Tuesday, July 28, 2026Subscribe
Meridian48 — Tech news, summarised
AI, business, devices, policy — global tech, summarised every 30 minutes.
Daily puzzle · One guess per day

Guess The AI

See a paragraph. Guess which model wrote it: Claude 4.7, GPT-5, Gemini 3 Pro, Grok 4, or DeepSeek V4. Build your streak. Share your result. New puzzle every day.

Today's puzzle · Tuesday, July 28
Which AI wrote this?
Streak 0
Accuracy 0%
Played 0
The prompt
Summarise the key risks of using AI for medical diagnosis
The response
Three principal risks dominate the clinical AI literature. First, model performance often degrades when deployed against patient populations that differ from the training distribution, which means an AI validated in a US academic medical centre may behave unpredictably in a Pakistani public hospital. Second, the interpretability problem: clinicians cannot easily verify why a model reached a specific conclusion, which complicates both medical judgement and liability. Third, automation bias: when AI provides confident outputs, even experienced clinicians demonstrate measurable tendency to defer to the model's recommendation, including in cases where the recommendation is wrong. Research from the Royal College of Radiologists (2025) suggests the third risk may be the most underrated in current procurement decisions.

Pick the model that wrote it

Pick one to lock in your guess. You only get one try per day.

Samples are paraphrased examples of each model's characteristic prose style based on real outputs. We rotate puzzles daily from a curated pool. New entries are added every two weeks. Have a great example we should include? Email it in.

Why this game exists

After three years of using these tools every day, we noticed something: even experienced builders can't reliably tell which model wrote a given output. We default to whichever model we're paying for and stop noticing the texture differences entirely.

But the differences are real, and they matter for picking the right tool. Claude's instinct for balanced framing makes it the best general-purpose advisor. GPT's structural templates make it the cleanest at marketing copy and email drafting. Gemini's academic register suits research synthesis. Grok's irreverence works for opinion content where personality matters. DeepSeek's technical density is ideal for engineering documentation.

Playing this daily for a month genuinely teaches you to feel the difference. Then you'll start picking better.

Frequently asked questions

How does the daily puzzle work?

Every 24 hours, the puzzle rotates to a new paragraph written by one of five frontier AI models. You see the original prompt and the model's response. You have one guess. Get it right and your streak increases; miss and it resets to zero.

How are the samples written?

Each sample is a real (or representative) output from the model it credits, chosen to highlight that model's characteristic prose style. We rotate the pool every two weeks. Claude defaults toward balanced 'first X then Y' framing, GPT toward emoji-laced bullet structure, Gemini toward formal academic prose, Grok toward conversational irreverence, DeepSeek toward dense paragraph-level analysis. After a wrong guess we reveal which tell you missed.

Can I play yesterday's puzzle?

Not within this tool — yesterday's puzzle is gone once midnight UTC rolls over. This is deliberate: it keeps the streak meaningful. If you missed a day, your streak resets and you start fresh.

What if my guess is partially right (e.g. right vendor, wrong model)?

It counts as wrong. The five options are different enough that distinguishing Claude Opus from Claude Sonnet isn't part of the puzzle — only vendor-and-model-tier identification is.

Why this game?

Because most builders use one or two models without ever really getting a feel for the others' default registers. After playing for a few weeks, you'll genuinely start to recognise the texture of each model — which makes you better at picking the right tool for each task.

How is this scored / shared?

Your streak and accuracy live in your browser's localStorage; nothing is sent to a server. The 'Share' button copies a short result string (today's date, win/loss, streak) you can paste anywhere.

Related on Meridian48

The 48° Brief — Weekly

The week in AI, tech business, and the tools worth knowing.

One email, every Friday morning. Curated and written by Faizan Khan. No filler. No tracking pixels. Unsubscribe in one click.

Join the first readers of Meridian48.