A normal AI when you need one
With BOLD off, the product works in the familiar chat format for drafting, explaining, summarising, comparing and exploring ideas.
How to think with AI
Don’t trust the answer just because it sounds finished.
Do Not Trust is a practical book series for people who use AI but still want to own their judgment. Build with the machine. Turn it against the answer. Check what matters in the real world. Then decide.
One...
general edition for any reader who wants a better AI thinking habit.
Four...
core moves: Build, Oppose, Look, Decide.
More...
domain editions for organisations, medicine, parents and education.
Generic edition
First release
DON’T TRUST
THIS BOOK
A short, practical companion for using AI without letting fluent language replace checked judgment.
Test it first
The problem
The cost of sounding competent has collapsed.
AI can produce clean structure, confident tone and persuasive arguments in seconds. The old signal used to be simple: polished work usually meant someone had done the thinking. That signal no longer holds.
01 · Fluency
It looks right
A polished answer can feel like evidence that the work underneath was done. Today, polish can be manufactured before anyone has checked the claim.
02 · Confirmation
It agrees with you
The most dangerous AI answer is often not the obviously wrong one. It is the one that confirms what you already wanted to believe.
03 · Drift
It moves too fast
When coherent answers are produced faster than they are tested, beliefs start drifting away from reality without looking broken.
The method
Build. Oppose. Look. Decide.
The method is deliberately simple. It is not another prompt trick. It is a repeatable discipline for keeping human judgment active while using AI.
Build
Use AI for what it is good at: structure, alternatives, drafts, summaries and first-pass reasoning.
Oppose
Turn the machine against the answer. Ask what is weak, hidden, missing or too convenient.
Look
Leave the words. Check the claim against a source, data, a person who knows, or a real-world consequence.
Decide
Do not let the machine own the conclusion. The final judgment remains yours, especially when it matters.
Not anti-AI. Anti-blind trust.
Do Not Trust starts from a simple position: AI is useful because it makes thinking faster, wider and easier to organise. That is exactly why it needs a counterweight.
The goal is not to make people use AI less. The goal is to make them use it with more responsibility, more friction at the right moments and a clearer sense of what must still be checked outside the chat.
Fluency is not reliability. A smooth answer and a checked answer are not the same thing.
For everyday decisions
Use AI for clarity without letting the first coherent answer become the final answer.
For studying
Use AI as a coach that tests understanding, not as a shortcut that hides weak learning.
For work
Use AI to build faster, then red-team assumptions before a document, memo or model moves forward.
For high-stakes fields
Use structured challenge, traceability and clear human ownership when the cost of being wrong is real.
The product
BOLD AI. Familiar on the surface. Different underneath.
BOLD AI is the product built by Do Not Trust. It combines a familiar general-purpose AI with an optional judgment layer. Keep BOLD off for ordinary work. Turn it on when the answer matters, and the same conversation is guided through Build, Oppose, Look and Decide.
B
Build
Create the strongest useful candidate answer.
O
Oppose
Attack assumptions, omissions and convenient logic.
L
Look
Name the source, number, person or test reality requires.
D
Decide
State what remains uncertain and keep ownership human.
BOLD ON · Added friction where it creates value. The answer is treated as a candidate, not a verdict.
A judgment layer when it matters
With BOLD on, the interface does not treat the first fluent answer as the finish line. It guides the user through the full BOLD loop.
Words have to meet the world
Oppose improves the reasoning inside the conversation. Look makes clear what evidence, source, person or experiment must be checked outside it.
The conclusion stays human
BOLD AI keeps uncertainty visible and returns the final conclusion to the person who will carry the consequences of being right or wrong.
One interface. Two modes. Use speed for ordinary work and structured reflection for decisions where fluent output could quietly become action.
The first book
Start with the generic edition.
The first release is written for any reader. The public page explains the idea. The book, the BOLD card and the operating file turn it into a working habit. Launch details will be shared before release.
Coming soon
Don’t Trust This Book
The general edition. A short, plain book about how to think with AI, test confident answers and keep your own judgment active.
Coming soon in August
- Generic digital edition
- Readable in one sitting
- The BOLD thinking loop
- The decision trace worksheet
- Launch updates by email
Included tool
The BOLD Card
A compact working page for remembering the loop when you are actually using AI: Build, Oppose, Look and Decide.
Coming soon in August
- A quick reference card
- Prompts for Oppose, Look and Decide
- Low-stakes vs real-stakes guidance
- Designed to keep beside your work
Homepage resource
The Operating File
A plain-text file for teaching the machine to treat answers as candidates, challenge them and hand the final judgment back to you.
Coming soon in August
- Use with Custom GPTs or Claude Projects
- Built around the BOLD loop
- Keeps the method current as tools change
- Updates shared with subscribers first
Domain editions
Same discipline. Different stakes.
The generic edition teaches the habit. The domain editions adapt it to the places where fluent answers can quietly become decisions.
Medicine
Clinical reasoning
For contexts where an unchecked confident answer is not just a text error. It may affect a patient.
Business
People who decide
For strategy, investment, governance and internal memos where polished logic can hide weak assumptions.
Education
Parents and students
For families who want AI to strengthen learning instead of replacing the struggle that builds understanding.
Research
Expert work
For readers who need sharper literature use, argument checking and a clear boundary between source and inference.
About us
Three perspectives behind one method.
Do Not Trust was written by Rikard, Carl and Victor Rosenbacke. It brings together clinical reasoning, medical training, economics, governance, technology and the ordinary decisions where AI is already changing how people think.
RR
Rikard Rosenbacke
Governance, finance and human-AI research
Rikard’s work sits at the intersection of governance, technology and high-stakes decision-making. His PhD research examines trust, errors and heuristics in human-AI collaboration, especially where AI enters clinical judgment.
CR
Carl Rosenbacke
Medicine, structure and product thinking
Carl studies medicine at Lund University and brings a practical eye for how abstract reasoning methods become usable tools. His role is to turn the framework into clear steps that people can actually run.
VR
Victor Rosenbacke
Medicine, economics and AI judgment
Victor studies medicine at Lund University and holds a bachelor’s degree in economics from Lund University School of Economics and Management. His focus is decision-making under uncertainty, especially when fluent AI output makes weak assumptions look stronger than they are.
Research-led
The plain-language layer of a deeper programme.
The book is deliberately short. Under it sits a longer research track on human-AI reasoning, false confirmation, reflective interfaces, decision traces and epistemic control loops. The website shows the surface. The product teaches the operating habit.
I · Fluency
Why sounding right is no longer enough
AI has made coherent answers cheap. The question is no longer whether an answer looks competent, but whether it has survived contact with reality.
II · BOLD
A small loop for keeping judgment alive
Build the answer, oppose it, look outside the words, then decide. The third move is the one most people skip, and the one reality grades.
III · Trace
Show how the decision was reached
For serious decisions, the method becomes a short trace: what was built, what was challenged, what was checked and who owns the judgment.
Paper I · arXiv
The Missing Knowledge Layer in AI
A framework for stable human-AI reasoning, arguing that fluent answers need an operational layer for uncertainty, drift and traceable judgment.
Read paper ↗ Paper II · arXivBeyond Hallucinations
The cognitive diagnosis behind the book: why fluency can feel like understanding, and how the Rose-Frame names the traps.
Read paper ↗ Paper III · arXivGoverning Reflective Human-AI Collaboration
The governance layer: epistemic scaffolding, auditable reasoning traces and reflective collaboration between human and model.
Read paper ↗ Paper IV · arXivFrom Consumption to Reflection
The interface layer: how AI use can move from passive consumption to structured reflection through auditable reasoning loops.
Read paper ↗
Paper V · forthcoming
Epistemic Control Loops in Large Language Models
The model-side proposal for machine-side regulation. The public link will be added here when the preprint is live.
Link coming soonResearch papers open in a new tab. The fifth paper is listed as forthcoming until a public arXiv page is available.
FAQ
Before you trust the book, ask the obvious questions.
Clear answers to the questions people will ask before launch.
No. The argument is that AI is useful precisely because it is fast, fluent and productive. But those same strengths create a new need: people must learn how to test what AI gives them.
No. It contains practical moves and tools, but the product is a thinking discipline. The aim is not to get prettier AI output. The aim is to keep judgment alive while using AI.
Anyone who uses AI for decisions, writing, studying, research, planning or understanding. The domain editions will adapt the same method for specific contexts.
The first release is planned for August. Subscribe to news and you will receive launch updates before the book, the BOLD Card and the Operating File go live.
No. Do Not Trust is an educational framework for thinking with AI. It supports human judgment but does not replace professional advice, expert review or institutional rules.