The Ethics Layer in Human Data

Human data, as a field, didn't start as one. It assembled itself from pieces.

Linguists who had been doing corpus annotation for decades. Content moderators who were already reviewing harmful material at scale. Crowdworkers on mechanical labelling images for computer vision researchers. Software engineers who shifted into pipeline tooling. People with domain expertise in medicine or law who found that AI companies would pay for their judgment. All of these streams converged into something that now gets called "human data" or "AI data," but for years, nobody built unified professional standards for the work itself.

The models got their guidelines early. Responsible AI frameworks, model cards, safety evaluations, red-teaming protocols. The infrastructure around what the model does after training has been documented, debated, and regulated in increasing detail since the last decade.

What took longer was the other side of the pipeline. The part where a person sits down at a screen, opens a task queue, reads a set of instructions, and produces the labels that shape how the model learns. That part operated for years on ad hoc norms, company-specific SOPs, and very little external accountability.

That gap is closing now, and not voluntarily.

What regulation is actually demanding

The EU AI Act's Article 50 transparency obligations became legally binding on August 2, 2026. The scope goes beyond labelling AI-generated content. Every provider of a general-purpose AI model is now required to publish a detailed summary of the datasets used for training, following a template the AI Office released in July 2025. That template asks for data modalities, sizes, sources, languages, acquisition dates, and how copyright opt-outs were handled. Non-compliance carries fines of up to 3% of global annual turnover or EUR 15 million.

For annotation vendors, the pressure is arriving through supply chain ethics directives rather than AI-specific law. Sama announced in 2024 that its policies and practices met the requirements of the EU's Corporate Sustainability Due Diligence Directive (CSDDD), scoring 98.6% on a CSR audit aligned with CSDDD standards for a major automotive supplier. That audit covered human rights, health, safety, and business ethics embedded in company practices. Sama was then chosen as the benchmark against which the customer would measure all other annotation vendors.

That last detail is the one worth paying attention to. When a buyer uses one vendor's ethics audit as the scoring rubric for all other vendors in the same category, supply chain ethics become a competitive filter.

The problems that got us here

The ethical problems in AI training data are not abstract. They are documented, measured, and in several cases currently being litigated.

On the data procurement side, the question of what you can legally scrape and use for training has moved from grey area to active courtroom. Thomson Reuters v. Ross Intelligence in February 2025 produced the first final judgment on AI training data copyright, ruling that using Westlaw headnotes to train a competing legal AI was not fair use. The EU Copyright Directive now requires AI developers to check whether a data source carries a copyright reservation and exclude or license that content before using it in training. Web scraping and unlicensed data mining are no longer a grey area in Europe.

On the labour side, the picture is worse. SOMO's March 2026 report found that Amazon, Google, Meta, Microsoft, and Nvidia collectively use at least 30 intermediary companies for data work. Several have been accused of paying below minimum wages, preventing collective organising, and providing no social protections. In January 2026, workers at Covalen's Dublin offices, providing AI training services for Meta, went on strike demanding union recognition and better wages.

Oxford's Fairwork project surveyed over 700 workers on digital labour platforms and found that none of the 15 assessed platforms scored better than the bare minimum on fair pay, conditions, contracts, management, and representation. Zero out of fifteen.

The psychological toll sits alongside the economic one. Equidem's "Scroll. Click. Suffer." report interviewed 113 data labellers and content moderators across Kenya, Ghana, Colombia, and the Philippines and documented over sixty cases of serious psychological harm, from PTSD and insomnia to suicidal ideation. In Colombia, 75 out of 105 workers declined to be interviewed. In Kenya, 68 out of 110. The reason was fear of violating NDAs. As Brookings noted, workers also reported forced unpaid overtime, no fixed salary, and withheld payments.

These are not edge cases. The data annotation market is projected to grow from USD 1.2 billion in 2024 to USD 10.2 billion by 2034. The scale of the workforce producing this data is growing faster than the protections around it.

What annotation programmes should actually do

The regulatory and ethical pressure points converge on a set of practices that are no longer optional for any programme running at scale. Some of these are institutional. Some are things I do as a lead. Some are things I did as an individual contributor before I ever ran a team.

Contributor well-being protections need to be structural. Content warnings before tasks involving graphic material. Enforced session limits for psychologically taxing work. Access to mental health support that is genuinely independent of the employer, not a counselling service that reports back to management.

As a lead, I build this into onboarding. Contributors know before they accept a project what content categories they will encounter. They know they can opt out of specific task types without it affecting their standing. That information has to come before the first task, not after someone has already been exposed to something they weren't prepared for.

Fairwork's five principles (fair pay, fair conditions, fair contracts, fair management, fair representation) are the closest thing the industry has to an independent labour standard for data work. Programmes that cannot demonstrate compliance with something in this range are carrying regulatory and reputational risk that will only grow. Privacy controls on PII need to be built into the annotation workflow itself, not added after.

Guideline quality is the one that gets overlooked most often. When instructions are ambiguous, contributors fill the gap with assumptions, and those assumptions get baked into the training data. I spend more time on guideline writing and calibration than on any other part of programme setup, because a clear guideline is the single best protection against both bad data and contributor frustration.

Why it matters at the individual level too

Most of the conversation about ethics in annotation happens at the programme level. Vendor audits. Regulatory compliance. Institutional policy. That makes sense. The structural problems need structural solutions.

But there is a version of this that applies to the person doing the work, and it often gets skipped.

In annotation, the feedback loop between your actions and their consequences is almost invisible. You log in, open a task queue, label data for a few hours, get paid, and log out. The model that trains on your labels ships months later. You never see the downstream output. You never know whether your judgment on a borderline case shaped the model's behaviour in a way that mattered. That distance makes it easy to treat the work as purely transactional.

I think that's a mistake, and I have seen it play out. Contributors who click through tasks without engaging with the instructions, who label to match patterns they think the client wants rather than applying genuine judgment, produce data that actively degrades model quality. In RLHF and preference work, where the labels directly encode human values, the quality of attention the annotator brings to the task is the quality of the signal.

This does not mean every annotator needs to become an ethicist. It means building a baseline awareness. Know what the data you are labelling will be used for. Read the guidelines with actual care rather than skimming for the minimum viable interpretation. If a task asks you to label personal information without clear consent framing, or evaluate content that is causing you psychological distress without support structures, or produce labels you know are inaccurate because the rubric is broken, say something. Flag it. Escalate it. The Equidem and Fairwork findings exist in part because workers at scale did not feel they could push back. A culture where flagging a broken process is normal and safe to do is not something that gets built by regulation alone. It gets built by the people who do the work deciding that their judgment matters.