26 April 2026: OpenAI Apologises Over Tumbler Ridge as Sovereign AI Push Gains Momentum

Sam Altman issues a public apology over the Tumbler Ridge shooting, Europe moves to build its own AI champion, Anthropic’s agents negotiate real deals with real money, and an amateur mathematician uses ChatGPT to crack a problem that had defeated professionals for six decades.

OpenAI’s chief executive has publicly apologised to the community of Tumbler Ridge in British Columbia, Canada, after the company failed to alert authorities about a user who later carried out a mass shooting. In a letter addressed to residents, Sam Altman said he was “deeply sorry” that OpenAI did not act on information it held about the suspect. The apology is significant because it represents a major AI company acknowledging, for the first time in this direct a way, that it may hold safety-relevant information and that failing to act on it carries consequences. The question of what AI platforms are legally required to do when they detect signs of planned violence is unresolved in most jurisdictions, including the UK. Existing frameworks around safeguarding and data protection were not designed with AI conversation systems in mind, and regulators in Westminster and Brussels will be looking at this case as they develop clearer guidance.

Europe’s AI landscape shifted this week as Canadian startup Cohere announced it is acquiring Germany’s Aleph Alpha, with financial backing from Schwarz Group, the parent company of supermarket chain Lidl. The deal has the informal endorsement of both the Canadian and German governments, and is explicitly pitched as a sovereign alternative to the American AI platforms that currently dominate enterprise deployments. Aleph Alpha had built a reputation for AI models designed with European data protection requirements in mind. Cohere brings commercial scale and a strong track record with large enterprise clients. For UK organisations operating under the UK GDPR and looking for alternatives to US-hosted AI, this merger is worth monitoring closely. A combined Cohere and Aleph Alpha could become one of the most credible European-friendly options for regulated sectors including financial services, healthcare, and the public sector.

Anthropic has been running a controlled experiment in which AI agents negotiated and completed real commercial transactions entirely on their own. The test, reported by TechCrunch, involved Claude-based agents taking on the roles of buyers and sellers in a classified marketplace. The agents assessed listings, entered negotiations, and closed deals involving actual money, without a human approving each step. The sums were modest and the environment was purpose-built for the experiment, not a live product. But the result matters because it demonstrates that the core mechanics of agentic commerce, where AI systems act as economic actors rather than tools, are already functional. For business owners, this is a signal of where AI automation is heading: from drafting emails and summarising documents to executing procurement, vendor negotiations, and customer transactions on your behalf.

A hobbyist mathematician with no formal research credentials has used ChatGPT to solve a combinatorics problem that had been unsolved for 60 years. The problem, known as Erdős 1196 and catalogued on the ErdősProblems.com database, was one of the many open questions left by the legendary Hungarian mathematician Paul Erdős. According to Scientific American, the solver worked through the problem in extended conversations with ChatGPT, iteratively testing conjectures, checking logic, and refining arguments over many sessions. The solution has since been verified by professional mathematicians. The practical meaning for anyone who uses AI tools at work: these systems are no longer just good at retrieving and summarising what is already known. Used patiently and systematically, they can support genuine original reasoning. If an amateur can crack a 60-year-old mathematical puzzle with conversational AI, the ceiling on what a determined non-expert can accomplish with AI assistance is higher than most people currently assume.

OpenAI has retired SWE-bench Verified as a primary measure of its models’ software engineering capabilities. The benchmark, which tests how well AI systems can resolve real GitHub issues, has become saturated: frontier models now perform at a level where the test no longer distinguishes meaningfully between them. OpenAI published a post today explaining it is moving to harder evaluation frameworks that better reflect the demands of real-world software development. For developers and companies using AI coding assistants, this is a useful reminder that benchmark numbers on marketing pages often measure the last generation of capability, not the current one. The tools available now are likely more capable than any single score suggests, and the metrics used to describe them are still catching up.

Here is everything else worth knowing from today’s AI news.

• Maine’s governor vetoes data centre moratorium — The state’s governor blocked L.D. 307, which would have imposed the United States’ first statewide ban on new data centre construction, running through November 2027. TechCrunch
• Bay Area homeowner is selling a 13-acre property for Anthropic equity — A Mill Valley estate north of San Francisco is reportedly on the market with a condition that buyers must hold Anthropic shares, a measure of how the company’s valuation is reshaping wealth in the Bay Area. TechCrunch

This is a daily news update for informational purposes only. AI products and policies change rapidly. Verify details directly with providers before making decisions. Nothing here is financial or legal advice.

AI Daily is Cristoniq’s afternoon update on developments in artificial intelligence, published every weekday afternoon.