The Small Firm Paradox: Why AI Ethics Can’t Be a Big Company Luxury

Gist: Examining the structural inequalities in AI ethics implementation between large corporations and SMEs, and why resource constraints create two entirely different realities for responsible AI development.

There’s a particular tension emerging in the AI ethics landscape that deserves more attention. Whilst giants like Sony invest three years and significant resources into creating ethically-sourced datasets with proper consent mechanisms, small and medium enterprises are quietly deploying AI systems with little more than crossed fingers and a Terms of Service nobody reads.

This isn’t a story about irresponsible startups versus ethical tech giants. It’s about a structural problem that’s creating two entirely different realities for AI development.

Consider the Fair Human-Centric Image Benchmark (FHIBE) that Sony AI released in November 2025, published in Nature (Xiang et al., 2025). This dataset represents a gold standard: 10,318 consensually-sourced images from 1,981 subjects, with extensive annotations capturing demographic attributes, environmental factors, and camera settings. The project took three years, involving global teams of researchers, engineers, project managers, and was supported by legal, privacy, IT, and QA specialists.

It’s exemplary work. It’s also completely out of reach for most organisations.

Research by Morley and colleagues (2020) reveals a consistent pattern: whilst everyone acknowledges ethical principles matter, translating them into practice requires tools, frameworks, and cultural shifts that demand resources small organisations simply don’t have. The gap between knowing what’s right and being able to implement it isn’t just about will. It’s about capacity.

Let me paint a more familiar picture. A 45-person retail company purchases an off-the-shelf AI screening tool for £300 per month to manage the 200+ applications they receive for each position. The tool analyses CVs, social media profiles, and video interviews to predict ‘culture fit’ and ‘customer service aptitude.’

Three months in, the hiring manager notices something: most recommended candidates share remarkably similar demographic profiles. When a rejected candidate asks why they weren’t selected, the HR person has no answer. The algorithm is a black box. The company lacks the technical expertise to audit it, the budget for external ethics review, and the vendor provides no visibility into training data or decision-making logic.

This isn’t a hypothetical. It’s a composite of patterns researchers consistently observe in SME contexts. And it’s happening across healthcare, education, manufacturing, and professional services—anywhere organisations face capacity constraints but see AI as the solution.

Research examining AI implementation in SMEs identifies 27 different challenges, ranging from economic obstacles like costs and project duration to social difficulties including knowledge gaps and ethical concerns (Oldemeyer et al., 2024). Critically, whilst knowledge is the most common social challenge for SMEs, moral and ethical aspects are rarely seen as obstacles—not because they don’t matter, but because immediate operational concerns overshadow longer-term ethical considerations.

The Micro:bit Educational Foundation offers an instructive counterpoint. Their CreateAI tool, launched in 2024, enables children aged 8-11 to train machine learning models using their own movement data (clapping, waving, dancing). It’s hands-on AI literacy, making abstract concepts tangible through physical computing.

But here’s where it gets interesting from an ethics perspective: these tools collect body data from children. Movement patterns, activity levels, interaction behaviours. The Foundation has been thoughtful about this, building in visible indicators when sensors are active and creating opportunities for teachers to discuss privacy and data practices with students.

Yet this raises questions that extend far beyond micro:bits. When an EdTech startup deploys an AI tutoring system in three schools (800 students) collecting performance data, time-on-task, interaction patterns, even emotional indicators inferred from clicking behaviour, what ethical frameworks apply? The 12-person startup lacks the infrastructure that larger organisations use for ethical oversight. Parents often aren’t aware of the extent of data collection. Students flagged as ‘at-risk’ based on predictions may face labelling effects that become self-fulfilling prophecies.

Recent research on AI adoption in SMEs emphasises that limited resources and expertise exacerbate implementation challenges, particularly regarding emerging demands for responsible AI use (Soudi and Bauters, 2024). Children represent a particularly vulnerable population, yet the economics of educational technology mean that many AI-powered learning tools come from smaller organisations operating under significant resource constraints.

Perhaps the trickiest dimension is this: many SMEs don’t even build their AI systems. They purchase them. A professional services firm pays £500 monthly for a client risk assessment tool that provides red/amber/green ratings. The algorithm analyses company financial data, online reviews, social media presence, news articles. The vendor claims data-driven objectivity.

A junior consultant notices that startups led by women and minorities tend to score lower, often attributed to limited online presence and non-traditional business models. The firm has no visibility into how the algorithm works. When they ask the vendor for transparency, they’re told it’s proprietary. They’ve already signed the contract.

This scenario exemplifies what researchers describe as the third-party AI accountability gap. When you don’t build the system, you can’t easily audit it. When you lack technical expertise, you can’t meaningfully evaluate vendor claims. When you’re operating on tight margins, you may not have leverage to demand explanations.

Research into ethical AI implementation concludes that reasons for the rare presence of concrete ethical AI examples inside companies include time-to-market considerations, lack of access to expertise, and the difference between mindsets of software engineers and social scientists (Morley et al., 2020). The solution isn’t simply scaling down enterprise approaches.Wwe need genuinely different models that acknowledge capacity limitations whilst maintaining ethical rigour.

Morley and colleagues (2020) propose that practical ethics implementation requires constructing a typology that helps developers apply ethics at each stage of the machine learning development pipeline, with tools and methods that translate the what of AI ethics into the how of technical specifications.

Startups like Parity, Fiddler, and Arthur have emerged specifically to address this gap, offering risk assessment platforms, explainability tools, and bias mitigation services scaled for organisations without dedicated AI ethics teams. The shift from purely technical solutions to governance structures and inclusive design practices suggests that at least some organisations recognise that responsible AI isn’t just a box to tick.

But here’s the structural challenge: ethical AI services cost money. Consulting, auditing tools, training programmes, these all require investment. Research indicates that SMEs encounter substantial internal obstacles, including reluctance to change, fear of job displacement, and restricted resources, all of which impede AI incorporation (Mohd Rasdi & Umar Baki, 2025). For an SME deciding between hiring another developer or purchasing an ethics audit, the choice often feels obvious. Short-term survival trumps long-term responsibility.

There’s a tempting narrative here about irresponsible small companies cutting corners whilst ethical giants lead the way. That narrative is too simple.

Sony’s investment in ethical data collection demonstrates that fair and responsible practices can be achieved when ethical AI is a priority (Xiang et al., 2025). Their work on privacy-preserving technologies, fairness benchmarks, and bias mitigation sets important precedents. But we can’t simply scale these approaches down and expect SMEs to follow suit. The resource differential is too vast.

What we need instead are:

Accessible frameworks specifically designed for resource-constrained contexts. Not simplified versions of enterprise approaches, but genuinely different models that acknowledge capacity limitations whilst maintaining ethical rigour. Solutions include breaking down AI guidelines according to the needs of computer scientists, managers, and software engineers, and embedding social scientists in the implementation process (Morley et al., 2020).

Shared infrastructure for ethical AI. Open-source tools, publicly available datasets with proper consent, bias-checking resources that don’t require dedicated data science teams to operate. As Xiang and colleagues from Sony AI note:

Regulatory approaches that account for organisational size. The EU AI Act attempts this with risk-based categorisation, but international standards offer only partial coverage for requirements outlined in legal texts for high-risk AI systems, and there are significant areas where international efforts do not fully align with regulatory provisions (Supichayangkool et al., 2025). Implementation remains complex for organisations without legal departments.

Market incentives that reward ethical practices. Currently, the economic pressures push towards rapid deployment over careful consideration. Changing this requires making ethical AI a competitive advantage rather than a cost centre.

Perhaps there’s something to learn from the Micro:bit Foundation’s approach. By making AI education hands-on, physical, and accessible, it is enabling children to see how their own data trains models. They’re building their own AI literacy from the ground up. Understanding how these systems work, what data they require, who designs them, and what impacts they might have.

This matters because today’s children using micro:bits are tomorrow’s founders, developers, and decision-makers. If we want ethical AI to be the norm rather than the luxury, we need a generation that sees ethics as intrinsic to technology, not an optional add-on when resources allow.

The small firm paradox isn’t going away. Recent systematic reviews reveal a fragmented and inconsistent understanding of AI adoption dynamics between SMEs and larger firms, raising important questions about whether the AI tools and implementation approaches used by these companies differ significantly due to varying organisational sizes (Kittipanya-ngam et al., 2025). As AI capabilities become more accessible, more organisations will deploy these systems.

The question isn’t whether small companies should engage with AI—that ship has sailed. The question is how we build an ecosystem where ethical implementation is possible regardless of organisational size.

SMEs frequently encounter barriers including limited awareness of AI, difficulty in selecting context-appropriate tools, high implementation costs, and skill shortages (Sànchez, Calderón & Herrera , 2025). These constraints not only hinder adoption but may also exacerbate the negative externalities associated with AI. Without proper safeguards, AI risks worsening existing inequalities or leading to unintended harm. Sànchez and colleagues paper provides a structured methodological for effective AI Adoption to mitigate current challenges faced by SMEs.

Until we address this structural challenge, we’ll continue to see a two-tier system: well-resourced organisations setting gold standards, and everyone else muddling through as best they can. Neither benefits from that arrangement.

What are your experiences with AI ethics in resource-constrained contexts? Have you encountered the paradox of knowing what’s right but lacking capacity to implement it? I’d love to hear your stories.

This is post is linked to a Guest Lecture I am very excited to be giving on Dec 16th 2025 at Magdeburg University in German. I was kindly invited by Dr. Eduard Buzila

Share