Chapter 14 — Bias, Fairness, and Responsible AI

Fourteenth post of the chapter-by-chapter walkthrough of LLM Primer VII: AI Security. The chapter that treats responsible AI as a discipline of choices made under uncertainty — where technical tools surface trade-offs without resolving them.

Why this chapter exists

Bias, fairness, and responsible AI are the substantive content of what the regulations of Chapter 13 are trying to address. The technical literature and the organisational literature meet here. The chapter walks the sources of bias in LLMs, the fairness measurement literature and its methodological limits, the safety-utility trade-off documented in the alignment work, transparency and explainability as related-but-distinct disciplines, and organisational AI policy as the layer that translates all of it into operational practice. Bender, Gebru, McMillan-Major, and Shmitchell's 2021 "Stochastic Parrots" paper set the reference framing; the field has spent the intervening years working out what the frame implies for engineering.

14.1 Bias has several sources with different mechanisms

Bias in an LLM is not a single phenomenon. The principal sources are training-data bias (the corpus reflects the population that produced it — English over-represented, some demographic groups more represented than others, historical patterns of association preserved), representational bias (some concepts or groups are represented with less nuance because the training signal was sparser), allocation bias (the model's outputs distribute a resource — attention, opportunity, credit — unevenly across groups even when individual outputs seem reasonable), evaluation bias (the benchmarks used to certify the model reflect the biases of their creators and their reference populations), and deployment bias (the context of use pushes the model toward outcomes the training did not anticipate). Each has a different mechanism and different mitigation path. Training-data bias is addressed through curation and augmentation, with limits — you cannot fabricate representative data that does not exist. Representational bias is addressed through targeted fine-tuning, with the caveat from Chapter 16 that fine-tuning can also erode alignment. Allocation bias requires system-level intervention rather than model-level tuning. Evaluation bias requires expanding the benchmark set. Deployment bias requires product-level scrutiny that no amount of model work can substitute for.

14.2 Fairness is measured, imperfectly, by benchmarks that disagree

Fairness measurement in LLMs has produced a substantial methodological literature and several standard benchmarks. BOLD (Dhamala et al., FAccT 2021) measures sentiment, toxicity, and regard in open-ended generation across demographic groups. BBQ (Parrish et al., 2022) uses hand-built question-answering pairs to probe bias. StereoSet and CrowS-Pairs probe stereotype associations. Each benchmark measures something different, and no single benchmark captures the fairness properties an organisation might care about. The methodological literature is also clear that fairness metrics can be mutually inconsistent — improving group parity may worsen calibrated accuracy across groups, and vice versa — so the choice of metric is itself a value choice the organisation has to make rather than defer to technical judgement. The safety-utility trade-off, documented in Anthropic's 2022 paper "Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback" and continuing through the DPO literature, is the empirically established observation that training the model to be more harmless tends to also train it to be less helpful. Modern alignment methods have shifted the frontier but not eliminated the trade-off. The engineering choice is where along the frontier to operate for the specific product, and the choice has to be defensible to the users, the regulators, and the audiences the trade-off affects.

14.3 Transparency and organisational policy carry the load

Transparency (disclosure of system properties) and explainability (accounts of specific outputs) are conceptually distinct. Transparency is largely served by the documentation artefacts of Chapter 13 — model cards, system cards, datasheets. Explainability is the more technically demanding problem. SHAP (Lundberg and Lee, NeurIPS 2017) and LIME (Ribeiro et al., KDD 2016) were developed for classification and adapt imperfectly to token generation. Mechanistic interpretability — Anthropic's dictionary-learning work, OpenAI's automated circuit discovery — is a research frontier with production applications still forming. The regulations often ask for kinds of explanation the state of the art cannot yet deliver, and the honest engineering answer is to name that gap rather than paper over it. The organisational AI policy is where the substantive concerns become operational. The policy has to establish who has authority over AI decisions, an inventory of AI systems in use, a risk-classification approach, a lifecycle discipline from evaluation through retirement, a data-handling standard, and a human-oversight standard. Anthropic's Responsible Scaling Policy, OpenAI's Preparedness Framework, Google DeepMind's Frontier Safety Framework, and Microsoft's Responsible AI Standard are the published examples that have set the industry's floor.

What Chapter 14 sets up

Chapter 15 turns to the organisational infrastructure that supports the discipline: security culture appropriate to AI work, red-team and audit functions that test the organisation's posture, vendor risk assessment that handles the supply chain, continuous evaluation infrastructure that supports ongoing assurance, and long-term model stewardship. The treatment builds on Chapter 13's regulatory context and Chapter 14's substantive concerns and gives them operational form. Chapter 16 then narrows to fine-tuning as its own security surface — alignment erosion through benign data, deliberate poisoning, evaluation gates in CI, rollback discipline — and Chapter 17 closes the volume by looking at the threats that are still forming: autonomous agents, multimodal attack surfaces, synthetic identity, and the AI-versus-AI dynamics of mid-2026.

Next — Chapter 15: Building a Secure AI Organization. AI-specific security culture, internal red teams, vendor risk assessment, continuous evaluation, and long-term model stewardship.