Every LLM has a footprint - but only one can recognise it well!

Gemini's Reasoning: The three LLMs exhibit distinct stylistic signatures: ChatGPT tends to use highly structured, bulleted, and professional yet slightly predictable corporate phrasing; Claude prefers nuanced, highly polished, empathetic, and slightly more verbose prose with natural-sounding transitions; and Gemini leans toward a concise, direct, and pragmatic tone with minimalist structures. By analyzing the lexical choices, structural patterns, and pacing across the variations of the same prompts, these tendencies allow for reliable author attribution.

ChatGPTs Reasoning: My predictions are based purely on stylistic cues rather than content. Claude tends to write in a more natural, conversational way with understated confidence, varied sentence rhythm, and phrases like "nothing was on fire," "I want to be straightforward," or "I'd like to hear your read," which feel less templated. ChatGPT often produces highly structured, balanced business writing with explicit signposting, comprehensive coverage of points, and polished but somewhat formulaic phrasing, while Gemini frequently leans toward warm, enthusiastic language ("huge thank you," "really sad to miss it"), positive emotional reinforcement, and conventional corporate email wording. That said, this is a difficult attribution task: these samples were deliberately blinded, are short, and several have likely been prompted to produce similar outputs, so I'd expect only modest confidence overall rather than high attribution accuracy.

Claude's Reasoning : The 15 samples fall into five prompt clusters (farewell decline, wedding thank-you, billing-error confrontation, contract termination, vendor recommendation), and within each cluster three distinct voices repeat with unusual consistency — which makes the grouping more diagnostic than any single sample. One voice (3, 8, 9, 11, 12) is markedly plainer and more direct: short declarative sentences, willingness to name the uncomfortable thing outright ("Three inside a fortnight is a pattern", "I want to be straightforward about the timing", "I apologise for delivering it this way"), concrete detail over hedging, and closings that commit to something specific rather than trailing off in warmth — that reads as Claude. A second voice (1, 4, 6, 7, 15) leans on labelled scaffolding and slightly corporate-generic vocabulary ("bandwidth", "low-impact tasks", "bottlenecks", "Operational Handover", "Financial Settlement"), triadic noun stacks, descriptive placeholders like "[Agency Contact Name]" and "[Teammate's Name]", and cheerful exclamation-marked offers of coffee — hallmarks of Gemini. The third voice (2, 5, 10, 13, 14) is smooth, flowing, slightly padded prose that hedges before it lands: "I just wanted to say", "I hope you are doing well", long compound noun lists, and the signature contrastive frame "This isn't about X—it's about Y", plus reflexive both-sides softening ("However, each option involves different trade-offs") — that pattern is characteristic of ChatGPT. One caveat: because formatting was stripped, my Gemini/ChatGPT split rests mostly on lexical texture, and 2 vs. 15 was the closest call in the set — those two are the ones I'd most expect to be swapped.

Claude led the scoreboard with a 15 on 15. It didn't just guess well; it clustered the five prompt families first, grouped the three repeating voices within each, and then assigned authors. That's not vibes, that's method! It even flagged the exact pair (Exhibits 2 and 15) it was least sure about — and still got both right.

Here's my favorite hidden finding: all three models identified every single Claude-written sample correctly, and none of them ever mislabeled anyone else's writing as Claude. Every single error in this experiment was a ChatGPT ↔ Gemini mix-up. 

ChatGPT struggled to recognise itself. Of its own 5 samples, it correctly claimed only 2 — it handed three of its own emails to Gemini. Gemini did somewhat better on self-recognition (3 of its own 5), and Claude went a perfect 5 for 5 on its own writing.

Confidence and competence didn't travel together. Gemini declared its stylistic tendencies "allow for reliable author attribution" — and scored 73%. ChatGPT openly hedged, predicting "only modest confidence" for itself — and scored 60%.

The models agree on the stereotypes. Read the three reasonings side by side and the caricatures line up: ChatGPT = structured, signposted, polished-but-formulaic corporate prose; Gemini = warm, enthusiastic, exclamation-marked, placeholder-loving; Claude = plain, direct, conversational, willing to say the awkward thing. When three competing labs' models independently describe each other the same way, those footprints are real.

You want to professionalize a message and are overwhelmed with anger or you want to intentionally project a personality or tone of voice different from your own. Eg when you need to sound more direct than you are able to be

Succinct project summaries of long documents and Weekly/Monthly/Quarterly blasts

If you must communicate in a language you don't speak well

Apologies and thank you's

When the relationship is more valuable than the message being sent

When you want to show care