Hallucination Hype vs. Reality
OpenAI has unveiled GPT-5.5 Instant as the new default model for ChatGPT, boasting a 52.5% reduction in hallucinated claims on high stakes prompts in medicine, law, and finance. The company also touts a 37.3% drop in inaccurate claims on user flagged conversations. But let’s be real: these are internal evaluations from a company with a massive incentive to polish its image after years of AI making stuff up. Without independent audits, these numbers are just marketing dressed as metrics. The model’s system card, linked in their announcement, offers some methodology, but it reads more like a press release than a rigorous scientific paper.
Fewer Emojis, More Control, Same Old Questions
Beyond hallucinations, GPT-5.5 Instant is supposedly more capable at everyday tasks like analyzing image uploads and deciding when to search the web. OpenAI claims responses are ‘tighter and more to the point,’ and that the model will avoid ‘gratuitous emojis.’ A welcome shift, but hardly a revolution. The real news is the ‘memory sources’ feature that lets users see what context ChatGPT pulled from previous chats or Gmail to craft a response. This is a nod to transparency, but it also raises fresh privacy concerns. Users can now delete or correct stored information, but this feels like a band aid on a system that’s been hoarding personal data by default. No CVEs were mentioned in the article, but AI systems like this often rely on frameworks with known vulnerabilities; for instance, CVE-2025-31147 highlights risks in unsecured API endpoints that could expose user memory data. OpenAI is rolling this out to Plus and Pro users first, leaving free and business tiers waiting again.
Source: Theverge
