To make regulatory decisions, policymakers need to define problems and priorities in a process called agenda setting (Baumgartner, 2015). Issues raised in the agenda setting phase inform the evidence to be collected from empirical research (Head, 2010; Pawson et al., 2011) to shape the space of policies envisioned (MacKillop & Downe, 2023; Schiff, 2024).

Artificial intelligence (AI), while long studied in the academic sphere, is, from the point of view of public policy, an emerging technology with as-yet-unknown risks. In the public consciousness, two vivid risks people envision from AI are existential threats to humanity (Cameron, 1984) and the risk of being replaced by machines (Autor, 2015; 2022). In the past few years, multiple groups have introduced AI incident trackers (Abercrombie et al., 2024; Hutiri et al., 2024; McGregor, 2021; OECD, 2024) and taxonomies (Arda, 2024; Cattell et al., 2024; Critch & Russell, 2023; Shelby et al., 2022; Weidinger, 2022) to analyse the potential harms and risks of AI.

In this paper, we focus on labour relations with AI (Arntz et al., 2016), sometimes called the ‘future of work’ (Laker, 2023). Specifically, we investigate how two popular AI incident trackers represent labour-related risks for AI in the workplace.

To determine how well AI incident trackers inform policymaking, we study how well the knowledge organization of different incident trackers reveals labour-related risks for AI in the workplace. We search for and analyse labour-related incidents in two AI incident trackers, the AI incidents monitor and the AI incident database, with a focus on how AI is impacting and is expected to impact workers within the United States.

The AI incident trackers do not facilitate ready retrieval of labour-related incidents. Being unable to readily see the incidents related to labour makes it impossible to understand what problems are associated with the AI-related risks and harms to labour.

The two AI incident trackers we examined used limited existing labour-related terminology. In OECD AIM, we found only one relevant term—the stakeholder workers—which is applied not only to AI end-users whose labour is replaced or augmented by AI systems, but also to AI producers (e.g., workers under pressure to engineer AI systems).

In AIID we identified three relevant term sets—‘data input’, ‘AI task’ and ‘end user amateur/expert’— from the CSET taxonomy manual for incident classification. However, ‘data input’ did not consider the human labour and roles involved in producing input data (such as training data that the AI system was trained on). ‘AI task’ considers tasks such as ‘resume reading’ or ‘chatbot’, which can be viewed as labour replacement (Hoffmann et al., 2023). The inherent problem of staffing a support hotline for eating disorders with chatbots instead of a human call centre workforce, for example, is that this violates our expectations that therapeutic applications may require emotional labour (Posada, 2020), but ‘AI task’ elides emotional labour.

While the OECD AIM has a taxonomy for industry sectors, its reliance on news reports risks overrepresenting some sectors such as entertainment. Workers stakeholder incidents tend to concentrate in the arts, entertainment, and recreation industry sector, covering workers’ copyright infringement, data protections and personal likeness. And the challenges of generative AI in the workplace have been raised by recent Hollywood strikes calling for controls over licensing creative content. More attention should be paid to understanding AI’s effects on different sectors, especially given the likelihood that some sectors are overrepresented. It was unclear when reconciling how classification terms like worker and ‘future threat’ should be interpreted when combined. Should policymakers rely on OECD AIM’s classification technique when it excludes workers as stakeholders and simultaneously designates incidents in industrial sectors pertaining to earthborn land uses (like real estate, energy, agriculture, etc.) as ‘future threat?’ Representing incidents in this manner presents its own risk and high stakes: Can policymakers properly gauge why heightened future risks in a sector should not portend in harm for workers?

The heavy reliance on news coverage, common to current AI trackers (Turri & Dzombak, 2023), leads to some challenges. News reports can frame different stories with the same evidence, by choosing what to ignore, background or highlight. When Amazon’s implementation of a predictive hiring algorithm resulted in gender discrimination (Dastin, 2018), some news reports attributed the harm to sexist input data (Kraus, 2018) or mischaracterised output from reinforcement learning algorithms as outperforming human judgment (Short, 2018) and free from ‘user abuse’ (Doctorow, 2018; Papadopaulos, 2018). Current AI incident report approaches group incidents geographically and temporally, which reduces the structured data available about where and when incidents took place. Consequently, in AIID, determining which incidents are labour related requires a substantial amount of judgement. When a single news report frames the news event as a labour issue, it makes the case that the incident is labour-related. Yet each incident collects multiple news reports, and labour was often in the background or used as supplemental frame. Finding incidents for labour- related policy topics was different from finding mentions of keywords. Some keywords have high precision, retrieving only labour-related incidents, but low recall, giving a limited picture of the risks and harms. For instance, scheduling only matched incidents about controversial algorithmically driven worker shift management software. Keywords with stronger recall, like ‘compensation’ and ‘hiring’, tended to be less precisely labour-focused, with matching incidents often referring to internal AI enterprise practices and consumer settlements.

Power relationships between the working class and management (and by extension, capital) are salient in the labour-related incidents we flagged: who introduces AI into the workplace? Who is responsible for AI when it backfires? Automated staffing decisions using AI-supported predictive analytics led to multiple problems (e.g., Southwest Airlines’ flight cancellations (Sider, 2022), Tesla’s factory delays (Duhigg, 2018)). Lack of human override and management’s technical under- comprehension exacerbated these problems. Values need to be taken into consideration to determine which labour conflicts themselves qualify as harms.

Future work should examine how the distribution of news reports used in AI incident trackers vary over time, across industries and in relation to AI principles and news media sources and audiences. The ratio of news reports to incidents varies, with some incidents (such as the Hollywood strikes) heavily reported, without regard to the actual exposure to AI-related harms in a given industrial sector. Comparing how news reports frame the same AI incident (e.g., with framing analysis) would be valuable. Likewise, researchers should seek to understand how different groupings into incidents can contribute to policymakers’ problem definitions, perhaps by examining the variation in how news reports are grouped in different AI incident trackers.

Future research could systematically identify which query terms to use, including testing stemming for words such as hiring, firing, and scheduling as well as considering additional terminology such as ‘crowd labor’. The use of terminology in AI governance could also be fruitfully examined, e.g., ‘trustworthy AI’, ‘ethical AI’, ‘AI for good’, ‘beneficial AI’, and ‘responsible AI’ (Stix, 2022).

Current definitions for harm or hazard (Placani, 2017; Rowe, 2021) need to be revisited when evaluating the relationship between AI and labour to enable incident risk prevention (Meyer, 2023). Incident reporting, which has roots in risk management and safety (Johnson, 2003), may not be sufficient for understanding AI risks to workers. Alternative conceptual frameworks for risk management from epidemiology, audit culture and social work may be more suitable for centring workers.

Future work should investigate what harm, incident, potential harm, or hazard mean in the context of labour. For instance, are ‘disruptions’ ever supposed to be weathered as part of events contributing to the normal functioning of the economy? Tradeoffs must consider multiple perspectives (e.g., worker, manager, capital) in the political economy of labour.

Sociotechnical perspectives will be needed. Above all, it is important to consider ‘how algorithms may reshape organizational control’, as Kellog et al. (2020) review. Sociotechnical analysis of fairness in machine learning (Selbst et al., 2019) can inspire new approaches for attending to the power dynamics of AI and labour, drawing on fields such as organizational behaviour and management science, human resources management, labour law and the sociology of labour.

Future classifications of AI and labour should document when a system extracts labour value unjustly. Key examples would be when a ‘worker’ can claim direct credit or likeness to the source data (such as Scarlett Johanssen’s voice appropriated by OpenAI (Pisani & Albert, 2024)) or when input data is created with the express purpose of serving an AI system (Satariano & Mozur, 2023) (see our content moderation/data labour category in Table 3). Taxonomies for organizing how AI interacts with prior types of ‘people problems’ inside the workplace (Moore, 2019) may be helpful since both AI and AI incident reporting ultimately depend on how people are organised to share information within a firm. Standardised terminologies and definitions for workplace safety can be informed by International Labour Organization worker protections requiring information disclosure in the event of mass dismissal and redundancy (De Stefano, 2019). Incidents are made visible based on the power structures upheld by a given taxonomy: naming and power is worth specific consideration.

The two incident trackers we investigated do not adequately capture the nuanced impacts of AI on labour. Particular attention needs to be paid to power imbalances that may increase risks for harm in the workplace. Better definitions are needed to capture labour-related AI incidents, to help policymakers gather evidence to anticipate and mitigate the risks and harms threatening workers across diverse industries. AI incident trackers’ reliance on news reports and limited vocabulary for identifying worker-related harms and risks leads to gaps in understanding AI for policy formulation and problem definition.

Are the labour-related risks of AI intelligible in the incident trackers with a level of detail and reliability that could support comprehensive policy and problem definition for workers’ protection from the harms of AI? Our answer is no.