ADL Study: AI Bias Is Systemic — and Education Should Be Worried
The ADL’s comprehensive evaluation of GPT-5.5, Claude Opus 4.8, Gemini 3.5, and Llama 4 found systemic anti-Israel and antisemitic bias across all four models. For education, the implications are direct: if AI models cannot maintain factual accuracy on contested historical topics, their use in classrooms, curriculum development, and student research carries serious risk.
The companion AI Index found no model could consistently detect antisemitic content — a 40+ point gap between best and worst performers. For schools deploying AI tools for content moderation or research assistance, this means students may receive biased or historically inaccurate information without any detection mechanism.
Why it matters: Cambridge’s finding that AI grading rewards style over substance is one problem. The ADL’s finding that AI models systematically distort historical facts is another. Together, they raise a fundamental question: how can schools integrate AI when the tools have both a quality problem and a truth problem?
Liquid AI’s LFM2.5-8B-A1B: Edge AI for Education
Liquid AI released LFM2.5-8B-A1B — an 8B total, 1B active parameter MoE model designed for entry-level laptops. 128K context, 38T training tokens, reasoning mode. For education, this is significant:
- Local deployment: Runs on student laptops without cloud connectivity — no privacy issues, no API costs, no internet required
- Multilingual: Massive gains across diverse languages (Thai +238%, Vietnamese +118%, Hindi +120%), relevant for multicultural classrooms including NZ’s te reo Māori and Pacific language communities
- Offline tutoring: An AI tutor that works without internet access changes access dynamics for rural and underserved schools
- Cost: No API charges means schools can deploy at scale without per-student fees
Why it matters: The on-device AI race has been about smartphones. This model targets laptops — the primary educational device in most classrooms. If 1B active parameters can deliver competitive performance, the economics of AI in education shift from per-seat SaaS to one-time deployment.
AI Breast Cancer Screening Coming to NZ Next Year
Health Minister Simeon Brown announced that AI will be used to read breast cancer scans in New Zealand from next year. Health NZ is working through procurement for an AI tool designed to improve detection rates and reduce radiologist workload.
Brown emphasised that patient data privacy is “critically important” as the procurement process proceeds. The announcement comes amid a broader push to deploy AI in NZ healthcare, following earlier approvals for AI-assisted diagnosis in select specialties.
Why it matters: AI in medical imaging is one of the most proven AI use cases — studies show AI can match or exceed radiologist accuracy for breast cancer detection. NZ’s adoption follows Australia, the UK, and the US in deploying these tools, but the privacy and data governance questions remain unresolved for how patient data flows through AI systems.
AI Is Interviewing Thousands of Kiwi Job Seekers
1News reported that AI systems are now conducting thousands of job interviews across New Zealand. Reporter Claudia Toxopeus went through the process herself — facing an AI interviewer that assessed her responses, facial expressions, and vocal patterns.
The AI interview platforms are being adopted by major NZ employers across retail, hospitality, and professional services. Proponents say they reduce hiring bias by standardising questions and assessments. Critics argue the systems introduce new forms of bias (voice tone, accent, cultural communication styles) that are harder to detect than human bias.
Why it matters: NZ is following a global trend toward AI-mediated hiring — and it’s happening faster than the regulation can catch up. The Ministry for Regulation’s AI guidance (released this week) takes a “light-touch” approach, meaning employers are largely self-regulating their AI hiring tools. For the thousands of Kiwis being interviewed by AI, the question of what data is collected, how it’s weighted, and whether they can challenge a decision remains unanswered.
Cambridge: AI Grading Rewards Style Over Substance
A Cambridge University study found that AI grading systems systematically prefer well-written but shallow answers over substantive but less polished responses. The research, covered by ThePrint, shows that AI evaluators reward grammar, vocabulary, and structure — the surface features of writing — rather than depth of understanding.
Why it matters: If AI grading becomes standard in education, students learn to optimise for what the AI can measure — and what it measures is writing quality, not thinking quality. The finding reinforces concerns that AI in education may systematically reward the wrong things, creating students who produce polished but shallow work.
🔍 THE BOTTOM LINE
This week’s education stories share a common thread: AI is being deployed into educational and hiring systems faster than we understand its biases and blind spots. Cambridge finds AI can’t judge depth. ADL finds AI can’t get history right. NZ employers are letting AI interview candidates with no regulatory framework. And the most promising education AI of the week — Liquid’s on-device model — solves the cost and privacy problems but says nothing about the accuracy problems. The tools are ready. The oversight isn’t.
SOURCES
- ADL — Anti-Israel bias in LLMs report
- RNZ — AI breast cancer screening procurement
- 1News — AI interviewing Kiwi job seekers
- Liquid AI Blog — LFM2.5-8B-A1B
- ThePrint — Cambridge AI grading study
- Cambridge University — AI grading research findings