The brief · 23 April 2026

The AI brief, 2026-04-23

Busy day: GPT-5.5 lands in ChatGPT and Codex, Insilico keeps up the momentum seen on April 21 with a drug candidate designed in Abu Dhabi, and two US bills set concrete limits on AI in clinical settings and the anonymity of its work.

2 min read J / K to navigate

Taiwan builds its own banking model to cut ties with US tech giants

Sixteen Taiwanese banks have pooled resources to fund an AI model tuned to local regulation, something ChatGPT or Gemini don't handle natively out of the box. This shows that even tech-rich countries see generalist models as falling short for banking. If Taiwan pulls this off, other regional markets are likely to copy the approach instead of paying for a license from OpenAI or Google.

practitioners › First release expected late 2026, focused on banking knowledge, backed by the Taiwan FinTech Alliance (16 institutions).

Bloomberg / Taipei Times →

GPT-5.5 rolls out in ChatGPT and Codex

OpenAI retrained its flagship model from scratch, not just tweaked a few settings. That translates into a ChatGPT that's noticeably stronger at coding, research, and data analysis, plus a context window bumped up to 1 million tokens. In practice, you can hand it an entire codebase or a stack of documents without chunking anything.

practitioners › GPT-5.5, 1M-token context window, rolling out now in ChatGPT and Codex.

OpenAI →

Delaware bars AI from being licensed as a nurse or doctor

It sounds obvious, but it wasn't legally settled before this law. Delaware shuts down any attempt to have an AI agent hold a protected clinician title, even as clinical AI tools already work alongside human staff. This sets a marker for other states still weighing the question.

practitioners › HB 191, signed April 23, 2026 by Delaware's governor, bars non-human entities, including AI agents, from being licensed as a nurse, physician, or physician assistant, or from using those protected titles.

Delaware General Assembly →

US bill would force AI to sign its own work

The Protecting Consumers from Deceptive AI Act would direct NIST to set watermarking and provenance-tracking standards for AI-generated content. If it passes, every AI-produced image or text could carry a verifiable mark of origin. The open question: will mainstream models actually comply, or find ways around it.

practitioners › Bill filed April 23, 2026 on Congress.gov, tracked as H.R. 8479.

Congress.gov →