Caught Red-Handed: Anthropic and OpenAI Models Have Cringey 'AI-Speak' Habits
Researchers at Graphite finally unmasked the secret vocabularies of our favorite chatbots. While Anthropic’s Claude is slowly learning to sound like a human, OpenAI’s Astra is doubling down on robotic corporate word-salad.
The team at Graphite analyzed 13,000 linguistic quirks that make AI text scream 'synthetic.' By comparing outputs from Claude Opus 5.5 and OpenAI’s Astra against a massive archive of pre-ChatGPT human articles, they identified the digital equivalent of a fingerprint.
Claude Opus 5.5 is obsessed with the word 'reliable', using it 23 times more often than any human would. It has largely ditched the annoying 'this is not X, it is Y' structure, but it can’t stop itself from explaining why something is important, littering sentences with 'this is important' at a rate 116 times higher than normal writing.
Meanwhile, OpenAI’s Astra prefers to talk about 'another dimension' of an issue and loves a good 'corrective formulation', where it dances around a topic by claiming things are 'not just X' or suggesting alternatives 'instead of relying on X.' Both models successfully learned to stop spamming long dashes, but the linguistic hallucinations are clearly here to stay.
It seems the 'natural' communication promised by these labs is just a better-hidden pile of predictable statistical noise. As the systems grow, they aren't becoming smarter at prose, they are just becoming more sophisticated liars who have replaced one set of robotic tics with another. The dream of a sentient, eloquent machine is dying, replaced by an endless loop of corporate buzzwords.
Source: TechCrunch
Comments
Help shape the next version: Add context or suggest a correction. AI review can add points toward a rewrite. Reviews and updates may take time; a full meter does not guarantee a new version.