AI’s Implications for Governance: A Closer Look at Plausible Scenarios

• June 8, 2023
• Artificial Intelligence

Building upon the previous two conceptual framings of governance, in this article we will discuss how AI technologies (using GPT as a focused example¹) can be used in specific scenarios to enhance government’s performance as well as facilitating open collaborations among various stakeholders in governing our society.

AI for High-Capacity Governments

The emergence of ChatGPT, as well as other GPT-backed applications, opens up the possibility of making government operations even more user-friendly for both citizens and government employees. Below we outline five different ways that these applications could be employed in service of achieving government goals more efficiently.  

For example, one of the frequently mentioned scenarios since the launch of ChatGPT is the idea that every government employee could now be equipped with an intelligent assistant. The utilization of AI tools in automating repetitive tasks can speed up transactions and provide more accurate assessments of policy outcomes, thereby reducing the workload of government employees. However, beyond that, there are other potential scenarios in which we may find GPT-backed tools useful.

1. Conduct sentiment analysis for textual data. 

One exciting potential of AI is the ability to automate sentiment analysis, especially for detecting biases or inappropriate language uses in reports, filings, publications, or records processed by governmental agencies. There have already been experiments utilizing early GPT models to detect hate speech (see Fanton et al., 2021; Tekiroglu et al., 2020; Zhou et al., 2021), and the detection accuracy seems to increase significantly as the GPT model evolves. For instance, with specified instructions, GPT-3 is able to detect hate speech with over 75% accuracy (Chiu et al., 2022). The latest GPT-4 model, and any subsequent versions, are likely to further improve detection accuracy and reduce latency when processing large volumes of textual data. 

Not only will GPT’s analysis capacity contribute to fighting against verbal offense and inappropriate language, but it also proposes insights derived from human responses for policy-making. For instance, recent research on Human-Computer Interaction (HCI) has demonstrated ChatGPT’s capability of summarizing sentiments precisely from unorganized human responses (Tabone & de Winter, 2023). A similar practice example is from Malta’s national AI strategic plan, which explains how governmental agencies used AI models to discover industry trends and users’ sentiments for future tourism planning² (Malta AI Strategic Plan, 2019).

2. Improve discipline training for law enforcement.

An extended use of AI’s analysis capacity is to conduct comprehensive checks for transcripts from body camera videos.³ AI makes it possible to provide rapid language analysis for audio recordings, reporting positive and efficient interactions for rewards while flagging inappropriate and abusive languages right away. In addition, the analyzed results might provide critical insights for designing training programs and help with deciding whether a training program is required for a specific individual. There is no doubt that AI-backed body camera analytics may control the intentional use of profanity, racial slurs, or insults. In the meantime, it also remains in question how body camera analytics may reshape the process of law enforcement. The accuracy of language analysis, especially when the conversation takes place in usually complicated situations, might still be a challenging task since the system involves hardwares, algorithms, data storage and transmission. Inaccurate language analysis or overly strict criteria could easily result in unnecessary problems for law enforcement staff. It is worth conducting close evaluations for pilot projects that have been adopting an AI-backed body camera analytics system.

3. Detecting foodborne public health risks.

Maintaining public health is one of the foremost charges of the government. Yet curtailing foodborne diseases is traditionally difficult, as many cases are unreported and therefore not documented in the databases managed by governmental authorities. Crowdsourcing foodborne illnesses from social media has been important in public health research (see Sadilek et al., 2017; Tao et al., 2021),  but GPT tools may contribute to reducing food risks in the following ways: first, by facilitating the crowdsourcing process from text data available online; second, by verifying information included in the filed complaint (e.g., description accuracy, business name, and location, food image, etc.), and third reminding the general public, including both consumers and business owners from the food industry, of the latest information related to public health risks, disease spread, and policy changes or updates.

4. Verifying applications for public subsidies.

With necessary instructions, AI could be the rudimentary reviewer for almost all subsidy applications by filtering some bogus applications with illogical information. This function might be particularly useful during a crisis, such as when many small businesses applied for public grants or loans during the COVID-19 pandemic while the government was also in a staff shortage. For instance, an instructed AI tool could prevent the massive overpayment of flawed New York City grant claims (see Cowley, 2021).

Additionally, many people and families who are eligible for multiple benefits are not enrolled in all programs they could be, therefore leaving money and other forms of assistance on the table. By utilizing a similar filtering model as above, AI tools could flag applicants who are eligible for multiple benefits and even submit applications to those programs.

5. Assisting call centers to respond to service requests.

It is easy to imagine a powerful chatbot like the ChatGPT may provide useful assistance to call centers by responding to some frequently asked questions. Of course, the challenge remains in ensuring factually accurate answers rather than just fulfilling the responsibility of generating responses. On the other hand, an instructed AI tool may also help match individual requests with the correct call centers, for instance, separating calls for 911 and 311. That may significantly optimize the workflow in call centers. However, if it is an emergency situation, it might still be a challenge that AI needs to ensure no-latency communication when processing callers’ requests.

6. Improving data extraction and processing through drone imagery.

Considering imagery can be a very reliable representation of reality, numerous tech companies have made substantial investments in research related to imagery processing. While it is risky to assert that recent advancements in AI have completely revolutionized imagery processing methods, there are notable user-friendly applications emerging in the realm of public management and urban planning. For instance, Esri, a leading company in geographic information software development, explicitly showcases how its product suite, ArcGIS Imagery, incorporates AI to streamline the processing of drone imagery. The implementation of AI deep learning models demonstrates impressive capabilities in identifying and categorizing specific features, such as building footprints, parked cars, and cracked pavement. In theory, this automation function could serve as a convenient tool for local governments to swiftly conduct surveys using drone imagery.

AI for Engaged Governing Processes

In contrast to earlier chatbots that focused on improving communication between citizens and the government and reducing the administrative workload on public organizations, GPT presents possibilities to reshape the relationships between non-governmental actors and public actors. Fundamentally, GPT may provide unparalleled assistance to non-government individuals seeking information. 

1. Improving information accessibility and algorithmic legibility. 

GPT-backed tools possess a notable characteristic of presenting information in a user-friendly manner, making them valuable learning aids for intricate, complex, or technical subjects. For instance, while the US census data is publicly available, it can be difficult to select, process, and visualize outputs. The recent Census GPT addresses these challenges by allowing individuals to search the census database using natural language, such as asking: “locate areas in a specific city with the most elevated average household income and exhibit the proportion of each racial group in that location.” AI tools are adjusting how people engage in policy, giving them the power to gain knowledge on complex datasets even if they do not have expertise in data processing. 

Despite the increased availability of digital applications that enhance the information legibility, few of these tools or webpages disclose the data processing methods, which typically require advanced technical knowledge. AI tools make it possible to quickly translate coding languages to conversational language for everyone to stay informed about the methods of generating and processing data.⁴ In either situation, AI tools may offer significant assistance for citizens to overcome technical difficulties in accessing critical information, enhancing informational accessibility and openness to the wide public. 

2. Assistance in seeking and understanding professionally demanding information. 

GPT’s ability to lower barriers to entry or understanding is impressive, particularly when tasked with lowering barriers to acquiring professional information. There are multiple AI tools that focus on analyzing and interpreting various types of textual documents to help citizens stay informed about important information that might be overlooked due to the use of formal and technical languages.⁵ One unique case is Clarity, a tool that aggregates mainstream news and analyzes how these news sources are divided based on the left, center, and right leaning perspectives using ratings and attention scores. The tool, while not aiming at eliminating biases, may provide crucial assistance for citizens to gain a fuller and clearer picture of unspoken biases in daily news.

A rather alarming scenario, yet one worth assessing, is GPT’s performance in addressing legal questions. Through a series of experimental exams, Choi et al. (2023) conclude that ChatGPT-3 (without any additional modifications related to the legal exam) has demonstrated its capability of earning a JD degree from a selective law school. That being said, it might not be difficult to consider GPT as a legal consultant for some basic-level advice. However, the potential risk could also be significant. How can those without legal research and reasoning skills evaluate the reliability of GPT’s advice? That is not only a question for individual users but also for regulatory agencies.

3. Navigator for bureaucratic procedures. 

One may also imagine a navigator-like tool that walks less tech-savvy citizens through  applications for specific public services or supports, such as housing grants, education grants, and medical care support qualification processes. Such a tool might be handy for aging adults, as well as low-income and English-language learning communities. However, it requires extra caution when designing such a tool to ensure its role as a “navigator” who assists in drafting applications rather than a “ghostwriter” for applications. 

The recent increase of GPT and its extended applications proved some of the pre-GPT expectations for AI, such as paving the road for optimized government-citizen interaction and communication (Androutsopoulou et al., 2019), as well as fostering citizen trust through improving engagement and transparency in governance (Dwivedi et al., 2019). The proliferation of GPT-backed applications suggests an empowerment process that may increase citizens' sense of ownership of their city. The underlying driver, unsurprisingly, is AI’s capacity to inform citizens, raising their awareness of many subjects they might have found difficult to acknowledge or understand and therefore producing more engagement and resident buy-in.

  Conclusion

The rapid development of the GPT model and related applications have even made speculative discussions challenging as it is difficult to predict what is emerging or what new tool will surface the next day. Every single innovative change might lead to a series of innovations, and all these incidents may aggregate, resulting in large-scale social, economic, and political impacts. In other words, not only is GPT itself already an innovative technology, but also that it reinforces the innovation cycle in almost all aspects of today’s society. GPT, along with its successors and competitors, are driving a process that increasingly democratizes public and private organizations’ capacity for innovation and citizens’ skills to explore and understand their position, and rights,in the society they belong to. 

Meanwhile, AI poses as much excitement as concern in today’s society. We can reasonably expect  to see AI initiate some foundational changes to our existing governing mechanisms. While the previous discussion focuses on positive, forward-looking scenarios, they all entail perils that could threaten public interest and civil rights. To counter these would require the government to adapt and regulate, quickly, with an active and agile attitude. Learning about and understanding AI tools is a prerequisite to responding smartly to emerging technical challenges and opportunities in this fast-changing world.

¹  We are fully aware that there are various ChatGPT alternatives, each of which might specialize in different fields. In this short article, we take the GPT model as well as its related tools as the representatives of the still evolving AI technologies.

² Although Malta did not use a GPT model, it stands as an early example of conducting sentiment analysis through AI models.

3 There is an early stage company, Truleo, which focuses on body camera analytics leveraging AI algorithms. For more information: https://www.axios.com/2023/01/30/police-tyre-nichols-bodycam-footage.

⁴ Most of the leading AI tools to explain coding are designed for professionals (e.g., Denigma AI), but it is fairly straightforward for anyone to use the tools to interpret lengthy codes and understand how data is being processed and presented.

5 There is an abundance of AI tools related to document processing, such as document analysis (e.g., ExplainThis, ELI5, Distillr, etc.), legal document review (e.g., Magon, Latch, etc.), invoice generation (e.g., Tonkean Invoices), patent application (e.g., PatentPal), and so forth. However, most of these tools are at the early stage and will require cautious and comprehensive testing and evaluation in practice.

  Bibliography

Androutsopoulou, A., Karacapilidis, N., Loukis, E., Charalabidis, Y., 2019. Transforming the communication between citizens and government through AI-guided chatbots. Government Information Quarterly 36, 358–367. https://doi.org/10.1016/j.giq.2018.10.001

Chiu, K.-L., Collins, A., Alexander, R., 2022. Detecting Hate Speech with GPT-3.

Choi, J.H., Hickman, K.E., Monahan, A., Schwarcz, D., 2023. ChatGPT Goes to Law School. https://doi.org/10.2139/ssrn.4335905

Cowley, S., 2021. S.B.A. Overpaid $4.5 Billion on ‘Illogical’ Small Business Grant Claims. The New York Times.

Dwivedi, Y.K., Hughes, L., Ismagilova, E., Aarts, G., Coombs, C., Crick, T., Duan, Y., Dwivedi, R., Edwards, J., Eirug, A., Galanos, V., Ilavarasan, P.V., Janssen, M., Jones, P., Kar, A.K., Kizgin, H., Kronemann, B., Lal, B., Lucini, B., Medaglia, R., Le Meunier-FitzHugh, K., Le Meunier-FitzHugh, L.C., Misra, S., Mogaji, E., Sharma, S.K., Singh, J.B., Raghavan, V., Raman, R., Rana, N.P., Samothrakis, S., Spencer, J., Tamilmani, K., Tubadji, A., Walton, P., Williams, M.D., 2021. Artificial Intelligence (AI): Multidisciplinary perspectives on emerging challenges, opportunities, and agenda for research, practice and policy. International Journal of Information Management 57, 101994. https://doi.org/10.1016/j.ijinfomgt.2019.08.002

Fanton, M., Bonaldi, H., Tekiroglu, S.S., Guerini, M., 2021. Human-in-the-Loop for Data Collection: a Multi-Target Counter Narrative Dataset to Fight Online Hate Speech, in: Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). pp. 3226–3240. https://doi.org/10.18653/v1/2021.acl-long.250

Malta AI Strategic Plan, 2019. Malta the ultimate AI Launchpad: a strategy and vision for Artificial Intelligence in Malta 2030. Malta, Parliamentary Secretary for Financial Services, Digital Economy and Innovation.

Sadilek, A., Kautz, H., DiPrete, L., Labus, B., Portman, E., Teitel, J., Silenzio, V., 2017. Deploying nEmesis: Preventing Foodborne Illness by Data Mining Social Media. AI Magazine 38, 37–48. https://doi.org/10.1609/aimag.v38i1.2711

Sharma, G.D., Yadav, A., Chopra, R., 2020. Artificial intelligence and effective governance: A review, critique and research agenda. Sustainable Futures 2, 100004. https://doi.org/10.1016/j.sftr.2019.100004

Tabone, W., de Winter, J., 2023. Using ChatGPT for Human–Computer Interaction Research: A Primer.

Tao, D., Zhang, D., Hu, R., Rundensteiner, E., Feng, H., 2021. Crowdsourcing and machine learning approaches for extracting entities indicating potential foodborne outbreaks from social media. Sci Rep 11, 21678. https://doi.org/10.1038/s41598-021-00766-w

Tekiroglu, S.S., Chung, Y.-L., Guerini, M., 2020. Generating Counter Narratives against Online Hate Speech: Data and Strategies. https://doi.org/10.48550/arXiv.2004.04216

Zhou, X., Yong, Y., Fan, X., Ren, G., Song, Y., Diao, Y., Yang, L., Lin, H., 2021. Hate Speech Detection Based on Sentiment Knowledge Sharing, in: Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). Presented at the ACL-IJCNLP 2021, Association for Computational Linguistics, Online, pp. 7158–7166. https://doi.org/10.18653/v1/2021.acl-long.556