An Alien in the Newsroom: AI Anxiety in European and American Newspapers

Reviewer’s comment

It is of particular interest the semi-automated method that is

performed, although it would be good to know more about the

limitations, reliability and how human supervision has been

applied to counteract this automation. For example, when the

authors used GPT-4o model to detect the negative scenarios

identified by Cave et al. (2020), how was this process? How did

they ensure that this analysis matched the same definitions for

each scenario? Which were the instructions? Has manual coding

been carried out to compare results?

Answer

We specify the prompt used: "This is a news headline related to artificial intelligence. Read it slowly and answer one by one the following questions. AI or the AI model mentioned in the headline is represented as a technology that...

a) can dehumanize us and make us lose our essence and values.

b) can uprise and escape human control.

c) can make humans obsolete and replace them.

d) is dangerous because it can be used to discriminate, kill, disinform, steal, etc.

Answer all the questions separated by commas. Answer only ""yes"" or ""no"" for each one, without further explanation. For example, your answer could look like this: ""yes, no, no, no""."

We contrasted the results with the output of several sentiment analysis methods and manually coded the headlines that presented divergent classification across methods.

Reviewer’s comment

As a minor improvement suggestion, it is recommended that

authors add “Conclusions” as a section in their manuscript (either

in a separate section or together with “4. Discussion”).

Some conclusions would improve with a more reflective

approach based on previous literature: is there any indication of

why regional media “present a growing tendency to anxiety?” Is

this anxiety really a perception of journalists, or also a narrative

used by the media that instrumentalizes this anxiety through

clickbait techniques to attract the attention of the audience and

monetize content? These can also serve as future lines of study.

The authors adequately describe the limitations of the study and

propose some lines of future studies.

Answer

We have integrated the conclusion in the discussion section. We also define two research questions that strengthen the article’s coherence and clarity. The reason for the growing tendency to anxiety in regional news outlets are now specified in the final section as part of the answer to those research questions. We find a correlation between a sudden increase in the number of news stories about AI in regional news media and a higher level of anxiety. This suggests that lower scores in the anxiety index don’t depend on the prominence, the prestige, or the reach of a news outlet, but rather on its sustained commitment to the coverage of AI and the corresponding expertise in the newsroom.

Reviewer’s comment

The references are relevant and up-to-date, but the list should be

checked to match the guidelines of the journal.

Answer

The references’ format has been adapted to the journal’s guidelines.

Reviewer’s comment

First, I would like to see a clearer discernment of AI anxiety

between the media and the public. On page 3, for example, the

authors delve into the gap between portrayals of AI in the media

and the social views. This raises two questions: 1) what exactly

does the “media” entail? I ask this because I was reading, up

until now, with the assumption that this is the news media, but

then the references to popular media (movies, such as the

terminator) made me question this assumption. I am also not

clear of what AGI refers to here. A lot is discussed in a short time

and I would like to see a bit more elaboration on the terms, here.

2) The paper was set up with the premise of media coverage

leading the public perception of AI. How then do you explain

these discrepancies of media and public perceptions of AI? Why

is there a positive tone about AI in the digital media but a

negative one on social media? This needs some working out and

could also be tied into the first point above.

Answer

We have reworded several passages to avoid the misunderstanding between news media (the focus of this article) and other media, such as movies or science fiction. We clarify the meaning of “AGI”: artificial general intelligence. In page 3, we explicitly present the concept of “AI anxiety” as a way to explain and move beyond discrepancies between media and public perception of AI. The examples of this section demonstrate that the key differentiation in analysing coverage of AI is not between positive and negative, but rather between a sober and realistic representation of the technology on the one hand, and an exaggerated and distorted one (both for good and bad) on the other. The concept of ‘AI anxiety’ accounts for that difference by focusing on the wrong conceptualization of what AI is and can be. 

Reviewer’s comment

Second, I see some major issues with the set up of the

theoretical framework and the conceptual definitions and

measurement of the content. From page 3 onwards, I had a bit of

a hard time with the theory. It read somewhat like blocks of

summary of various studies and texts, listing multiple theories

and factors of AI anxiety in their complete form, but little

interpretation of integration is done with regard to the

overarching RQ, which I was also missing. How do all of these

theories fit together? Instead of listing all of the findings and

conepts, can you interpret the overall findings of these studies in

relation to each other and integrate them into your interpretations

for the research question you are trying to answer. This happens

only at the very end, where the authors write “all of the

aforementioned studies assume that the media play a key role in

public perception of AI…etc” but this synthesizing of the literature

only happens at the very end. I would like to see this more as a

narrative and the literature that is cited interpreted in light if the

research goal throughout, rather than at the end.

Answer

We have added two research questions in the beginning of the paper: How has the launch of ChatGPT influenced the level of AI anxiety in news media coverage across national and regional newspapers? (RQ1) What are the key factors contributing to AI anxiety in regional and national media? (RQ2). Moreover, we have reorganized the theoretical framework to connect each theory with these questions. Instead of listing findings and concepts related to AI anxiety, we now redirect our literature review to the goal of identifying four causes of AI anxiety: three related to technology misconceptions (socio-technical blindness, anthropomorphism, future orientation) and one related to journalistic practices (clickbait).

Reviewer’s comment

Third, the section of “the alies vs the sociotechnical AI narrative”

is a bit confusing to me. What is the purpose of this section? It

ends with a lot of questions, but it is not clear whather the

authors aim to answer these questions in the paper. I am left

unsure of what the aim of the methods is, and whether I should

read it with these questions in mind and, if not, what the purpose

of these questions is. There are more lists of concepts (a, b, c,

d,..) such as in the sections before, but no explanation of what

they contribute or how the reader should interpret them.

Answer

We have eliminated this section and integrated it into the end of the previous one, as a way to exemplify the idea of AI anxiety.

Reviewer’s comment

Overall, for the theoretical framework, I would suggest defining a

central Research Question, and then centerering the theory

around that RQ, making it clear how it contributes to it. I would

also, as the theory progresses, develop sub-RQs, so that it is

clear how the theory builds up to more specific questions that the

methods and materials then aim to answer. As is, the section

reads as a summary of literature, but the aim or the

conceptualization of the research terms of it is not clear, so that it

needs major re-working.

Answer

The whole theoretical framework has been subjected to major reorganization, including the definition of two research questions, the connection between them and the literature review, and a clearer explanation of what theories are taken from previous research and how the contribute to our analysis.

Reviewer’s comment

In the methods, I see some justification issues. Primarily, I would

like to see at least some justification for the country selection,

and more importantly, the selection of the unit of measurement.

The theory section cites many interpretations of AI, but these are

taken from discourse about the topic. Why does it suffice, for this

study, to look at only the headlines? This is a crucial point and, in

my opinion, warrants more attention in the methods (and

perhaps theoretical) discussion of the paper.

Answer

The country selection has been clarified: these are the top countries in the SCIMago classification (Trillo-Domínguez et al. 2023) corresponding to the end of the period to be studied (summer 2023). From each country, the main newspaper with international coverage and the main newspaper with regional coverage according to the same ranking were chosen. We have also provided with reasons and precedents to chose headlines as the unit of analysis. Here we follow previous studies on headlines about multiple topics such as the Covid-19 (Aslam et al. 2020), fake news (Calvillo and Smelter 2020), partisan news (Ross et al. 2021), and the subject of our analysis, AI coverage (Roe and Perkins 2023, Ouchchy, Coin, and Dubljević 2020). Even if the headline is only a partial and sometimes misleading component of the whole news story, it is this limitation that makes it a valuable unit for analysing the potential shortcomings of AI representation (Leufer 2020, Becket et al. 2023) and offers unique features that are particularly relevant to our study. For example, headlines in online news outlets may act as clickbait to make the reader access the whole article (kuiken et al. 2017) and show higher levels of anthropomorphism (Cheng et al. 2024).

Reviewer’s comment

In the results, some interpretation is needed. It starts of quite

stark, with descriptives about pre-and post-chatgpt Ai coverage.

Perhaps this could be introduced and framed as an introduction

to the results because (as far as I know) this was not the aim of

the paper – and it is perhaps not all too surprising of a finding to

see that AI coverage increased after the release of ChatGPT. The

authors address some differences between the country cases (or

outlet cases) but why is this interesting? I am not saying that it

isn’t, but if this is going to be a part of the paper, then justification

is needed on why these differences matter. Again, more

interpretation for the reader is key.

Answer

We have reduced this section by eliminating the figure with the growth per country. However, we maintain the differentiation across news outlets and specify why this is relevant to answer our RQ: It is to be expected that the newspapers that have suddenly increased their coverage on AI will offer less balanced and nuanced stories than those that had been following the topic since before the launch of the popular chatbot, a hypothesis confirmed in the following sections.

Reviewer’s comment

Section 3.2 of the results is, from what I understand, the main

focus of the paper. But again here, only one three sentences are

dedicated to the overall results (pre-table 4). I would like to see

elaboration on this – what is it that you found? “a clear

dominance of positive versus negative headline, both before and

after the launch” – what does this mean? Are they getting more

negative and how does that answer your overall RQ? Then we

can talk about it more in detail (like below).

Answer

We have reduced the sentiment analysis section and clearly connect its main findings with the research question, as emotions detected in the corpus anticipate major trends of the AI anxiety index proposed in the article.

Reviewer’s comment

One minor point: The point of the graphs per months is to

visualize the change after the launch of ChatGPT, I would guess.

It would be helpful to indicate in those graphs where that launch

was (like in Figure 5), so that it is easier to interpret for the

reader.

Answer

The line indicating the launch of ChatGPT has been added.

Reviewer’s comment

Overall, the results sections was hard to get through, for me.

There are so many detail results and numbers, but because of

the lack of clarity in the beginning of the paper, I was not sure

what I was looking for in terms of answers. I appreciate the

visualizations and the many analyses that were run, but without

proper interpretation, I am not entirely sure what the answer to

the question of how the media portrays AI is. This is especially so

given the many cases that are looked at, but neither the selection

of the cases nor the discrepancies in the findings for the cases

(and the discrepancies across variables) is discussed properly.

Answer

The link between results and the research questions is clearer now. We have explicitly group the findings around RQ1 and RQ2. Discussion paragraphs were included to provide more context on the main insights. In particular, we highlight the correlation between a steeper increase in AI coverage after the launch of ChatGPT and a larger increase in the anxiety index. The three newspapers that increased their production the most (Münchner Merkur, Manchester Evening News, San Francisco Chronicle) also show significant gains in the index. On the opposite side, smaller changes in the coverage of Le Monde, The Guardian and Frankfurter Allgemeine Zeitung are associated with a decrease in the anxiety index.

Reviewer’s comment

The paper has a lot of potential, but it suffers from a lack of

structure and clear aim and conclusion. This runs throughout the

paper, starting with the Intro and a missing definition of a

Research Question, throughout the theory and missing sub-

questions and conceptualizations of measured variables, as well

as the methods and results. With some major revisions, I think

this can be a great contribution to knowledge on the media

coverage of AI, but the authors need to be much more clear

about the aim, procedure and results of the paper.

Answer

As with the previous sections, the discussion has been largely rephrased to establish a clearer connection between the introduction, the research questions and the results.

Reviewer’s comment

First, I would like to see a clearer discernment of AI anxiety

between the media and the public. On page 3, for example, the

authors delve into the gap between portrayals of AI in the media

and the social views. This raises two questions: 1) what exactly

does the “media” entail? I ask this because I was reading, up

until now, with the assumption that this is the news media, but

then the references to popular media (movies, such as the

terminator) made me question this assumption. I am also not

clear of what AGI refers to here. A lot is discussed in a short time

and I would like to see a bit more elaboration on the terms, here.

2) The paper was set up with the premise of media coverage

leading the public perception of AI. How then do you explain

these discrepancies of media and public perceptions of AI? Why

is there a positive tone about AI in the digital media but a

negative one on social media? This needs some working out and

could also be tied into the first point above.

Answer

We have reworded several passages to avoid the misunderstanding between news media (the focus of this article) and other media, such as movies or science fiction. We clarify the meaning of “AGI”: artificial general intelligence. In page 3, we explicitly present the concept of “AI anxiety” as a way to explain and move beyond discrepancies between media and public perception of AI. The examples of this section demonstrate that the key differentiation in analysing coverage of AI is not between positive and negative, but rather between a sober and realistic representation of the technology on the one hand, and an exaggerated and distorted one (both for good and bad) on the other. The concept of ‘AI anxiety’ accounts for that difference by focusing on the wrong conceptualization of what AI is and can be. 

Reviewer’s comment

Second, I see some major issues with the set up of the

theoretical framework and the conceptual definitions and

measurement of the content. From page 3 onwards, I had a bit of

a hard time with the theory. It read somewhat like blocks of

summary of various studies and texts, listing multiple theories

and factors of AI anxiety in their complete form, but little

interpretation of integration is done with regard to the

overarching RQ, which I was also missing. How do all of these

theories fit together? Instead of listing all of the findings and

conepts, can you interpret the overall findings of these studies in

relation to each other and integrate them into your interpretations

for the research question you are trying to answer. This happens

only at the very end, where the authors write “all of the

aforementioned studies assume that the media play a key role in

public perception of AI…etc” but this synthesizing of the literature

only happens at the very end. I would like to see this more as a

narrative and the literature that is cited interpreted in light if the

research goal throughout, rather than at the end.

Answer

We have added two research questions in the beginning of the paper: How has the launch of ChatGPT influenced the level of AI anxiety in news media coverage across national and regional newspapers? (RQ1) What are the key factors contributing to AI anxiety in regional and national media? (RQ2). Moreover, we have reorganized the theoretical framework to connect each theory with these questions. Instead of listing findings and concepts related to AI anxiety, we now redirect our literature review to the goal of identifying four causes of AI anxiety: three related to technology misconceptions (socio-technical blindness, anthropomorphism, future orientation) and one related to journalistic practices (clickbait).

Reviewer’s comment

Third, the section of “the alies vs the sociotechnical AI narrative”

is a bit confusing to me. What is the purpose of this section? It

ends with a lot of questions, but it is not clear whather the

authors aim to answer these questions in the paper. I am left

unsure of what the aim of the methods is, and whether I should

read it with these questions in mind and, if not, what the purpose

of these questions is. There are more lists of concepts (a, b, c,

d,..) such as in the sections before, but no explanation of what

they contribute or how the reader should interpret them.

Answer

We have eliminated this section and integrated it into the end of the previous one, as a way to exemplify the idea of AI anxiety.

Reviewer’s comment

Overall, for the theoretical framework, I would suggest defining a

central Research Question, and then centerering the theory

around that RQ, making it clear how it contributes to it. I would

also, as the theory progresses, develop sub-RQs, so that it is

clear how the theory builds up to more specific questions that the

methods and materials then aim to answer. As is, the section

reads as a summary of literature, but the aim or the

conceptualization of the research terms of it is not clear, so that it

needs major re-working.

Answer

The whole theoretical framework has been subjected to major reorganization, including the definition of two research questions, the connection between them and the literature review, and a clearer explanation of what theories are taken from previous research and how the contribute to our analysis.

Reviewer’s comment

In the methods, I see some justification issues. Primarily, I would

like to see at least some justification for the country selection,

and more importantly, the selection of the unit of measurement.

The theory section cites many interpretations of AI, but these are

taken from discourse about the topic. Why does it suffice, for this

study, to look at only the headlines? This is a crucial point and, in

my opinion, warrants more attention in the methods (and

perhaps theoretical) discussion of the paper.

Answer

The country selection has been clarified: these are the top countries in the SCIMago classification (Trillo-Domínguez et al. 2023) corresponding to the end of the period to be studied (summer 2023). From each country, the main newspaper with international coverage and the main newspaper with regional coverage according to the same ranking were chosen. We have also provided with reasons and precedents to chose headlines as the unit of analysis. Here we follow previous studies on headlines about multiple topics such as the Covid-19 (Aslam et al. 2020), fake news (Calvillo and Smelter 2020), partisan news (Ross et al. 2021), and the subject of our analysis, AI coverage (Roe and Perkins 2023, Ouchchy, Coin, and Dubljević 2020). Even if the headline is only a partial and sometimes misleading component of the whole news story, it is this limitation that makes it a valuable unit for analysing the potential shortcomings of AI representation (Leufer 2020, Becket et al. 2023) and offers unique features that are particularly relevant to our study. For example, headlines in online news outlets may act as clickbait to make the reader access the whole article (kuiken et al. 2017) and show higher levels of anthropomorphism (Cheng et al. 2024).

Reviewer’s comment

In the results, some interpretation is needed. It starts of quite

stark, with descriptives about pre-and post-chatgpt Ai coverage.

Perhaps this could be introduced and framed as an introduction

to the results because (as far as I know) this was not the aim of

the paper – and it is perhaps not all too surprising of a finding to

see that AI coverage increased after the release of ChatGPT. The

authors address some differences between the country cases (or

outlet cases) but why is this interesting? I am not saying that it

isn’t, but if this is going to be a part of the paper, then justification

is needed on why these differences matter. Again, more

interpretation for the reader is key.

Answer

We have reduced this section by eliminating the figure with the growth per country. However, we maintain the differentiation across news outlets and specify why this is relevant to answer our RQ: It is to be expected that the newspapers that have suddenly increased their coverage on AI will offer less balanced and nuanced stories than those that had been following the topic since before the launch of the popular chatbot, a hypothesis confirmed in the following sections.

Reviewer’s comment

Section 3.2 of the results is, from what I understand, the main

focus of the paper. But again here, only one three sentences are

dedicated to the overall results (pre-table 4). I would like to see

elaboration on this – what is it that you found? “a clear

dominance of positive versus negative headline, both before and

after the launch” – what does this mean? Are they getting more

negative and how does that answer your overall RQ? Then we

can talk about it more in detail (like below).

Answer

We have reduced the sentiment analysis section and clearly connect its main findings with the research question, as emotions detected in the corpus anticipate major trends of the AI anxiety index proposed in the article.

Reviewer’s comment

One minor point: The point of the graphs per months is to

visualize the change after the launch of ChatGPT, I would guess.

It would be helpful to indicate in those graphs where that launch

was (like in Figure 5), so that it is easier to interpret for the

reader.

Answer

The line indicating the launch of ChatGPT has been added.

Reviewer’s comment

Overall, the results sections was hard to get through, for me.

There are so many detail results and numbers, but because of

the lack of clarity in the beginning of the paper, I was not sure

what I was looking for in terms of answers. I appreciate the

visualizations and the many analyses that were run, but without

proper interpretation, I am not entirely sure what the answer to

the question of how the media portrays AI is. This is especially so

given the many cases that are looked at, but neither the selection

of the cases nor the discrepancies in the findings for the cases

(and the discrepancies across variables) is discussed properly.

Answer

The link between results and the research questions is clearer now. We have explicitly group the findings around RQ1 and RQ2. Discussion paragraphs were included to provide more context on the main insights. In particular, we highlight the correlation between a steeper increase in AI coverage after the launch of ChatGPT and a larger increase in the anxiety index. The three newspapers that increased their production the most (Münchner Merkur, Manchester Evening News, San Francisco Chronicle) also show significant gains in the index. On the opposite side, smaller changes in the coverage of Le Monde, The Guardian and Frankfurter Allgemeine Zeitung are associated with a decrease in the anxiety index.

Reviewer’s comment

The paper has a lot of potential, but it suffers from a lack of

structure and clear aim and conclusion. This runs throughout the

paper, starting with the Intro and a missing definition of a

Research Question, throughout the theory and missing sub-

questions and conceptualizations of measured variables, as well

as the methods and results. With some major revisions, I think

this can be a great contribution to knowledge on the media

coverage of AI, but the authors need to be much more clear

about the aim, procedure and results of the paper.

Answer

As with the previous sections, the discussion has been largely rephrased to establish a clearer connection between the introduction, the research questions and the results.