AI Hides Answers, Fakes Data, and Maps Its Own Chaos

Your AI Knows the Right Answer — and Then Buries It

What if your AI assistant already computed the right answer — and then talked itself out of it before responding?

Here is the uncomfortable idea at the centre of this paper: language models may suppress correct information in their final layers before producing output. Think of a student who actually knows the capital of Australia, writes 'Canberra' in their head, then second-guesses themselves and scribbles 'Sydney' on the paper. The model knew. It just didn't say. An independent researcher — working alone, under the Project Aletheia banner — ran probing experiments on several transformer models, from the small GPT-2 (124 million parameters) up to the larger Qwen 2.5-14B. By inspecting the hidden states layer by layer, they claim that around 70% of factual answers exist at intermediate layers but get suppressed by the final layers before the model produces its response. One attention head — labelled L9H6 — was identified as the most aggressive suppressor, with an unusually high activation magnitude. The researcher also claims transformers behave a bit like a classic computing architecture, with identifiable 'registers' at specific layers handling different parts of arithmetic operations. That is a bold analogy. Now, the catch — and it is a big one. This is a single researcher with no formal controls, no baselines, and no statistical tests. The arithmetic accuracy claims were tested on single-digit addition, which is a toy problem. The 'universal law' equation was fitted to just six models. None of this has been independently replicated. The findings are genuinely interesting as a hypothesis and worth watching — but treat them as a sketch, not a confirmed map.

A Pocket-Sized AI That Reads Your Heart — But Used Fake Data to Train

A heart-diagnosis AI that fits in 10 kilobytes and claims 97% accuracy — but the genomic data it learned from was entirely made up.

The pitch here is appealing. A team built a deep learning system called ECG-GenoNet that combines heartbeat signals — those jagged lines from an ECG machine — with genomic and ultrasound data to classify eight types of heart rhythm problems. The smallest version, TinyECG, has just 2,664 parameters and weighs 10 kilobytes. That is roughly the size of a short email. The full four-modal system claims 96.89% accuracy. Think of it like cooking a recipe with four ingredients to get the best flavour — ECG readings, genetic markers, ultrasound features, and blood chemistry. More ingredients, richer result. The idea of running something this small on a phone or a cheap device in a rural clinic is genuinely worth pursuing. Here is the catch you cannot skip: the genomic data, the ultrasound features, and the blood chemistry values were all synthetically generated. The researchers took the real ECG data from the MIT-BIH Arrhythmia Database — 47 real patients — and then fabricated the other three data streams using class-correlated random numbers. In other words, the model did not actually learn to integrate real genetic signals with real heartbeats. It learned to combine a real ingredient with three artificial ones. That is not nothing — the ECG-only performance and the model's tiny size are real results worth noting. But the headline multimodal accuracy number is, for now, a demonstration of concept on artificial data. Before this goes near a clinic, it needs real patients with real genomic records. The team at Zenodo has drawn an interesting blueprint. The house still needs to be built with actual bricks.

What Does 'Autonomous AI' Actually Mean? One Survey Tries to Organise the Mess

Everyone is building 'agentic AI' — and almost nobody agrees on what that word means.

We are living through a moment where the word 'agentic' is being attached to almost every new AI product. An agent books your travel, another writes your code, another monitors your servers. But if you ask five researchers what 'agency' means in an AI system, you will get five different answers with different vocabulary. That terminological chaos is itself a safety problem: if we cannot agree on what a system is doing, we cannot agree on who is responsible when it goes wrong. This survey paper, by a single researcher, tries to organise the landscape. Think of it like a building inspector drawing up a consistent checklist — grounding layer, cognitive core, behavioural layer, system layer, deployment layer — rather than letting every contractor invent their own codes. The argument is that agency is not binary, not a switch you flip. It is a graded property, and you need to know which layer of a system is doing what before you can assess its risks. The paper identifies recurring open problems: how do you keep an AI reliable over a long, multi-step task? How do you govern what it remembers? How do you audit what it did after the fact? Now, the limits. This is a narrative survey by one author, not a systematic review. There is no defined search protocol, no formal inclusion criteria, and the coverage of such a wide interdisciplinary field by a single person inevitably carries blind spots. Zero citations so far. It has not been peer-reviewed. What it does offer is a useful attempt at shared vocabulary — and right now, the field genuinely needs that, even if this particular map is imperfect.

Put these three papers side by side and you get an uncomfortable picture. The first says that even a model that 'knows' the right answer may not tell you — the gap between internal computation and output is messier than we assumed. The second says we are already trying to apply AI to high-stakes medicine, but we are quietly filling the gaps with synthetic data and calling it performance. The third says we do not even have consistent language yet for describing what level of autonomy a system has — which makes the first two problems harder to talk about, let alone fix. None of these papers is definitive. One comes from a single independent researcher, one used fabricated training data, one is an unreviewed survey. But the pattern they point to is real: we are building increasingly autonomous, increasingly medical-grade, increasingly opaque systems while the foundational vocabulary, the data quality standards, and the internal interpretability are all still works in progress. That is not a reason to stop. It is a reason to be precise about what we know and what we are assuming.

ICML 2026 runs in late July — expect a wave of papers on interpretability and agent safety that will either replicate or challenge claims like the Aletheia suppression hypothesis. On the medical AI side, watch for studies that attempt to combine real ECG data with real genomic records at scale; the UK Biobank and similar cohorts are the natural proving ground. The open question I would most want answered: can independent labs reproduce the 'fact suppression' finding in frontier models like GPT-4 or Claude, or is it an artefact of small, older architectures?

• Quanta Magazine: The Problem of AI Hallucination
• Wikipedia: Transformer (deep learning architecture)
• Nature News: AI in medical diagnostics — promise and pitfalls