AI2Day Daily Brief — 9 Sept 2026

Good morning, it is Wednesday the ninth of September, and the biggest story today is Google DeepMind dropping what it calls a complete prediction map of every possible single-letter change in human DNA, free for researchers to use right now.

Google DeepMind released AlphaGenome Atlas yesterday, a database covering all nine billion possible single-letter changes in the human genome. The dataset weighs in at roughly one petabyte, which DeepMind says is more than thirty times the size of the AlphaFold protein database that won the Nobel Prize in Chemistry in twenty twenty-four. A companion scoring tool called the AlphaGenome Variant Impact score collapses the predictions down to one number per variant, covering both protein-coding and non-coding DNA. That matters because testing these variants in a lab, one by one, would take lifetimes. Early uses include Broad Institute researchers flagging a variant in the D-N-M-one gene linked to epileptic encephalopathy, later confirmed in the lab, and a University of Exeter researcher applying it to whole-genome data from over fifty-four thousand UK Biobank participants. Atlas is free for non-commercial research now, with a commercial version on Google Cloud coming.

Two stories about the same tool in one morning is a signal of how significant people think this is. Story three is a different kind of AI problem entirely.

A new research paper published via Hugging Face finds that standard AI safety training can cause a model to refuse perfectly ordinary questions simply because they sit near sensitive topics. The team, from Multiverse Computing, tested on Qwen three, an eight-billion-parameter language model, using political content as their test case. They found that over-refusal on safe prompts jumped from two percent to seventy-four percent on one popular benchmark after safety training. There was also a data problem: a single-shot training approach was silently dropping nearly twenty percent of training examples before the team caught it. Their fix, adding what they call boundary pairs, matched prompts where one should be blocked and one should not, cut false refusals near the problem zone from roughly thirty-three percent to just over four percent. The headline finding is that measuring only how often an AI blocks bad content gives an incomplete and potentially misleading picture of whether it actually works.

Worth noting for anyone evaluating AI tools internally: refusal rates in one direction do not tell you what is happening in the other. Next, Australia.

Australia's Labor government is planning legislation that would require digital platforms to offer users a genuine opt-out from algorithmically curated feeds. The proposal would cover social media, search engines, and AI chatbots operating in the country. Choose the opt-out, and you would get content in a different order, presumably chronological or unranked, rather than whatever the platform predicts will keep you engaged longest. The prime minister has said publicly that the tech industry will push back hard. The opposition has raised a different concern, arguing the scheme could open a door for government influence over what users see. The policy is still a proposal at this stage, but Australia has moved quickly on platform legislation before, so this is one to track.

Regulatory pressure on recommendation systems is building in several jurisdictions at once, so the Australian move is worth watching for what follows elsewhere. Finally, a workflow story for anyone in video production.

Adobe is adding a feature called Generative Media to Premiere Pro, its professional editing software. It lets editors generate AI video clips and audio directly inside the editing timeline without switching to another application. Inside the tool, editors can choose from five AI video generators: Adobe Firefly, Google Veo, Runway, Luma, and Kling. Spot a gap in your cut, highlight it, and Premiere can generate a context-aware clip or sound to fill it, already editable in place. New AI audio tools, currently in beta, can separate overlapping speakers, automatically duck music when someone is talking, and let editors adjust dialogue and background sound independently. Adobe's plain-language AI Assistant is also coming to After Effects in beta, letting editors rearrange projects and build effects without writing code. Adobe's own characterisation, for what it is worth, is that this eliminates the round-trip to external tools.

Today's sponsor is Zero2Hero: an AI poker coach sits at your table, and you can ask it mid hand whether the player across from you is bluffing. It runs in your browser on any computer. Windows, Mac, Linux, even a Chromebook. Nothing to install. Lifetime access is $99, in US dollars, until 15 September, at zero, the number two, hero, dot A I.

For the full breakdown on all of these stories and more, the Monday show AI Today Weekly goes deep, so catch that if you want the longer read. That is your AI briefing for today. Every story is at A-I-2-Day dot live. That is A, I, the number two, D-A-Y, dot live. We are back tomorrow morning. Do us a favour before you go: hit that thumbs up, and subscribe for next week.