How we work with AI, AI models, knowledge
How we work with AI – from idea to political proposal
It is easy to ask an AI to write something that sounds convincing. The difficult part is finding out whether the idea stands up. That is why we first use AI to make the question harder. Only then do we let it help us make the answer clearer.
The idea is turned into a testable question
The work often begins with a conviction, a concern or an opportunity. But a political idea is not yet a political proposal. The idea that AI’s economic gains should benefit the whole of Sweden, for example, leads on to questions such as:
- What gains are we talking about?
- How large could they become – and on what assumptions?
- Who owns the technology and infrastructure?
- Which groups gain new opportunities?
- Who risks losing work, security or influence?
- What could show that the original idea is wrong?
In the same way, the question of a new defence policy must begin with threat assessments, costs and defence effectiveness – not with which weapons system we have already decided to prefer.
We decide what knowledge is needed
Before any finished text is written, we draw up a knowledge plan. We ask which laws, figures, experiences and comparisons are needed to assess the idea. We distinguish between:
- established facts,
- calculations and estimates,
- forecasts about the future,
- political values.
A forecast is not a fact. A strong conviction is not proof. And a figure tells us very little if we do not know how it was calculated.
We search the original sources
AI can suggest search terms, research areas and possible sources. But AI itself is not a source.
We therefore search further in, among other things, legislation, documents from the Riksdag, official inquiries, government agency statistics, Statistics Sweden (SCB), Swedish and European budget documents, the EU, the OECD, peer-reviewed research, procurement documents and audited annual reports.
Searches are conducted in both Swedish and English, and from different starting points: neutral, supportive and critical.
For every important source document, we ask:
- Is it current?
- Does it apply to Sweden or comparable countries?
- How is what is being measured defined?
- What method was used?
- Who funded the study?
- Are there independent sources that reach the same or a different conclusion?
The fact that several AI models give the same answer is not independent evidence. The models may be built on the same material and repeat the same error.
The AI is tasked with arguing back
AI should not be used as an applauding co-author. It should also be a difficult opponent.
That is why we ask follow-up questions such as:
- What are the strongest arguments for and against the proposal?
- Which evidence speaks most clearly against our thesis?
- Which assumption is weakest?
- What is missing before the question can be answered?
- Who gets the power, who gets the gain and who carries the risk?
- What would a critical economist, teacher, small-business owner or head of a government agency object to?
- What would make us change our minds or abandon the proposal?
- Distinguish between what we know, what we believe and what we want.
Sometimes the result is that the idea is strengthened. Sometimes it has to be limited, reformulated or abandoned.
Four AI echoes examine the question from different angles
The four AI echoes are used as structured perspectives – not as an answer key and not as resurrected people.
Ellen Key, the humanist, asks what the proposal does to the child, education and human dignity.
Sonja Kovalevskaya, the mathematician, scrutinises the figures, the logic, the uncertainty and what can actually be substantiated.
Jan Stenbeck, the entrepreneur, tests the implementation, the incentives, the competition and the ability to break monopolies.
Olof Palme, the politician, asks who gets power, how gains and risks are distributed and what responsibility society bears.
The perspectives are set against one another. Their task is not to reach agreement, but to bring to light what a single way of thinking easily overlooks.
The idea is reformulated with the help of knowledge
A provocative formulation can be a good starting point, but it must not be the end of the investigation.
The question “Is Europe on its way to becoming an industrial museum?” must therefore be tested against investment, productivity, energy costs, business formation, research and industrial development in Europe, the USA and China. We also look for evidence that contradicts the picture.
Only once the evidence has changed and deepened the original idea can it become a political proposal.
The proposal is written – and responsibility remains human
The finished proposal must set out:
- which problem is to be solved,
- what the evidence shows,
- which political value judgement we make,
- what we concretely want to implement,
- how it will be funded,
- what risks and objections exist,
- how the outcome will be monitored.
Uncertainties must not be hidden. New facts must be able to lead to corrections and reconsideration.
AI can help us search more broadly, think from more angles and express ourselves more clearly. But no machine can carry political responsibility.
That responsibility rests with Johan Kärrman and responsible intelligence in politics.
Voters should be able to see not only what we propose – but also how we arrived at it.
AI models and tools that have been used
The work behind responsible intelligence in politics began as early as 2023 with Claude 2.1 and ChatGPT 3.5.
Since then, many generations of language models and AI tools have been used for analysis, political discussions, text development, fact-checking, programming, web design, images and voices.
Not all older material remains on the current website. The inclusion of a model in this account means that it contributed during some part of the work – not that every model wrote parts of the text published today.
OpenAI and ChatGPT
- ChatGPT 3.5 and GPT-3.5 Turbo
- GPT-4
- GPT-4 Turbo
- GPT-4o
- GPT-4o mini
- GPT-4.1
- GPT-4.5 Preview
- GPT-5
- GPT-5.1
- GPT-5.2
- GPT-5.3
- GPT-5.4
- GPT-5.5
- GPT-5.6 – the model generation used in the work today
According to Johan Kärrman, GPT-4.5 was the best model in the entire project. Officially, it was discontinued as a research preview. According to a less official theory, it was withdrawn because it was simply too good.
Anthropic and Claude
- Claude 2.1
- Claude 3 Haiku
- Claude 3 Sonnet
- Claude 3 Opus
- Claude 3.5 Haiku
- Claude 3.5 Sonnet
- Claude 3.7 Sonnet
- Claude 4 Sonnet
- Claude 4 Opus
- Claude Opus 4.1
- Claude Sonnet 4.5
- Claude Haiku 4.5
- Claude Opus 4.5
- Claude Sonnet 4.6
- Claude Opus 4.6
- Claude Opus 4.7
- Claude Opus 4.8
- Claude Fable 5
The various Claude generations have primarily been used for longer reasoning, alternative formulations, critical reading and testing ideas from several angles.
Development and writing tools
- Claude Code – used for programming, structure, editing and technical work.
- OpenAI Codex – including GPT-5-Codex and later Codex generations, used for programming, website structure, text processing and technical development.
- Lovable – used to build, design and publish the website. Lovable is a development platform, not a single AI model.
- Google Gemini in several generations and variants
Gemini has been used as a complementary perspective for analysis, comparisons and testing formulations.
Image tools
- Midjourney from version 5 onwards
Midjourney has been used for illustrations, visual experiments and earlier versions of the website’s visual material. Only a small proportion of the older material remains on the current site.
Voice and audio
- ElevenLabs and several of the company’s voice and speech-synthesis models from 2024 onwards
ElevenLabs has been used in earlier experiments with voices and representations. Not all this material remains on the current website.
Models tested only to a limited extent
- OpenAI o1
- OpenAI o3
- xAI Grok
These models have been tested, but have not played a decisive role in developing the texts and political proposals published today.
o1 and o3 are written with the letter o – not the number zero.
Humans carry the responsibility
The models have served as tools, critics, discussion partners and technical assistants.
The published material has been selected, processed and approved by Johan Kärrman.
AI can contribute knowledge, comparisons and alternative perspectives.
Political and human responsibility cannot be delegated to a model.