines are intricately linked to curiosity through their wording, being crafted to tap into the readers’ natural desire to uncover answers and explore the unknown, it has become clear that studying clickbait headlines from a linguistic perspective is important because it allows for an understanding of the specific linguistic strategies that capture readers’ attention, shape perceptions, and influence reader engagement. Moreover, linguistic research on clickbait can reveal how language manipulates information, potentially impacting public perception and trust in the media. It also helps understand how digital communication has evolved with the influence of social media algorithms, which reward engagement, thereby encouraging language use that maximizes clicks.

Therefore, the selected headlines were further analyzed through a qualitative linguistic lens to uncover the features that spark the readers’ curiosity and emotional engagement. The analysis first focused on the lexical and grammatical characteristics, followed by the analysis of the stylistic elements, and continued by comparing the human-authored clickbait headlines and the ChatGPT-generated ones to uncover what makes certain headlines more compelling, providing insights into the mechanics of reader engagement and the effectiveness of various headline strategies.

The decision to analyze clickbait through lexis, grammar, and stylistics provides a nuanced framework for understanding the specific linguistic strategies that engage and influence readers. Each level of analysis sheds light on how language is deliberately structured to capture attention in the digital media landscape.

The effectiveness of clickbait relies heavily on its vocabulary choices, which evoke curiosity, emotion, and urgency. The foundational work of Biber et al. (Biber et al. 1999) on lexical patterns highlights the strategic selection of verbs, nouns, and adjectives to create dramatic effects. The psychological dimension of emotionally charged vocabulary is supported by Pennebaker et al. (Pennebaker et al. 2003), who emphasize that words with strong emotional valence increase engagement. In the context of clickbait, this aligns with the use of high-arousal words to stimulate curiosity and anticipation.

In terms of grammar, grammatical structures in clickbait headlines are meticulously designed to draw readers into the narrative. The analysis by Biber et al. (Biber et al. 1999) on tense and aspect could inform how the present tense, commonly used in clickbait, creates a sense of immediacy and relevance. The work of Beaugrande and Dressler (De Beaugrande and Dressler 1981) focuses on principles of cohesion and coherence in texts, which can be directly applied to understanding the use of ellipsis and interrogative structures in clickbait headlines. An ellipsis creates a gap that engages the reader’s cognitive processes, requiring them to “fill in the blanks” with information implied by the surrounding text. In the same line, Halliday and Matthiessen (Halliday and Matthiessen 2013) offer a comprehensive framework for analyzing language through its ideational, interpersonal, and textual metafunctions, which can serve as a foundation for studying clickbait. The ideational metafunction helps identify how dynamic verbs, noun phrases, and lexical choices in headlines convey curiosity and emotional engagement. The interpersonal metafunction examines how imperatives, direct address, or rhetorical questions establish a connection with readers. Finally, the textual metafunction reveals how features such as ellipsis or thematic structure enhance cohesion and intrigue, effectively capturing audience attention.

Lastly, stylistics enhances clickbait’s appeal by making it more vivid, memorable, and emotionally engaging. For instance, Lakoff and Johnson (Lakoff and Johnson 1980) offer a foundational understanding of how metaphor shapes perception, cognition, and communication. Their insights are particularly relevant to the study of clickbait headlines, which frequently use metaphorical language to evoke emotional responses, create intrigue, and frame information in a compelling way. Similarly, the stylistic principles of Tannen (Tannen 1989), particularly her analysis of involvement strategies in discourse, align with the use of direct address and pseudo-conversational tone in clickbait. Furthermore, Jeffries and McIntyre (Jeffries and McIntyre 2010) provide tools for analyzing how linguistic choices shape meaning and impact, making them highly relevant for studying clickbait. Their focus on the interaction between form and function in texts aligns with examining how stylistic elements like metaphors, lexical emphasis, and structural patterns create intrigue and emotional resonance in headlines. Their discussion of reader interpretation and narrative engagement helps understanding how clickbait strategically employs stylistic devices to capture attention, evoke curiosity, and manipulate reader expectations.

The analysis of the linguistic and stylistic features of the clickbait headlines is detailed in the following section.

• Lexical and grammatical level

At the lexical–grammatical level, verbs and verbal structures play an essential role in creating a sense of urgency and drama.

➢

Verbs and verbal structures

The dynamic verbs found in the clickbait headlines in the analyzed corpus (such as “amenință” (“threatens”), “explodează” (“are exploding”), and “dezvăluie” (“reveals”) imply swift, dramatic, or perilous actions, evoking a sense of intrigue. These verbs indicate that something significant is currently occurring, motivating the reader to seek further information.

E.g.,

“Putin amenință cu atacuri preventive! Rusia ar putea lansa prima lovitura nucleară, schimbând doctrina militară!” [“Putin threatens with preventive attacks! Russia could launch the first nuclear strike, changing the military doctrine!”]

“Extremiștii proruși din Bulgaria explodează în popularitate!” [“Bulgaria’s pro-Russian extremists are exploding in popularity!”]

“Olaf Scholz dezvăluie: Rusia a renunțat la planul nuclear!” [ Olaf Scholz reveals: Russia has given up its nuclear plan!”]

Modal verbs such as “ar putea lansa” (“could launch”) introduce uncertainty and possibility, encouraging speculation and enhancing curiosity. They imply a catastrophic scenario that captures attention and drives the reader to delve into potential details.

E.g.,

“Putin amenință cu atacuri preventive! Rusia ar putea lansa prima lovitura nucleară, schimbând doctrina militară!” [“Putin threatens with preventive attacks! Russia could launch the first nuclear strike, changing the military doctrine!”]

Likewise, the use of phrases such as “a trasa o linie roșie” (“to draw a red line”) or “a da în judecată” (“to take to court”) creates a framework of tension and mystery, drawing attention to seemingly suspenseful events. “Drawing a red line” or “taking to court” are associated with conflict and tense situations, which can stimulate interest in the development of these events.

E.g.,

“Olaf Scholz dezvăluie: Rusia a renunțat la planul nuclear! Află cum comunitatea internațională a trasat o linie roșie!” [“Olaf Scholz reveals: Russia has abandoned the nuclear plan! Find out how the international community has drawn a red line!”]

“Șoc în Elveția! Mii de femei senior dau în judecată guvernul din cauza schimbărilor climatice—Europa așteaptă o decizie istorică” [“Shock in Switzerland! Thousands of senior women take the government to court over climate change—Europe awaits historic decision”]

➢

Nouns

Nouns are strategically selected to heighten emotions. Words such as “bombă” (“bomb”) and “dezastru” (“disaster”) exaggerate the scenarios, making them appear more dramatic. These terms emphasize sensationalism and capture attention, as seen in the following headlines.

E.g.,

“Bombă la vama Siret! Vameșii români, complici într-o rețea internațională de fraudă cu tutun, dezvăluită de Kovesi!” [“Bomb at the Siret customs! Romanian customs officials, accomplices in an international tobacco fraud network, revealed by Kovesi!”]

“Dezastru în București! Zeci de copaci prăbușiți de furtună din cauza nepăsării autorităților!” [“Disaster in Bucharest! Dozens of trees collapsed by the storm due to the carelessness of the authorities!”]

Other nouns suggest mystery and revelations. Nouns such as “adevărul” (“the truth”), “secretul” (“the secret”), and “povestea nespusă” (“the untold story”) attract attention because they suggest information that is not known to the general public.

E.g.,

“Secretul din spatele numelor Kira și Ianis Hagi! Află povestea nespusă a copiilor lui Gică Hagi care i-a adus succesul!” [“The secret behind the names Kira and Ianis Hagi! Learn the untold story of the children of Gică Hagi who brought him success!”]

“Marian Godină dezvăluie ADEVĂRUL despre dosarul penal al Liei Bugnar: Ce nu știe publicul despre lege!” [“Marian Godină reveals the TRUTH about Lia Bugnar’s criminal case: What the public does not know about the law!”]

➢

Adjectives and adverbs

Adjectives and adverbs enhance the emotional weight of headlines. Words such as “șocantă” (“shocking”) or “incendiară” (“incendiary”) are employed to heighten the headline’s impact, implying that the information presented is highly surprising or even extraordinary.

E.g.,

“Gigi Becali, donație șocantă de 800.000 de euro pentru o biserică în Africa! Află de ce a ales Rwanda!” [“Gigi Becali, shocking donation of 800,000 euros for a church in Africa! Find out why he chose Rwanda!”]

“Bătălia incendiară de la Vocea României! Iulian Nunucă a lăsat juriul fără cuvinte, iar Bogdan Jeler surprinde cu o decizie șocantă!” [“The incendiary battle from the Voice of Romania! Iulian Nunucă has left the jury speechless, and Bogdan Jeler surprises with a shocking decision!”]

Superlatives in the corpus emphasize the uniqueness or intensity of certain elements, creating a heightened sense of drama. They are instrumental in capturing readers’ attention by implying that the event or person being described is extraordinary or uncommon.

E.g.,

“Cel mai murdar om din lume a murit la 94 de ani! Vei fi șocat de motivul pentru care nu s-a spălat timp de 65 de ani!” [“The dirtiest man in the world has died at the age of 94! You’ll be shocked to find out why he hasn’t washed for 65 years!”]

“Incredibil! La 106 ani, Apo Whang-Od devine cel mai bătrân model pe coperta Vogue—povestea care a uimit lumea!” [“Unbelievable! At 106 years old, Apo Whang-Od becomes the oldest model on the cover of Vogue—the story that has stunned the world!”]

Likewise, adverbs of manner such as “brusc” (“suddenly”) or “incredibil” (“unbelievable”) are employed to highlight the novelty or intensity of the events. These words imply an unexpected shift, sparking curiosity about the reasons behind this decision.

E.g.,

“Șoc în junglă! Alex Bogdan părăsește brusc competiția “Sunt celebru, scoate-mă de aici”! Colegii sunt revoltați!” [“Shock in the jungle! Alex Bogdan suddenly leaves the competition “I’m famous, get me out of here”! His colleagues are outraged”]

“Incredibil! La 106 ani, Apo Whang-Od devine cel mai bătrân model pe coperta Vogue—povestea care a uimit lumea!” [“Unbelievable! At 106 years old, Apo Whang-Od becomes the oldest model on the cover of Vogue—the story that has stunned the world!”]

➢

Using numbers

Numbers in clickbait headlines are highly effective in grabbing readers’ attention and arousing curiosity. They provide a sense of clarity, specificity, and sometimes an element of surprise, lending the content an appearance of credibility and concreteness.

E.g.,

“Fața ascunsă a lui Tutankhamon dezvăluită după 3.300 de ani! Nu o să crezi cum arată legendarul faraon!” [“Tutankhamun’s hidden face revealed after 3,300 years! You won’t believe what the legendary pharaoh looks like!”]

“Dezastrul falimentelor din asigurări: Peste 1,6 miliarde de lei pierduți și riscuri uriașe pentru clienți! Află care companie este cea mai afectată!” [“The disaster of insurance bankruptcies: Over 1.6 billion lei lost and huge risks for customers! Find out which company is most affected!”]

➢

Using exclamation marks

Exclamation marks signal a shocking event, giving the impression of urgency and piquing curiosity.

E.g.,

“Bebeto Yildirim, milionarul viral pe TikTok! Află secretul care l-a transformat în cel mai urmărit ‘Sugar Daddy’ din România!” [“Bebeto Yildirim, the viral millionaire on TikTok! Find out the secret that turned him into the most followed ‘Sugar Daddy’ in Romania!”]

“România trădată în aderarea la Schengen! Manfred Weber, revoltat: “Nu este corect!”” [“Romania betrayed in joining Schengen! Manfred Weber, outraged: “It’s not fair!””]

• Stylistic level

Beyond the linguistic features previously mentioned, clickbait headlines often employ additional stylistic effects to spark the readers’ curiosity. These techniques make the information appear more enticing, provocative, and deserving of attention, as detailed in the following examples.

One of the stylistic techniques employed by ChatGPT was the use of metaphors and vivid images. Metaphors create vivid mental images, making headlines more memorable and dramatic. They invite readers to interpret the situation figuratively, which can heighten interest. For instance, “își întinde tentaculele” (“spreading its tentacles”) implies aggressive expansion, vividly depicting the Chinese influence, while “bătălia incendiară” (“incendiary battle”) evokes a heated confrontation, drawing attention through the intensity of the emotions portrayed.

E.g.,

“China își întinde tentaculele în Orientul Mijlociu! SUA avertizează: Ordinea internațională este în pericol!” [“China is spreading its tentacles in the Middle East! The US warns: international order is in danger!”]

“Bătălia incendiară de la Vocea României! Iulian Nunucă a lăsat juriul fără cuvinte, iar Bogdan Jeler surprinde cu o decizie șocantă!” [“The incendiary battle from the Voice of Romania! Iulian Nunucă has left the jury speechless, and Bogdan Jeler surprises with a shocking decision!”]

Another technique found in the analyzed corpus is direct questioning and the use of conversational terms. In Romanian, the T–V distinction to convey politeness plays a key role in communication, reflecting how we relate to the other person based on context formality and the level of respect we intend to show. Phrasings such as “nu o să crezi” (“you won’t believe it”) or “află cum” (“find out how”) directly address the reader, creating the impression that the headline is speaking to them personally. This direct approach is designed to arouse curiosity through a friendly, inviting tone.

E.g.,

“Fața ascunsă a lui Tutankhamon dezvăluită după 3.300 de ani! Nu o să crezi cum arată legendarul faraon!” [“Tutankhamun’s hidden face revealed after 3300 years! You won’t believe what the legendary pharaoh looks like!”]

“Celine Dion, diagnosticul care îi devastează cariera! Află cum îi afectează boala concertul din România!” [“Celine Dion, the diagnosis that devastates her career! Find out how the disease affects her concert in Romania!”]

Other headlines use contrasts and antitheses to suggest a sudden change or unusual situation, which makes them appealing. For example, the headline below shows a strong contrast between the idea of “eurodeputat” (“MEP”), a respectable civil servant, and “arestat cu 600.000 de euro cash” (“arrested with €600,000 cash”), suggesting corruption or irregularities. This brings an element of surprise and intrigue, making the reader want to know more.

“Scandal uluitor în Parlamentul European! Fost eurodeputat arestat cu 600.000 de euro cash ascunși acasă!” [“Amazing scandal in the European Parliament! Former MEP arrested with 600,000 euros in cash hidden at home!”]

Together, lexis, grammar, and stylistics reveal how clickbait operates on multiple linguistic levels to engage readers. Lexical choices set the tone and evoke emotion and grammatical structures establish immediacy and suspense, while stylistic elements ensure vividness and memorability. This multi-layered linguistic approach not only shapes reader behavior but also provides insights into how digital media manipulate language to drive consumption.

• Human-authored clickbait headlines vs. ChatGPT-generated clickbait headlines

In examining the respondents’ preferences regarding the clickbait headlines, a notable distinction emerged between human-authored headlines and those generated by ChatGPT. While 29.20% of the participants favored the clickbait headlines crafted by humans, a more significant 37.50% expressed a preference for the ChatGPT-generated headlines. The analysis delves into the characteristics that set these two types apart, seeking to understand why the AI-generated titles resonated more effectively with readers. Key differences may lie in the linguistic features, emotional engagement, and stylistic choices employed in the headlines.

Human-authored headlines have a more sober and informative tone, often focused on facts and analysis, even if they are meant to be clickbait. In the examples below, besides the clickbait element expressed through the interrogative form and the surprise “declarant mort” (“declared dead”) and drama (through EXCLUSIVE NEWS in capitals) elements, the headlines also promise to offer specific answers, e.g., when to read the Akathist or the artist’s answer to the blunder involving irony.

E.g.,

“Când este bine să citești Acatistul Nașterii Domnului. Este considerată cea mai puternică rugăciune pentru Crăciun” [“When is it advisable to read the Nativity Akathist. It is considered the most powerful prayer for Christmas”]

“Cornel Constantiniu a fost declarat mort, la Festivalul Mamaia. Ce replică a avut artistul după gafa Corinei Chiriac: “Bășcălia asta ce este?”/EXCLUSIV” [Cornel Constantiniu was declared dead at the Mamaia Festival. What was the artist’s response after Corina Chiriac’s blunder: “Are you making fun of me?”/EXCLUSIVE NEWS]

In contrast, the ChatGPT-generated headlines often leverage a more sensationalist and provocative tone, employing dynamic verbs, vivid adjectives, with exclamations and wording that suggest urgency or scandal. Headlines are designed to evoke intense emotions and curiosity. The examples below are very suggestive in this respect, especially when compared to the human-authored headlines; they use exclamations, hyperbolic words (shock), a conversational tone and shorter and more to-the-point sentences.

E.g.,

“Cea mai puternică rugăciune de Crăciun care îți aduce sănătate și pace! Află când trebuie să citești Acatistul Nașterii Domnului!” [“The most powerful Christmas prayer that brings you health and peace! Find out when to read the Nativity Akathist!”]

“ȘOC la Festivalul Mamaia! Cornel Constantiniu, declarat mort din greșeală! Adevărul despre gafa monumentală a Corinei Chiriac!” [“SHOCK at the Mamaia Festival! Cornel Constantiniu, declared dead by mistake! The truth about Corina Chiriac’s monumental blunder!”]

The presented examples illustrate the complex art of combining lexical–grammatical affordances with stylistic strategies to create attention-grabbing headlines, appealing to young audiences.

4. Discussion

The results of this study underscore the significant influence of headlines on readers’ decisions to engage with an article, confirming that headlines are among the main factors driving interaction with media content. About 80% of participants indicated that headlines play a critical role in their reading choices, highlighting the need for strategic headline creation in content production. Additionally, the importance of specific headline characteristics—such as clarity and relevance—was emphasized by most respondents, underscoring that these elements are fundamental in engaging audiences effectively.

In light of this, the study also explores the evolving role of AI in journalism, focusing on its potential to enhance the creation of attention-grabbing headlines within an increasingly competitive digital landscape. As AI technologies become integrated into media production, they offer new efficiencies in headline generation while simultaneously introducing concerns around authenticity and credibility—key pillars of journalistic integrity.

The findings reveal that AI has become a valuable tool in the newsroom, helping to streamline tasks such as data analysis and basic content generation. Yet, despite its convenience, AI brings with it a crucial dilemma. On the one hand, it can quickly generate content that is catchy and tailored to audience preferences; on the other, it risks eroding the very standards of ethical and diverse reporting that foster trust among readers. While sensationalist headlines are effective in drawing readers in, there is a growing discontent when these “clickbait” headlines do not deliver the promised content, ultimately leading readers to view such practices skeptically.

The analysis of the differences between human-authored and AI-generated clickbait headlines offers a valuable window into what truly engages readers and the ethical implications behind these attention-grabbing techniques. By studying how people respond to the subtle linguistic choices that distinguish these headlines, the power behind certain phrasing or structures can be better understood. For content creators, this knowledge is of utmost importance, as it helps pinpoint what drives reader curiosity—insights that can lead to more compelling yet responsible content creation strategies.

Ethics come to the forefront in such an analysis, given the often-controversial nature of clickbait. Headlines that use sensationalism run the risk of misleading or manipulating readers, which raises questions about their transparency. Comparing human and AI-crafted headlines allows creators to assess whether AI might unintentionally reinforce these concerns, or alternatively, if it can be trained to use engaging tactics responsibly.

Furthermore, these results also present the evolution of language models such as ChatGPT. By understanding why certain AI-generated headlines resonate with readers, developers can refine the model to better align with human preferences for authenticity, effectiveness, and ethical standards. This kind of analysis therefore supports the creation of tools that produce impactful yet trustworthy language—a priority for responsible AI.

Equally important, this analysis can also help readers in recognizing the methods that media use to grab their attention. Developing an awareness of these techniques supports media literacy, empowering audiences to critically assess the headlines and content they encounter.

Ultimately, exploring these differences between human-authored and AI-generated headlines sheds light on AI’s potential as a creative partner. It allows us to examine how AI might complement or even extend human intuition and creativity, particularly in fields such as journalism, where headline effectiveness is key. By combining effectiveness, ethics, and audience trust, this analysis reveals a thoughtful approach to using AI as a tool in modern content creation.

5. Limitations and Further Directions

The study presents several limitations that may influence the generalizability and applicability of its conclusions. Firstly, all research participants were Romanian students, which restricts the demographic and cultural diversity of the sample. Their preferences for AI-generated headlines over human-crafted ones may reflect the particularities of the Romanian cultural and educational context, limiting the applicability of the results to other cultural contexts. Additionally, the study does not include participants from other countries or with different media consumption habits, which could lead to varied reactions to AI-generated headlines. This lack of diversity among participants makes it challenging to extrapolate the results to a global level.

At the same time, it is to be acknowledged that further research is necessary in order to fully understand the rationale behind the respondents’ choices for a certain type of headline. Understanding why respondents prefer certain types of headlines—whether clickbait or informative—requires deeper exploration of the psychological, cultural, and contextual factors influencing their choices. Investigating these dynamics through behavioral studies, cross-platform analyses, and cultural comparisons can provide valuable insights into how headlines shape digital media engagement while balancing ethical journalism with audience expectations.

Moreover, the AI-generated headlines were produced using only a single ChatGPT model, without investigating the variability offered by other versions or alternative models. This approach reduces the ability to fully understand the potential of various AI models in generating engaging headlines. Participant preferences were evaluated through a questionnaire, a method that introduces a degree of subjectivity in assessing the headlines. In the absence of objective measures, such as click-through rates in real media contexts, conclusions regarding headline appeal are limited to participants’ reported perceptions.

These limitations highlight the need for future studies with more diverse samples and objective measures, allowing for a more comprehensive evaluation of AI-generated headlines.

This study serves as a catalyst for further exploration into the complex role of AI in media and journalism, offering a springboard for a range of investigative paths that delve deeper into both the opportunities and challenges AI presents. Its findings reveal how AI-generated headlines influence reader behavior, paving the way for nuanced research on AI’s impact on public trust in the news. As AI becomes increasingly embedded in newsrooms, one essential question for future studies could be posed: can AI be trained to create headlines that maintain or even enhance trust, while still captivating readers? This would be especially relevant in light of growing concerns over clickbait and sensationalism in digital media. Moreover, this research could ignite a new wave of longitudinal studies. By tracking the evolution of AI in journalism across regions and cultures, future research could shed light on the long-term impacts of AI on media quality, reader engagement, and even societal trust in journalism. Such comparative studies could uncover the cultural and regulatory factors that shape AI’s role in different contexts, revealing both commonalities and unique challenges.

It also highlights readers’ appreciation for clarity, relevance, and conciseness, pointing to fertile ground for research into how audience preferences adapt to repeated exposure to both AI-generated and human-created content. A possible area of focus could be the behavioral shifts in readers over time, exploring how continuous interaction with clickbait and straightforward headlines might alter trust in news sources.

Additionally, by comparing human versus AI-generated headlines, this study hints at the potential for effective human–AI collaboration in journalism. Instead of AI replacing human creativity, future research could explore how humans and AI might co-create content, balancing engagement with trustworthiness. This research could identify ways to preserve the human touch in journalism while utilizing AI to enhance productivity and reach.

Lastly, this study brings attention to the ongoing challenges posed by clickbait. While some see clickbait as a necessary evil in digital media, others view it as detrimental to credibility. This study could guide future research into policies that regulate clickbait, investigating the balance between crafting engaging headlines and upholding the ethical standards of journalism.