This image was partially generated using AI. 30 January 2024
The Italian state’s uneven adoption of AI for its own purposes is a useful case study in that it surfaces many of the challenges faced by democratic institutions across Europe. Generative AI portends a massive cultural shift, including in politics, and we can’t forget that it’s rolling out amidst the all-too-common backdrop of severe cuts to public administration staffing.
Bring in the Bots
Like many others, the Italian government led with the ‘customer service chatbot’ approach in 2019. It’s a relatively well-trodden application of AI – the Civic Tech Field Guide indexes nearly 60 other examples. From a cost perspective, customer service chatbots can reduce costly phone support services while reducing wait times.
The Italian Ministry of Labour and Social Policies launched its chatbot to help provide information on the nation’s guaranteed minimum income program (the Reddito di cittadinanza). The Ministry considered the chatbot a success, and later extended the tool to communicate with citizens inquiring about a transportation benefits program designed for low-income communities.
Starting the roll-out of AI citizen service within programs meant to benefit disadvantaged communities is a threat to equity. And routing citizens to AI conversational agents isn’t without risk. When Austria’s public employment agency introduced a ChatGPT-powered bot, a Der Standard investigation found it directed women toward lower-paid career options.
AI vs. the 1%
“Using AI for chatbots is an underestimation of what AI can do,” says Marco Bani, Tech Policy Analyst in the Italian Senate. With experience at both the Agenzia per l’Italia digitale (the government agency responsible for the digital transformation of the Italian government) and managing public policy at Amazon Web Services, he’s a vocal advocate for integrating AI into democratic processes.
Bani points to Anonimometro as a better illustration of the power of AI for internal government work. In a country that’s long struggled with high tax evasion rates, the new tool offers the nation’s tax fraud unit a rare advantage in their fight. The l’Agenzia delle Entrate and the Guardia di Finanza can automate analysing proprietary databases to cross-reference citizens’ income declarations with their other assets, such as bank accounts, asset management companies, insurance policies, and pleasurecraft.
The algorithmic approach replaces personally identifiable data with generated pseudonymic data, and runs the cross-database checks to flag potential fraud for the human members of the team to follow up. This approach masks individual citizens until there’s some evidence to warrant a closer look, earning the stamp of approval from The Italian Data Protection Authority. People always judge which cases warrant further investigation, a ‘human in the loop’ approach required by legislation. Anonimometro is contributing to significant improvements in Italy’s tax evasion rate, helping break the vicious cycle between lost tax revenue and declining government services.
Need for caution
Rolling out an AI-powered methodology on the most wealthy (who can more easily afford good lawyers to defend themselves) might be a better place to start than with disadvantaged communities. When the Dutch tax office introduced its algorithmic fraud-spotting program, their machine learning approach unjustly profiled tens of thousands of innocent people. Their only crime was representing traits that the tax system had determined was high risk of committing fraud, including having dual nationality and being low-income.
This is the example of government AI where Bani sees the most need for caution. “If they can see our habits, they can also profile citizens better. So they must be more transparent on algorithms and how they are going to process and control our data.”
Aqueducts Get an Upgrade
AI innovation is more complicated at Italy’s local level of government. With the global tech giants outbidding one another for AI talent, city and regional institutions rarely have the resources necessary to do the same. One exception is a project in Rome that combines low-cost sensors (remember the Internet of Things?) with machine learning to monitor the city’s water infrastructure for leaks. It turns out that the flow of water has a particular sound, and these sensors can be trained to detect when and where a pipe is leaking. This spares the municipal government from having to tear up the streets and disrupt traffic every time there’s a problem.
Situating AI projects in the context of government cutbacks
While government tech is often framed as innovation, much of the time it’s really about doing more with less. We must consider these government AI examples in the broader context of cuts to public services, and staffing levels in particular. The Italian civil service is still under a decade-old hiring freeze that requires four public servants to retire before a new one can be hired. Government AI advocates like Bani believe that in this context, “We have to try to use AI tools to improve productivity to be able to offer the same public services as in previous years.”
In the tax evasion and leaking pipes examples, it’s clear to see how AI can contribute to significantly better outcomes than traditional methods. But the chatbot example is trickier. First, they’ve been deployed to replace human support services for programs specifically aimed at low-income citizens who may be more comfortable with a phone than a chat widget, or have a complicated situation that warrants a conversation with a human.
No substitute for street-level bureaucrats
Anyone who’s attempted to solve a problem with, say, their telecom’s automated customer service channels knows there are finite limits to a pre-programmed chatbot’s utility. While they may help handle a large percentage of expected requests, it doesn’t take much to exceed their knowledgebase. It can be quite frustrating for a customer, or in this case, a citizen, to have to get past a bot regurgitating the answers to Frequently Asked Questions in order to reach someone who can truly help. For this precise reason, there’s already talk of requiring businesses to maintain at least one human-staffed customer support channel to solve problems chatbots cannot. The same rule should apply to government interfaces with the public, at the very least.
Michael Lipsky described the powerful role “street-level bureaucrats” have in aggregate. On a daily basis, they interpret their government’s policy at the immediate level, where it’s being applied to someone’s life. In many cases, the front-line bureaucrat is able to make an exception or otherwise adjust a too-rigid policy to span the chasm between policymakers and the reality of the person in front of them at the counter. What happens to the humane malleability of street-level bureaucrats as they are replaced by bots with limited agency and intelligence?
Staff experimentation
Over the previous three decades, the (still-ongoing) digital transformation of government has presented institutions with a choice: replicate what came before in a new medium (by simply digitising analog processes) or actually reimagine public services to make the most of the affordances provided by new technology.
In addition to lighthouse projects, individual government staff are experimenting with generative AI tools to become more efficient workers, often without formal permission or guidance. They’re employing the same consumer-level services as private sector workers, and try out new approaches in addition to their formal responsibilities. Even when individual experiments fail, the staffers’ firsthand experience using paradigm-shifting tech can help inform the government’s own adaptation to its evolution.
AI for the legislative process
Bani is one of those curious public servants who stays current with the rapidly-evolving generative AI ecosystem by writing about it and tinkering with it. He’s been working with consumer-grade generative AI tools to see if they might help policymakers draft better laws more efficiently.
“The most time-consuming part of drafting legislation is looking up all the references to previous laws,” he says. To see if Large Language Models could help expedite this work, Bani downloaded the Italian Parliament’s draft legislative proposals from the past ten years in the hopes of training a CustomGPT on them. (CustomGPT is ChatGPT’s new plugin product that allows anyone to create their own version of OpenAI’s flagship service).
But the output left something to be desired. “When I then tried to draft a law, the output was not useful at all. When you’re using ChatGPT to write something like an article, the final output might be rubbish but you might still get one usable line out of it. In the legislative draft proposal, it was just bad,” Bani laments.
LLMs as interns
“My next assignment is to draft a legislative proposal promoting quantum computing in Italy,” Bani says. “I tried to do that on ChatGPT two days ago and the output was terrible.” The model spit out general recommendations like a blog post might, with little imagination and none of the specifics a national policy will require. Language wasn’t the issue, as Italian is particularly well represented in ChatGPT’s training data. It’s the domain and legislative format expertise that’s missing.
For his next test, Bani’s going to try training the AI model on strong national quantum computing policies from the United States, Germany, and France to see if that improves the results. But his expectations are realistic. “I like the metaphor of LLMs as an intern,” Bani says. “You have infinite interns, but the quality of their work might not be good enough for every purpose.”
Legislative ChatGPT in action.
Will we see specialised AI models for government work?
It’s still an open question whether we’ll come to rely on a single general purpose AI model like ChatGPT, or specialised services tailor-made for our intended use case. Bani believes the latter will be necessary for government AI applications. “If the government continues failing to invest in AI applications, Italy is not going to provide services for the world in this field, but maybe the US, or Asian countries like Singapore, Korea, or Japan will, because they have a strong tech and public administration nexus.” As it stands now, the LLMs aren’t able to surface the right data in the right format. You can’t mix laws between countries, so specialized legislative LLMs might be required for this application.
Beyond Bani’s personal experiments, the Italy’s Chamber of Deputies has convened a series of meetings with the industry’s leading companies. They’ve also started to experiment with using AI to improve the legislative process, which Bani claims as a first amongst European parliaments. Together with research centers and the private sector, they’ve embarked on a 1-year experiment to develop an official AI tool for staff and members of the Italian Parliament.
Will the AI Act help drive government AI adoption?
In Bani’s estimation, no individual department or ministry in the Italian government is demonstrating the true potential of AI in the public sector. He doesn’t think we’ll see a thriving ecosystem of AI-fueled public services in the near future. The EU’s AI Act will require member states to institute AI offices, but Bani fears Italy will respond by creating a government body that checks the box on algorithmic regulation without advancing the country’s imagination of what the tech can do. He looks to the United States as the best current example, because industry can move faster with strong investment.
And what does the public think?
“We’re lagging behind,” Bani says. In general, the public isn’t exactly demanding for the government to bake AI into public services, and so politicians are not rushing to do so. Public awareness of AI is low, and where people are aware of it, “It’s a cultural war. AI is still magic for one person, worrying for another.”
