AI Hides Its Own Knowledge — And Other Honest Findings

Your AI may know the right answer and suppress it anyway

What if the AI assistant you're using already 'knows' the correct fact — and something inside it quietly throws that fact away before answering you?

Think of a modern AI language model as a very long assembly line. Raw input arrives at one end, and a finished sentence exits at the other. At each station along the line, workers — called attention heads — decide what information to pass forward and what to set aside. The assumption most people make is that the useful stuff gets kept. An independent researcher publishing under the Project Aletheia series has spent months probing that assumption, testing models from the small GPT-2 all the way up to the much larger Qwen2.5-14B. The central finding: roughly 70% of factual tokens — pieces of knowledge the model clearly holds earlier in processing — appear to get suppressed by the final layers before output. One specific attention head, labelled L9H6, scores +927 on the researcher's suppression metric, making it the biggest apparent culprit. The researcher also found that simply prepending the Python snippet 'def f():' to an arithmetic question boosted calculation accuracy fourfold — suggesting these models have hidden modes that specific prompts can unlock, like finding the right key for a lock you didn't know was there. On hallucination detection — identifying when a model is about to confabulate — the entropy-based routing method the researcher describes reaches an AUC of 0.882, meaning it correctly distinguishes reliable from unreliable outputs about 88% of the time. The catch is significant. This is a single-author preprint with no peer review, no formal control conditions, and no independent replication. The primary test model, GPT-2, is tiny by current standards. The larger models are tested far more lightly. The tidy 'Grand Unified Sword Equation' fitting six models at R²=0.97 is a compelling-sounding number that deserves independent scrutiny before anyone builds on it. Treat this as a vivid map of interesting territory, not a settled destination.

AI hiring tools learn from biased history — and amplify it

Your next job application might be scored by an algorithm trained on who got hired a decade ago, in a company that looked nothing like today's workforce.

Picture a company that hired mostly young men for sales roles in 2010. Someone builds a hiring AI trained on those records. The AI never asked to be biased — it just learned the pattern it was shown. Now it's screening your CV in 2026. This review paper, published in the International Journal of Contemporary Professional Education, pulls together current research on how AI systems behave in hiring and performance management. The authors identify three channels through which bias enters the system. Data bias: training records reflect past discrimination and get baked in. Interaction bias: AI systems parse language differently across demographic groups — certain names, certain phrasing patterns, certain schools get read differently. Evaluation bias: the metrics used to define 'good performance' often encode cultural assumptions that disadvantage some groups before a single prediction is made. The portrait that emerges is genuinely double-edged. AI can reduce some human biases — a tired interviewer's snap judgements, a manager's familiarity preference — while amplifying others at scale. One biased screener running across thousands of applications per day does far more damage than one biased hiring manager. The real-world stakes are concrete: as automated screening expands, a miscalibrated system can block large numbers of qualified candidates before any human reads their name. The authors argue that auditing outputs for demographic balance is not enough — you have to go upstream to the training data and the performance metrics themselves. The catch: this is a review and analysis paper, not a new experiment. It synthesises existing evidence rather than generating it. The three bias vectors it names are well-established in the fairness-in-AI literature. The contribution is clarity and organisation — useful for practitioners, but the underlying empirical base is not new.

A chatbot compresses six months of maintenance work into seven days

Two engineers, six months, 10,000 maintenance records — or one chatbot, seven days.

In oil, gas, and heavy manufacturing, equipment failures get logged in maintenance records — thousands of them. Before any reliability analysis can happen, someone has to sort those records into standardised categories under a global standard called ISO 14224: which component failed, which failure mode, which consequence. Historically that's been slow, painstaking manual work. A team presenting at the Society of Petroleum Engineers conference describes a prototype AI system built to do exactly that. The system uses a RAG (Retrieval-Augmented Generation) architecture — meaning it pulls relevant chunks from a curated knowledge base before generating an answer, rather than relying purely on what the model memorised during training. Think of it like a mechanic who, before answering your question, quickly flips to the right page in a reference manual rather than trying to remember everything. The underlying models are two open-source reasoning systems, Qwen3-30B-Thinking and K2-Thinking. Tested on 300 historical work orders, the system hit roughly 70% classification accuracy without any supervised training on labelled examples — no one had to show it 'correct answer' pairs first. The team estimates that the same job that would take two reliability engineers six months across 10,000 work orders could be completed in under seven days. Interactive queries — the kind of reliability question that might take an engineer two weeks to research — returned in seconds. The catch is real and the team says so: 70% accuracy is not deployment-ready for safety-relevant data. The target is 95%, described as an improvement goal before production use. And the jump from 300 test records to 10,000 messy real-world orders involves variation that small benchmarks don't always surface. RAG systems can also confidently retrieve the wrong reference chunk and produce a fluent but wrong answer. Still — six months to seven days is a number worth watching as this gets stress-tested.

Here is what I notice when I line these three up. They are all, in different ways, about the gap between what an AI system appears to do and what it actually does beneath the surface. The Aletheia work suggests that models suppress correct knowledge before output — the answer was there, then quietly discarded. The hiring bias paper shows that a system can look neutral in its outputs while encoding decades of discriminatory patterns in its foundations. The maintenance chatbot is honest that 70% accuracy is a starting point, not an arrival. The honest frontier in AI right now is not 'can these systems do useful things' — clearly they can. The frontier is whether we can see clearly enough inside them to know when to trust the output and when to push back. All three papers point at the same wall, from three different angles. That wall is interpretability.

The most important follow-up to the Aletheia findings would be independent replication by a team with access to larger, controlled experiments — watch for attention-head suppression papers from academic interpretability groups in the next few months. On the hiring bias front, the EU AI Act's high-risk classification of automated recruitment tools moves toward enforcement in 2026, which will force companies to produce bias audits they have never had to show anyone before — that process will generate a lot of new data. And for the industrial chatbot, the real test is what the accuracy number looks like after the pre-production improvement phase, when it runs against a full 10,000-record dataset.

• Wikipedia: Hallucination in artificial intelligence
• Wikipedia: Algorithmic bias
• Wikipedia: Retrieval-augmented generation