Human trafficking is a problem of global concern, with estimates suggesting that approximately 25 million men, women and children are trapped in these situations of severe labour exploitation worldwide. The eradication of these exploitative situations has been listed as one of the targets in the 2030 Sustainable Development Goals, yet much progress is needed to meet this goal of decent work for all. Identifying victims is a critical first step in supporting workers to exit these work environments. This paper describes the design and evaluation of Apprise, an expert system we developed and have been using in Asia Pacific to support stakeholders to identify victims of labour exploitation and human trafficking. It draws four key recommendations for using expert systems in humanitarian relief and development: a comprehensive understanding of who defines the 'good' in AI for 'social good' interventions; a consideration of human rights implications of AI systems; broad participation by stakeholders in the design of the system; and an understanding of the way that digital technology will be used to amplify human intent.
Over the last few years, the world has been shaped by major
migration and displacement events. While almost two thirds
of the 272 million international migrants migrated for work
purposes, significant numbers were displaced due to conflict
(e.g. within and from Syria, Yemen, and South Sudan),
extreme violence (e.g. Rohingyans fleeing to Bangladesh),
economic and political instability (e.g. in Venezuela), and
climate change (e.g. China and USA)
In popular media, the phrases ‘labour exploitation’,
‘forced labour’ and ‘human trafficking’ are used
interchangeably, so a clarification of these terms is essential
at this point. In this work we draw on Skȓivánková’s
continuum of exploitation (Figure 1) (2010) that defines
‘decent work’ and ‘forced labour’ as two ends of a
continuum. The International Labour Organization defines
forced labour as “all work or service which is exacted from
any person under the threat of a penalty and for which the
person has not offered himself or herself voluntarily”
These exploitative situations include “compromising conditions that deny fundamental principles and rights at work, put at risk the lives, health, freedom, human dignity and security of workers or keep households in conditions of poverty” (ILO 2015, 1). Using this continuum, we can see human trafficking as a process, consisting of a series of exploitative acts that move the worker towards a situation of forced labour.
Global estimates suggest that approximately 25 million
people are trapped in situations of forced labour or human
trafficking
Although the eradication of human trafficking and forced
labour is listed as Target 8.7 of the Sustainable Development
Goals, much progress is still required to meet this lofty goal,
specifically in the development of effective responses for
trafficking prevention and assistance to victims
This paper presents a case study of the use of Apprise, an AI-based screening tool to support FLRs to assess labour conditions and vulnerability in precarious work environments. It draws from our 4 years’ experience working with a wide variety of FLRs including: private auditors in supply chains across Asia-Pacific; Royal Thai Navy and Ministry of Labour Inspectors in fishing sector in Thailand; and NGOs, CSOs, and IGOs in fishing and sex work sector in Thailand. From this experience, the paper draws key recommendations for using AI more broadly in humanitarian relief and development.
This section describes Apprise, a screening tool that we
developed to support FLRs to proactively screening workers
for indications of labour exploitation and human trafficking.
This screening tool supports workers to raise concerns on
the illegal, undignified, or exploitative employment that
they are trapped in. If they choose to exit their current work
situation, workers then can have access to help, remediation,
and justice. As a complete description of this research is
outside the scope (and page limit) of this paper, we
summarise the key points here that are needed to frame the
subsequent discussion in the remainder of the paper,
referring interested readers to
We undertook a series of consultations with a broad range of direct (those who are intended to directly make use of a technology) and indirect (those who will be impacted by a technology) stakeholders, including: survivors of human trafficking; local and regional NGOs; inspectors from Department of Special Investigations; and intergovernmental organizations with mandates in migration and / or trafficking. The first stakeholder consultation identified the initial screening phase of victim identification as a critical first step in the process, where technology could potentially play a role. FLRs identified key problems that they face in initial screening as: language barriers / communication; lack of training or common understanding of the indicators of human trafficking; screenings being undertaken in uncontrolled environments, where victims’ privacy could not be guaranteed; and a lack of trust between all parties involved (workers, FLRs and translators - where available). On the latter of these points, stakeholders noted that migrant workers were often most susceptible to labour exploitation, and translators were required to be present in the initial screening process. Scheduling and resourcing problems were widely recognized in the field, as a FLR would not know which language workers would speak, so could not guarantee that they would have a translator with them who could speak the required languages. Stakeholders also noted that they could not verify the accuracy of translations in the field and spoke frequently of cases of translators who had been bribed by exploiters to mistranslate workers responses. Based on this consultation we identified the potential for a mobile-phone based, multilingual, expert system, to support FLRs and workers to communicate during the initial screening phase of victim identification.
From this, we developed Apprise, an app that is installed on the FLRs phone, but ultimately is a tool in the potential victims’ hands. When combined with a set of headphones, it offers workers privacy to answer audio questions that are played in their language. It contains screening questionnaires for different sectors of work (Figure 2(a)), with audio questions translated into the most common languages among migrant workers communities in that sector / region (selected by the worker, Figure 2(b)). Once the worker has selected their language, the app begins by playing an introductory video, describing the purpose of the interview, introducing the FLR, and demonstrating how to use the app (Figured 2(c)). It then asks for consent to continue to ask the worker a list of questions. Each list contains a mixture of positively and negatively worded yes/no questions (Figure 2(d)). Branching is included in the question list, enabling different information to be collected based on responses to previous questions. The last question in each list asks the worker if they would like help to exit their work situation. After completing these questions, a vulnerability calculation is performed, and an audio is played to inform the worker of the vulnerability of their situation. They are also given a chance to reconsider if they would like help to exit their work situation based on the outcome of the screening. At this point, the phone is returned to the FLR and the key vulnerabilities are summarised to inform FLRs follow-up investigations.
To assess the vulnerability of the situation, each question
has a predefined weighting and categorical information,
based on four principle dimensions of exploitation: unfree
recruitment, work and life under duress, impossibility to
leave, coercion / penalty or menace of penalty. Each
question is also aligned to one of the eleven indicators of
forced labour (as defined by ILO (2012b)). The knowledge
base takes into consideration the age of the respondent, to
pay particular attention to child labour (adapted from
Once interview responses have been uploaded (when network connectivity is available), FLRs and key personnel within their organization can access and analyze anonymous interview responses. These interview logs in themself can be a tool for policymakers to develop a nuanced understanding of the sector specific and evolving practices of exploiters. Reported practices of exploitation can be analysed by sector of work, location, language of interview and time, to support the development of evidence basedpolicy, which in turn can support the prevention of future trafficking situations by ensuring effective law enforcement and protection practices.
Apprise has been used in the field by frontline responders in Thailand (fishing, seafood processing, sex work), and across Asia-Pacific within supply chains of multinational corporations (manufacturing) since March 2018. This section draws together our findings, presenting key lessons that we learned that would be applicable to consider across many different humanitarian relief settings.
Like many ‘wicked problems’, labour exploitation and human trafficking are complex stories involving many characters, few of whom can be stereotyped as ‘all good’ or ‘all bad’. Different factors push and pull different actors to undertake actions that they would otherwise not consider under different circumstances. Care must be paid when designing systems to support triage of cases, to understand if this support is in the best interests of the recipients of the service.
This point can be most easily described with reference to
our research with sex workers in Thailand. We initially
began our study working with state actors, being asked to
support them to build their screening capacity. We soon
realized that due to widely reported corruption within
traditional justice channels
Building on this and stemming from the final question in each question list which asks if workers would like help to leave their work situation, sex worker organizations emphasized that technologies such as Apprise could only be useful if they enhanced their agency. Ultimately, they were interested in using Apprise if it could support them to enhance their understanding of the vulnerability of their work situation, rather than to choose for them what a ‘good’ outcome would be for someone who faces exploitation. One sex work described this further, saying that if the application indicates that she is in a vulnerable situation, but she is willing to stay, “would you respect my decision?”. This highlights the important concept of supporting the workers’ own agency, enabling them to override the suggestion that is provided by the expert system.
This example has illustrated several very different ideas of the ‘social good’ that triage tools could support within this one sector. This explicit understanding of the social good that our systems seek to support is fundamental to the use of AI, especially within humanitarian relief and development.
After calls in the previous section to consider the very different understandings of ‘social good’ that exist within a particular context, the question becomes how to determine the impact that a system may have on stakeholders.
In previous work we have called for the use of
international legal frameworks such as the International Bill
of Human Rights to assess the potential human rights
impacts of AI systems prior to rollout (and preferably
development) of a system
Within our work, we undertook a year-long consultative process to understand what exploitation looked like in different sectors. In formal work, labour law can be used as a basis for what is legal, and then by extension, what is exploitative and illegal. However, we also understood that law is slow moving and does not reflect the rapid changes and current practices of exploitation that can be experienced in the field. In our work in the fishing sector, we met with state actors, but also IGOs, CSOs, NGOs, survivors of exploitation, and vulnerable workers, to understand how people are currently being exploited. These findings informed what we referred to as proxy indicators for sectorspecific laws, which we then codified to form the expertise (or knowledge base) within the expert system.
In informal work, we again used these broad consultative processes to understand what exploitation looks like to workers within the sector. While not codified in the same way as formal work, there is often an understanding within the community of what acceptable labour conditions are. When using AI in humanitarian relief and development, we recommend this broad perspective and recognition of expertise.
Our final recommendation in this paper draws from Agre’s
assertion that digital technology in itself is neither a positive
or negative force, but at best will be used to amplify the
intentions of those that use it
Drawing from our experience with auditors in supply chains, we understand that we work with brands and corporate social responsibility experts that are inherently motivated to make positive change in the labour conditions of workers in their supply chains. In these cases, they aim to use Apprise to identify any exploitative conditions in order to address then and provide remediation. In other cases, the same technology could be used to provide inaccurate information about conditions of work. Employers could coerce workers to provide inaccurate responses to screening questions. Or workers themselves could try to use the interview to make false reports about the conditions of their work.
Within humanitarian relief and development contexts, it is important to remember that digital technology (and in this case, specifically AI systems) will not bring about systemic change in and of themselves. Any initiative must be supporting by robust transparency and accountability structures, to ensure that it is being used for the purposes for which it is intended. In the case of expert screening tools such as Apprise, this could include standard operating procedures detailing actions to be taken when a case is identified as showing vulnerability, as well as broader accountability to external parties.
This paper has presented insights derived from our four year engagement in Thailand and other countries within Asia Pacific, innovating and inventing technologies to support FLRs to screen workers in vulnerable situations. It described our system, Apprise that we developed and that has been used in the field to identify exploitation with fishing, seafood processing, manufacturing, and sex work. The paper draws key recommendations for researchers, academics, and program officers who are interested in using AI within humanitarian relief and development. First, by reflecting on the complex nature of problems that we see in humanitarian relief and development, the paper suggests that stakeholders take a step backwards, to reflect on who defines what ‘good’ is in AI for social good. Next, the paper suggests using key international frameworks such as International Bill of Human Rights, to identify the impact that an AI system would have in complex environments. The paper then suggests that for expert systems, a broad consultative process could be used when codifying the knowledge base, recognizing expertise from diverse actors. Finally, recognizing that technology is neither positive or negative, but can be used to amplify user’s original intent, this paper calls for comprehensive accountability structures as well as standard operating procedures to inform FLRs follow-up actions.