Officials at the Swedish Public Employment Service (Arbetsförmedlingen) started looking into the system after they noticed it was failing to generate letters to welfare claimants that had been expected. When they finished their review last year they found major shortcomings, with between 10% and 15% of the computer’s decisions likely to have been incorrect, SVT reported.
It is unclear whether it will be possible to identify and correct the erroneous decisions, and when exactly the problem started.
The government brought in the computer system to automate the process of checking that people receiving a certain type of unemployment benefit keep up their obligations – and issuing warnings or withholding their payments if they don’t. This was supposed to increase efficiency, but since last Autumn the system has been switched off and human bureaucrats brought back in to sort out the mess.
The agency did not inform the public about the problem, despite insiders knowing that thousands of decisions may have been wrong.
Adnan Habibija, Research Officer at the Swedish Trade Union Confederation (Landsorganisationen i Sverige), said: “It is unfortunate that a computer failure at the Public Employment Service has led to wrong payment decisions for many people. This may have led to uncertainty for many beneficiaries, whose economic situation depends on unemployment insurance payments.
“Mistakes can occur but we are surprised that the agency has waited so long time to inform beneficiaries, but also its own employees, the government and other stakeholders.”
In an, researchers Anne Kaun and Julia Velkova, who contributed to AlgorithmWatch’s recent , say the situation highlights the risks of a growing trend.
“Swedish public administration has begun to replace people with algorithms to decide on everything from welfare payments to child support and sickness benefit,” they write. “Citizens do not know if their decision has been taken by an algorithm or an officer.”
In the AlgorithmWatch report, Kaun and Velkova discuss the processing of social benefit applications in Trelleborg in the south of Sweden. The municipality uses machines to check applicants’ data with other government departments such as the tax office. It is one example of the automated decision-making that is being rolled out for a variety of purposes across Europe. The report highlighted the lack of oversight of automation that is being used to make judgments affecting millions of citizens.
Anposted on the Swedish employment service website in response to the SVT investigation says 500,000 automated decisions relating to warnings or sanctions were made by the flawed system.
Swedes taking part in certain government programmes for the unemployed are obliged to submit anevery month. This can be done by logging on to a website, through a smartphone app or by filling in a paper form. The agency said problems with the website and issues with registering manually submitted activity reports were among the factors leading to the computer wrongly issuing sanctions.
The employment service website posting said: “All persons who have been wrongly affected will have their cases corrected, provided they are in contact with the Employment Service.”
The same week as the problems with the algorithm became known to the public, the agency announced it had launched a second automated decision-making system. This new algorithm will decide who has the right to state-subsidised employment, which is offered to some long-term unemployed people and recent immigrants to the country. Within two years, the administration is expected to automate a further two services, SVT reported.
The news comes just three weeks after the employment service announced it was laying off up to 4,500 of its 13,500 employees. The service cited a budget cut of SEK 800 million (€75 million) between 2018 and 2019. Mikael Sjöberg, Director General, was quoted in asaying: “The renewal of the authority and the digitization we have carried out has laid a good foundation for continuing to deliver at a high level”.