Field notebook No. 001 - opened 14 Sep 2026

Cluefishing

AI is one of the best collaborators most of us have ever had. This is the site for using it out loud: noticing the tell when someone doesn't say so, and disclosing proudly when you do.

✓ Disclosed

"This cover image was generated with an AI image tool, then hand-edited in Photoshop for the final crop and colour."

Reader knows what they're looking at. Nothing to dig for.
⚑ Undisclosed

"Shot this myself over the weekend :-)" - no camera metadata, six-fingered hand in the background, posted as personal work.

This is the shape of a catch worth logging.
AI News - updated weekly

What AI is actually doing right now

The tell we look for on this site only matters because the tool underneath is genuinely good. Here's some of what it's doing lately, worth disclosing proudly rather than hiding.

A map of every possible DNA typo

Google DeepMind released the AlphaGenome Atlas, precomputed predictions for how all 9 billion possible single-letter changes in the human genome might affect gene activity, free for researchers studying genetic disease to search rather than compute themselves.

Source: deepmind.google →

A handheld device that finds a vein under pressure

AI-GUIDE, built by MIT Lincoln Laboratory with Massachusetts General Hospital, pairs a portable ultrasound with AI guidance so someone with minimal training can catheterize a blood vessel at the scene of an injury. It just won the Federal Laboratory Consortium's 2026 tech-transfer award.

Source: news.mit.edu →

The most detailed brain map yet

Google Research, HHMI Janelia, and collaborators at Cambridge published a complete wiring diagram of the male fruit fly's brain and nerve cord, 166,000 neurons and 125 million synapses, the largest connectome mapped to date.

Source: research.google →

Last week's edition, still worth a read

Verified math proofs in Lean, AI-designed antibiotics that cured drug-resistant infections in mice, and a CRISPR copilot from Stanford, three more genuine steps forward from the 14 September edition.

See the 14 September edition →
This edition: 15 September 2026 Full weekly archive →
Manifesto

Use the tool. Keep the credit honest.

AI has quietly become one of the best collaborators most of us have ever had. It can help build something faster, translate an idea, clean up a draft, or generate a texture nobody had time to paint by hand. That's worth being glad about, not defensive about.

The tool was never the problem. It starts the moment a person lets someone else believe an idea, a voice, or a piece of work was entirely theirs when a machine carried real weight in it.

We heard that failure mode get a name in a recorded conversation about AI ethics: clawfishing, presenting an AI-made creation or template as your own without disclosure. It's a good word for a real problem, so we built the mirror image of it. Cluefishing is the practice of noticing the tell, and of disclosing your own work before anyone has to go looking for it.

The idea still has to come from a person. AI can carry it beautifully. It shouldn't get silent credit for originating it.

- Geir Amdal, founder
Source note

The word "clawfishing" comes from a taped interview about AI ethics on a channel that also covers psychic and future-timeline claims. We're borrowing one word from that conversation, not endorsing the channeling material around it. Treat that origin as a spoken quote from one source, not a verified fact.

Field guide

Three entries to start with

001  

The Clawfishing Problem

+
What the word means, and why the mirror image needed a name too.

Clawfishing is simple to define and easy to miss in the wild: an AI-generated creation, or something built from an AI template, presented as if it came entirely from the person posting it. No disclosure, no credit to the tool, no indication that a prompt did some of the work.

It's not the AI use that's the issue. It's the silence around it. A drawing, a paragraph, a song, or a product photo made with AI assistance is fine to share; it's the implied claim, I made this myself, unaided, that turns ordinary tool use into something closer to plagiarism.

Cluefishing starts on the other side of that line: reading for the tell, and building the habit of disclosing before anyone has to ask.

002  

Provenance Is Not Paranoia

+
Metadata, content credentials, and why "where did this come from" is a normal question now.

Photographers have used EXIF metadata for decades to prove when and how an image was captured. The same idea is now being extended to AI content through standards like the C2PA content credentials effort, an industry coalition (Adobe, Microsoft, the BBC and others are publicly involved) building a way to attach a verifiable "how this was made" record to a file.

None of this requires trusting a single company's word. It just means provenance, the trail showing how something was made, is becoming a normal thing to check, the way you'd check a source on a news story. Asking for it isn't an accusation. It's due diligence.

Practically: keep your prompts, your drafts, and your source files. If someone asks how a piece was made, the honest answer should already be sitting in a folder somewhere.

003  

Six Places the Tell Shows Up

+
What to actually look at before you file a report.

1. Missing metadata. A photo with no camera information, or a document with no edit history, isn't proof of anything by itself, but it's worth a second look.

2. Suspiciously even rhythm. Human writing wobbles: sentence length varies, tangents creep in. Text that marches at the same pace for paragraphs can be a sign, not a certainty.

3. Stock transitions. Phrases that show up in almost every AI-drafted piece, moreover, it's important to note, in today's fast-paced world, aren't wrong on their own, but three of them in one paragraph is a pattern.

4. Confident errors. Clean grammar paired with a factual mistake a knowledgeable person wouldn't make is a classic tell.

5. Image artifacts. Garbled text in the background, extra or fused fingers, jewellery that doesn't quite close, lighting that doesn't agree with itself.

6. No visible process. A finished piece with zero sketches, drafts, outtakes, or revision history behind it, where the creator's usual work does show that trail.

One of these is a maybe. Two or three together are worth logging in the field notebook below.

Tackle box

Disclosure checklist

Run your own work through this before you publish. Answers stay in your browser only.

The log

Shared field notebook

● Coming soon

Spotted an undisclosed tell somewhere public? Soon you'll be able to log it here for others to see. The form below is a preview, it isn't connected to anything yet, we want a small moderation layer in place first so this stays a shared notebook and never a way to name people without cause.

When this goes live: describe patterns, not people. No names, no doxxing, no links to private accounts.

EXAMPLEexample

Freelance listing profile photo

No EXIF data, and the ear on the left side doesn't match the one on the right. Portfolio has no other shots in a matching style.

EXAMPLEexample

"Handwritten" newsletter intro

Every paragraph opens with a personal anecdote followed by "but here's the thing", same rhythm four issues in a row, no typos, ever.