Author: Chen Sha (Ph.D. candidate, School of Culture and Communication, Shandong University).
*: Young Journalists, Issue 19, 2023.
Introduction: While generative AI assists news production and promotes news reporting innovation, there are also problems such as providing inaccurate facts, opaque algorithms, and controversial application boundaries.
Generative AI, or "generative AI," is an AI technology that focuses on generating or creating new content that uses existing datasets such as text, audio, or images for machine learning and then generates entirely new content. [1] With the advent of GPT-4 and its derivative models, generative AI has become more and more powerful, and it can process a variety of text formats such as speech, images, and gestures in addition to text writing. It not only continues to be an important topic in news reporting, but is also becoming an important infrastructure in the news production chain, which means that news production will officially enter a full range of human-machine collaboration mode. How to use such technologies responsibly has become an urgent question to be explored.
Types of applications of generative AI in news production.
The use of AI to assist news production is an important path for digital transformation. Over the past decade, innovation in intelligent news production has gone through three phases: automation, augmentation, and generation. [2] Compared with the previous two stages, generative AI has a wider range of application scenarios due to the addition of pre-training, instruction tuning, and reinforcement learning based on human feedback in the algorithm system. Scholar Diakopoulos summarized 13 scenarios in which journalists use generative AI to assist news: content discovery, file analysis, translation, prompt guidance, social content creation, automated writing (structured data), automated writing (unstructured data), news lead discovery, summary generation, comment review, content format conversion, headline optimization, AB testing, personalized distribution, etc. [3] Different ** access to generative AI have different paths, and the use scenarios and functional applications will be different.
At present, there are three main types of ways to apply generative AI at home and abroad.
1) General tool type.
The relationship between the general tool type and the generative AI system is independent of each other, and like other users of the platform, they can use their functions to assist in news gathering and writing by logging in to the generative AI service platform. Usually, ** people only need to register a platform account to use the product. This method is low-cost, flexible, and the most common way to access generative AI. Taking ChatGPT as an example, reporters can input questions that need to be queried into the chat dialog, and give prompt engineering instructions in the dialog box, and ChatGPT can quickly provide information about their knowledge database.
In the era of picture reading and short editing, visual elements have become as important as information, and the social media platform represented by Douyin has developed an automated editing function in its editing application, which reflects that generative AI can not only transform complex events and esoteric data into easy-to-understand charts or images, but also create effects that cannot be achieved by traditional news through visual generation capabilities, providing more innovative possibilities for news production. [4] At present, the Xinhua News Agency's audio department has set up an AIGC application innovation studio, and "AIGC Tells the Truth" has published 5 creative short articles, with excellent quality and good communication effect.
2) Platform access type.
The platform access type is a generative AI system platform that provides API interfaces for content production, and provides content analysis, sentiment analysis, event extraction, summary generation, personalized recommendation, content review, visual generation and other services. In this type of human-robot collaborative production, the first person is mainly responsible for providing text data, and the machine is responsible for output and presentation. In the interaction between the machine and the end user, it plays the role of an intermediary between the end user and the generative AI interaction, that is, the original text of the news agenda is provided and set - the generative AI system platform is responsible for the technical implementation - the terminal display is provided, and its essence is the interaction between the end user and the generative AI platform.
The New York Times leveraged ChatGPT to create a Valentine's Day message generator with a combination of prompts, an interactive news campaign called "A Valentine, from A."i.to you", the poster news automatically generated by the news sharing function of the mobile terminal of the domestic surging news, are typical representatives of the platform access type. In traditional interactive news production, it is necessary to write and algorithm to design interactive topics for interactive news, but by intervening in the generative AI interface, generative AI can automatically generate interactive news products by simply issuing design instructions for interactive news. To a certain extent, generative AI can assist journalists in their creative expression in a relatively concise way.
3) Proprietary system type.
The construction of this type of application requires strong technical support, such as building a digital library and developing proprietary algorithms for itself. Its advantage is that, on the one hand, it can embed the best values into the system and generate content that conforms to the best positioning; On the other hand, the source database comes from the ** itself, and there are relatively few dirty data sources, which can reduce the risk of output bias. It usually has a high input cost and is suitable for specific large **groups**, such as business ** in the fields of economy, sports, politics, etc.
AngleKindling, developed by academic institutions to support multiple angles of news reporting, and BloombergGPT released by Bloomberg are this type of generative AI. The proprietary system collects a variety of reporting angles and methods of different news topics, which can be targeted training on the first-class data, and generate multi-angle industry reports through sentiment analysis, named entity recognition, news classification and Q&A. In other words, this proprietary generative AI can be invoked in the planning, gathering, editing, proofreading, reviewing, distribution and other processes of news production, providing distinct personalized services in all aspects of auxiliary news production, and meeting the needs of specific information, including: (1) proprietary interpretation of natural language, such as in the financial interactive system, typing "apple" will point more to Apple than to the fruit apple; (2) Links to proprietary databases, such as links to the current affairs transaction databases of listed companies; (3) Specific style news settings, such as the production of automated news writing suitable for ** positioning, etc.
The real-world problem of generative AI participation in news production.
Like humans, machines are not omnipotent, and the content they participate in production will also make mistakes, and different people have different perceptions of human-computer interaction. In practice, generative AI-assisted news production still has the following problems.
1) Risk of providing inaccurate facts.
Truthfulness and accuracy are the basic principles of news reporting, but no technology can be 100% accurate when translating a complex world. Schibsted, a Norwegian newspaper conglomerate, found that 1 in 10 content contained "illusions" or fictional elements in its experimental generative AI-assisted news summaries. [5] Similarly, when ChatGPT encounters data that the database does not have, it will not directly tell that it does not know, but will only make up an answer and provide false facts.
The reason for this is mainly due to the limitations of technology. At present, generative AI still has certain limitations in logical reasoning, reliability, robustness, accuracy, and security. [6] First, generative AI is essentially using machine learning to understand natural language, and the way it understands and the decision decisions are given by statistical probability, in other words, it is better at making statistical decisions than accurate logical reasoning, such as for mathematics or first-order logic, which often gives wrong answers. Secondly, although the knowledge produced by generative AI comes from large model language databases, these databases are only linked to existing network data, and the integrated knowledge is the knowledge that can be traced on the Internet at this stage, which is not updated in real time, and the feedback knowledge is limited in time. In addition, the network database itself contains a large amount of inaccurate data, which can also affect the accuracy of output decisions. At the same time, generative AI is mostly based on large language models, and the huge data volume exacerbates the data noise, which can cause the model to forcibly fabricate concepts that do not exist in the training data, and generate decisions that seem correct but are actually inaccurate.
2) "Algorithm transparency" is not standardized.
Since news production entered the algorithmic turn, the opacity of algorithms and algorithmic bias brought about by algorithmic black boxes have weakened the public's trust in news, and the academic community has called on intelligent news production to abide by the ethics of algorithmic transparency. Algorithmic transparency is a mechanism that attempts to elucidate the disclosure of information by algorithms. For end users, algorithmic transparency in human-robot collaborative news production means disclosing the existence of algorithms.
At present, most of the most ** algorithms lack transparency awareness, and in automated content generation, they either do not disclose the existence of algorithms, or do not know how to disclose them. In 2022, the American technology **cnet quietly published dozens of AI-generated reports, which were strongly criticized by the news community and users for not disclosing the author's algorithmic identity,[7] not only for spreading false facts, but also for the opacity of the algorithmic authorship. Due to the habit of trust, the public has not yet established a good trust relationship with generative AI technology. In the U.S., the survey found that 78% of people don't think it's a good thing to rely on software to write articles. [8]
3) There is a dispute over the applicable boundary.
The computing power of generative AI can improve the efficiency of news gathering and editing, but not all types of news are suitable for machine automation. For some ethical stories involving sensitive data, any small mistake can lead to systemic risk and damage to reputation.
Sensitive data involved in news gathering and editing includes confidential documents, as well as confidential information, trade secrets or personal data of news**, employees, customers or business partners, or other natural persons, and if these data are entered into the AI system, there is a risk of information leakage and infringement of the privacy rights of others.
Whether sensitive human topics, such as obituaries and serious issues with strong humanistic care, can be reported with generative AI is still controversial. ChatGPT's act of compiling the information of an incident at a Michigan college was met with students, who believed that the use of artificial intelligence to compile and send information about human tragedy was disrespectful to life and unethical. [9] In the era of intelligent media of human-machine collaboration, human beings do not have a high acceptance of information from automated production. In the current cognition of the public, they will still evaluate automated behavior from an anthropocentric perspective, especially in terms of human ethics, the public will generally think that the information produced by machines lacks "aura", and they will prefer un-mediated reports related to life issues.
Coping strategies for human-robot collaborative news production.
In the face of existing problems, it is not only necessary to correctly understand the content production mechanism and limitations of generative AI, but also to establish a series of specifications to optimize the human-machine collaborative production process.
1) Optimize the editing process.
In fact, in the process of news gathering and writing, reporters will also provide inaccurate information, and the solution is to set up editors to check whether the report is true and accurate. In view of the inherent defects of technology, what people can do is still to do a good job of controlling technical risks, pay attention to the importance of people in human-machine collaborative production, and optimize the process of human-machine collaborative news production.
The first key to optimizing the editing process is to improve the media literacy of the first people, so that they can correctly understand the content production mechanism and limitations of generative AI. AI systems on different platforms will have different functions and performance in content generation, so it is necessary to learn relevant knowledge in advance. In addition, gatekeeping the facts provided by intelligent systems and their reporting is a necessary process to reduce the generation of fake news, including verifying the information provided by generative AI, or setting up a dedicated AI editor to automate news production. Faced with the reality of large-scale generation of automated content, the Financial Times has set up AI editors to reduce the risk of automated content; The Newsroom requires that all AI-generated content be checked and revised by journalists if necessary. In its guide on human supervision and participation in generative AI production, the Dutch news agency ANP proposes a "human-machine-human" process collaboration model, stating that when using AI or similar technologies, it is necessary to "support the final editing, provided that the final inspection is carried out by a human after the fact", i.e. both ** and decision-making are supervised by humans. [10]
2) Disclosure of algorithmic information.
Since the release of ChatGPT, a number of ** have solicited users' opinions on AI-generated content in the user community, taking the "self**" platform Medium community as an example, many users mentioned the need for algorithm transparency and disclosure. Therefore, the establishment of transparent norms and the disclosure of algorithmic information are important ethics in the content generation stage of human-computer collaboration. There are two types of transparency policies that can be implemented for end users.
1.Inform the algorithm information. For example, Medium's first "Transparency, Disclosure, and Publication-Level Guidelines" policy for the use of AI language tools states that when creating any part of a submission using generative AI tools, creators must cite it as if they were any other **. [11]
2.Inform people of participation information, and classify and hierarchical management of automated participation content. For example, for purely automated content, it is actively identified as high-risk information, and it is indicated that it is automated production, which has not been manually reviewed, and there is a risk of inaccurate facts; For human-robot collaboration generated content, mark it as medium and low risk, inform human participation information, etc.
3) Standardize the applicable boundaries of the machine.
In the current human-robot collaborative news production, the first person should have the boundary awareness of the machine's participation in content generation, and coordinate the rights and interests of news parties, users, and users in the human-machine collaborative production.
Based on this, more intelligent newsrooms mostly limit generative AI to the production of non-sensitive news. Taking The Newsroom, which focuses on intelligent news production, as an example, the content on the platform is mainly the processing of existing reports, and in order to prevent risks, the platform also manages the risk according to the content theme, and does corresponding manual review for different risk categories. If an issue that is not suitable for machine content production is considered to be a high risk, it is forbidden to upload the relevant topic data to the system.
Conclusion. The application of generative AI in news production is not to let machines replace journalists to make decisions, but to let reporters define the angle, style, and value orientation of reporting, with the aim of improving the quality of content produced by human-machine collaboration. To this end, academia and the industry need to further explore more practical digital journalism ethics, standardize the use of technology, and prevent the risk of news algorithms.
This paper is the interim result of the 2021 Chongqing Municipal Education Commission Social Science Planning Project "Platform Expansion and Strategic Innovation of Chongqing Government Affairs Communication in the Era of Intelligent Media" (No. 21SKGH243).
References: 1]Gozalo-Brizuela R, Garrido-Merchan e C chatgpt is not all you need. a state of the art review of large generative ai models[j].
2]marina adami.is chatgpt a threat or an opportunity for journalism? five ai experts weigh in[eb/ol].（2023-03-23）[2023-04-10].
3]nick diakopoulos.what could chatgpt do for news production?[eb/ol].（2023-02-10）[2023-04-10].
4] AIGC has changed journalism [EB OL].WeChat***Tencent**Research Institute", 2023-08-28
5]johannes gorset.schibsted experiments with ai, finds 37+ ways it can help[eb/ol].（2023-03-27）[2023-05-01].
6] zhou j, ke p, qiu x, et al. chatgpt: potential, prospects, and limitations[j]. frontiers of information technology & electronic engineering, 2023: 1-6.
7]paul farhi.a news site used ai to write articles, it was a journalistic disaster[eb/ol].(2023-01-17)[2023-04-10].
8]artificial intelligence use prompts concerns[eb/ol].（2023-02-15）[2023-05-01].
9]jennifer korn.vanderbilt university apologizes for using chatgpt to write mass-shooting email[eb/ol].（2023-02-22）[2023-04-01].
10]hannes cools, nicholas diakopolulos.writing guidelines for the role of ai in your newsroom? here are some, er, guidelines for that[eb/ol].（2023-07-11）[2023-09-18].
11]scott lamb.how we’re approaching ai-generated writing on medium[eb/ol].(2023-01-27)[2023-04-10].
This article refers to the citation format:
Chen Sha. Application of Generative AI in News Production, Real-world Problems and Their Responses[J].Young Journalists, 2023(19): 57-59