Many AI projects are abandoned for practical reasons, not just ethics: a review and practitioner data show funding shortfalls, development hurdles and internal dynamics often halt systems long before deployment. Responsible AI efforts should therefore engage upstream decisions that determine which systems ever get built, not only how deployed systems behave.

Main Finding

The paper maps why organizations decide not to build or to stop building AI systems and shows that abandonment is driven by a broad set of levers beyond ethical concerns. Using a multivocal scoping review and empirical case collection, the authors develop a six-category taxonomy (ethical concerns; stakeholder feedback; development lifecycle challenges; organizational dynamics; resource constraints; legal/regulatory concerns) and demonstrate that many real-world abandonment decisions are motivated by practical, economic, technical, and governance factors as well as ethics. Early-stage (pre-deployment) decisions to not develop are an underexplored but important governance lever.

Key Points

• Taxonomy: Six categories of factors contributing to AI non-development or abandonment:
  • Ethical concerns (discrimination, privacy, misinformation, misuse, labor displacement, environmental impact, loss of agency, etc.)
  • Stakeholder feedback (engagement or external resistance from employees, users, civil society, domain experts, activists)
  • Development lifecycle challenges (unable to measure target, data problems, labeling/ground truth issues, poor model performance, integration/evaluation problems)
  • Organizational dynamics (misalignment with strategy, leadership sponsorship, changing priorities, client refusal, IP/asset concerns)
  • Resource constraints (insufficient technical expertise, compute costs/availability, development/maintenance cost, timeline length, outsourcing incentives)
  • Legal/regulatory concerns (AI-specific regulation, data protection, domain rules, liability, lack of guidance)
• Methodological emphasis: authors intentionally include grey literature, civil society resources, and journalism to capture non-academic, pragmatic drivers of abandonment that conventional academic searches miss.
• Empirical finding: while RAI research often foregrounds ethical risks as a reason to avoid AI, their case data (incident database + practitioner survey) show that many abandonment decisions are driven by non-ethics factors (cost, technical infeasibility, stakeholder pushback, regulatory uncertainty) or by mixtures of reasons.
• Framing: abandonment / non-development is not inherently bad — it can prevent harmful or poorly functioning systems, but can also foreclose beneficial systems that failed for solvable practical reasons.
• Research gap: RAI work should expand focus from mitigation of harms in deployed systems to the incentives, processes, and tools that shape early-stage decisions to build or not build.

Data & Methods

• Scoping multivocal literature review combining academic databases (Google Scholar, ACM Digital Library, Semantic Scholar) and grey literature (industry reports, white papers, tech journalism, civil society repositories).
• Search strategy used combinations of terms like [“not” OR “stop” OR “abandon” OR “halt”] AND [“develop” OR “build”] AND [“AI” OR “ML”]; initial review of first 50 results per source.
• Iterative process:
  • 48 initial sources selected
  • Snowballing added 32 sources
  • Category-specific searches added ~80 sources
  • Inductive thematic analysis by lead researcher to produce six-category taxonomy; categories iterated with team consensus
• Empirical evidence:
  • Collected real-world abandonment cases via a public AI incident database and an online practitioner survey (used to compare factors that drive abandonment pre- vs post-deployment). (The paper contains full sample sizes and case-selection details.)
• Analysis: thematic coding of literature and cases to map factors and compare the prominence and timing (pre-deployment vs post-deployment) of different drivers.

Implications for AI Economics

• Selection effects in deployed AI: Abandonment shapes which AI systems reach the market. Economic assessments of AI impacts (productivity, labor displacement, sectoral change) must account for the censoring effect — many proposed systems never reach deployment because of costs, feasibility, stakeholder resistance, or regulation. Ignoring abandonment risks overestimating diffusion and harms/benefits.
• Investment and R&D allocation:
  • Resource constraints (compute costs, expertise gaps, long timelines) and uncertain regulation materially affect firms' investment choices. These act as private-sector brakes on some AI development paths, changing which innovations receive funding.
  • The paper suggests that cheaper/easier-to-outsource projects are more likely to be pursued, affecting the distribution of in-house capabilities and market concentration (e.g., reliance on external vendors/cloud providers).
• Opportunity cost and sunk costs:
  • Abandoned projects represent sunk R&D expenditures and lost opportunity value (both social and private). Understanding drivers of abandonment can inform more efficient R&D portfolio choices and public subsidy design (e.g., when public funding should de-risk socially valuable but high-cost AI projects).
• Labor markets & organizational behavior:
  • Organizational dynamics and stakeholder resistance influence which internal AI capabilities firms build vs. outsource or drop. This affects skill development, labor demand for AI roles, and the supply of maintenance expertise.
• Regulatory design and signaling:
  • Regulatory uncertainty and liability concerns can deter development, producing a precautionary effect. Policymakers should balance preventing harmful systems with not stifling socially valuable innovation. Clear guidance, proportionate compliance costs, and pathway-specific risk assessments can reduce inefficient abandonment while enabling precaution where needed.
• Governance levers for economists and policymakers:
  • Make non-development visible and reportable: standardized reporting on abandoned AI projects would improve measurement of AI diffusion, R&D efficiency, and the incidence of precautionary non-development.
  • Targeted subsidies or technical assistance: to address resource or technical barriers for socially beneficial projects that are abandoned for economic reasons (compute credits, datasets, labeling support).
  • Incentives for early-stage assessment: fund or require early impact assessments and stakeholder engagement to surface high-risk projects before costly development, aligning private incentives with social risk mitigation.
  • Support for alternative pathways: where abandonment is driven by client or stakeholder refusal, policies could support alternative, lower-risk solutions (human-in-the-loop systems, regulation enabling liability-limited pilots).
• Research implications for AI economics:
  • Incorporate abandonment data into models of diffusion, welfare, and labor displacement to avoid biases from observing only deployed systems.
  • Study the relative elasticity of abandonment to changes in compute costs, regulation, or subsidies to estimate policy lever effectiveness.
  • Quantify the macroeconomic effects of systematic abandonment in particular sectors (e.g., public sector, healthcare) where social value vs private incentives diverge.

Suggested practical next steps for economists and policymakers: - Collect and standardize data on AI project starts, pivots, and abandonments (including reasons and costs). - Model how different policy interventions (grants, regulatory clarity, liability rules) affect abandonment probabilities and welfare outcomes. - Design interventions that lower socially inefficient abandonment (e.g., for high-value but resource-intensive projects) while preserving mechanisms that prevent development of harmful systems.

(For implementation details, sample sizes, and full case analyses see the paper’s methods and appendix.)