This is Scale & Strategy, the newsletter that's like having a Bloomberg terminal, but with personality.
Here’s what we’ve got for you today:
AI chatbots are getting easier to manipulate than most people assume.
OpenAI cracks an 80-year-old math assumption
AI chatbots are getting easier to manipulate than most people assume.
A BBC investigation found that tools like ChatGPT, Gemini, Claude, and Google’s AI Overviews can be pushed into repeating false claims pulled from weak or misleading sources online. In some cases, a single post was enough to shape a confident-sounding answer.
To test it, BBC journalist Thomas Germain published a fake article claiming he was a world champion hot-dog eater. Within 24 hours, major AI systems were repeating it as fact.
Funny example. Familiar failure mode.
The same tactic is already showing up outside the joke zone. It’s being used around supplements, investing, and retirement advice, where incorrect outputs can translate into real financial or health damage.
A big part of the issue is presentation. Instead of showing multiple sources side by side, these systems often compress information into a single response. That design favors fluency over verification, which makes it easier to accept the answer without checking where it came from.
Google has updated its spam policies to explicitly ban attempts to manipulate AI-generated answers. The company frames it as a clarification rather than a major shift.
Researchers point to a structural problem. These systems are built to synthesize information, not rigorously validate it. If enough low-quality content points in the same direction, it can start to look like consensus.
Some platforms are adding citations, warnings, and confidence labels. But the incentive structure hasn’t changed, and bad actors adapt quickly, spreading the same claim across multiple sites, videos, and influencer posts until it looks legitimate to retrieval systems.
For now, AI answers work best as a starting point, not a final authority, especially in areas like health, finance, or law where “sounds right” can still be flat-out wrong.
Framer helps teams design, build, and launch their marketing sites lightning fast. With the ability to publish hundreds of CMS pages in a single click, operate at a global scale with seamless localization, and even host unified content across multiple domains, teams have never been able to ship faster. Trusted by companies like Miro, Bilt, and Perplexity
Speed without chaos: ship pages and updates faster without turning the site into a fragile set of one-off hacks
Reduce dependency: shift routine brand and marketing work out of product engineering queues.
Production-grade foundation: Run real marketing systems (CMS, SEO, performance optimization) with governance and collaboration
OpenAI says an internal general reasoning model has produced a new solution that overturns a long-standing assumption tied to Erdős’ 1946 unit distance problem. The company is framing it as an early example of AI making a genuine mathematical discovery.
Erdős’ unit distance problem asks how many equal-length connections can exist between points in a plane. It’s a deceptively simple question that has shaped decades of work in combinatorics and geometry.
The new solution takes a different route, using algebraic number theory instead of the standard geometric and combinatorial approaches.
Mathematicians including Tim Gowers, Noga Alon, and Thomas Bloom have reviewed the proof and confirmed it holds.
The result came from a general-purpose reasoning model still in development, not a specialized system like DeepMind’s AlphaProof.
OpenAI previously walked back a claim that GPT-5 solved multiple Erdős problems after those turned out to be literature matches, not original proofs.
Why it matters: OpenAI researcher Alex Wei called math “a leading indicator of what’s to come.” The signal here isn’t just capability, it’s direction: general models starting to generate original results in mature fields, not just compressing what already exists.
Framer helps teams design, build, and launch their marketing sites lightning fast. With the ability to publish hundreds of CMS pages in a single click, operate at a global scale with seamless localization, and even host unified content across multiple domains, teams have never been able to ship faster. Trusted by companies like Miro, Bilt, and Perplexity
Speed without chaos: ship pages and updates faster without turning the site into a fragile set of one-off hacks
Reduce dependency: shift routine brand and marketing work out of product engineering queues.
Production-grade foundation: Run real marketing systems (CMS, SEO, performance optimization) with governance and collaboration
