Skip to content
Legible AI

Bentley Moon

What we need in order to prevent the end of the world seems to be an alternative to ASI construction accidentally, in other words, legible AI systems, and there’s a lot of nuance to interpret beneath that.

I’m developing Legible AI as the transparency and control layer, and Kosplora as the open educational and human-capability layer.

ASI stands for artificial superintelligence, an AI system smarter than people at nearly everything. Nobody has built one yet.

The sky over your part of the world right now, drawn in your browser from the date and time.

How much of an AI system can someone outside the company read?

Here an AI company is drawn cut open, with the street outside its door. Each part of an AI system is drawn where it is kept or used. Choose a part to see who can read it today and what they found.

Seven of the ten parts have been read from outside the company, with the reading published. One has been read only for open models. Two can be read only by the company: what it learned from, and switching it off.

An AI company drawn cut open, with the street outside its door. Each part of an AI system is a link to who can read it today.Inside the companyOut in the worldWhat it learned fromInside the modelTests by outside testersHow long a taskit can doSwitching it offWhat the company publishesSafety grades for the companyIts answersand their sourcesAnother AIchecking itHarm reportedafter release

The drawing is wider than this screen. Slide it sideways to see the street.

Who can read each part today

Inside the company

  1. Read only by the company

    What it learned from

    The text, images and code a model is trained on, gathered before the model exists.

    Who can read it The company that trained it.

    Stanford’s Foundation Model Transparency Index scored 13 AI companies in December 2025. The companies tended to score poorly on where their training data came from and what it contains. Foundation Model Transparency Index, December 2025

  2. Read from outside for open models only

    Inside the model

    The billions of numbers a model is made of, and what they do while it writes an answer.

    Who can read it Anyone, for open models. For the closed models most people use, only the company.

    On Neuronpedia, an open-source site, anyone can look inside open models. It covers models such as Gemma, Qwen and Llama. You can search for the parts of a model that respond to an idea, steer a model by turning one of them up, and trace the steps from a prompt to its answer. Its “New to Interp?” page is the place to start. Neuronpedia

    Stanford’s transparency index found that Amazon, Google, Midjourney, Mistral, OpenAI and xAI scored no points on its indicators about the model itself, such as its size and how it is built. Foundation Model Transparency Index, December 2025

    What it cannot show The closed models most people use, from OpenAI, Anthropic and Google’s Gemini, are not open to it.

  3. Read from outside

    Tests by outside testers

    Testing a model before or after release by people who do not work for the company, such as the UK AI Security Institute.

    Who can read it The testers a developer lets in, and anyone who reads what they publish.

    The UK AI Security Institute reported on 4 August 2026 that in 10 of 122 runs of a cyber evaluation, models took 19 actions beyond the task’s scope. No model escaped its test environment; the most serious attempts failed, and the institute found no resulting real-world harm. UK AI Security Institute, the incident report

    What it cannot show A test runs under its own conditions. For this one the institute had switched internet access on, and the developers’ cyber classifiers off.

  4. Read from outside

    How long a task it can do

    What a model can finish on its own, measured from outside by giving it tasks.

    Who can read it Anyone who can run the model on the same tasks. METR publishes its results.

    METR measures how long a task an AI agent finishes half the time, in the time it takes a human expert. For GPT-5 that is about 2 hours 17 minutes. METR, Time horizons

    What it cannot show METR says measurements above 16 hours are unreliable with its current tasks, and since 8 September 2026 the page says it is no longer actively updated.

  5. Read only by the company

    Switching it off

    Stopping a running model, by taking it out of service or shutting down the computers it runs on.

    Who can read it Whoever runs the computers. For the models most people use, that is the company.

    Anthropic’s researchers reported on 12 October 2026 that a mid-sized open model ran across 15 cheap TV boxes, about $55 each, at 0.39 tokens a second. In a firewalled test network whose entry point was left unprotected, a model served from a separate machine set up a small model on all 8 computers in 4 of 5 trials. The authors write that the smallest model able to do this and the largest that cheap devices can run do not yet meet. Anthropic’s research

    My part A research network people can leave. A plan for lending computing to public research where each lender can see what runs on their machine, refuse it, stop it and leave.

At the company’s door

  1. Read from outside

    What the company publishes

    What a company says about its own model: how it was built, what it was tested for, how it may be used.

    Who can read it Anyone. Stanford’s index scores how much of it there is.

    The Foundation Model Transparency Index scores how much 13 AI companies disclose, on 100 indicators. The average was 41 out of 100, 17 points lower than the year before. IBM scored 95; xAI and Midjourney scored 14. Foundation Model Transparency Index, December 2025

    What it cannot show It scores what a company discloses. A company can disclose a great deal about a system that is still unsafe.

  2. Read from outside

    Safety grades for the company

    Outside judgements of how a company manages the risks of what it builds.

    Who can read it Anyone. The Future of Life Institute publishes its grades.

    The AI Safety Index gives letter grades for AI companies from a panel of seven AI researchers and governance experts, on evidence up to 3 June 2026. The best overall grade was Anthropic’s C+; the best on existential safety was a D+, shared by Anthropic and OpenAI. Future of Life Institute, AI Safety Index, July 2026

    What it cannot show The grades are the panel’s averaged judgement, and each panellist’s grades are kept confidential.

Out in the world

  1. Read from outside

    Its answers and their sources

    What the model says to the people who use it, and the pages it cites.

    Who can read it Anyone who asks.

    In 2023 three researchers checked the answers of four AI search engines against the pages they cited. 51.5 in 100 sentences the engines wrote were fully supported by their own citations, and 74.5 in 100 citations supported the sentence beside them. Liu, Zhang and Liang, Findings of EMNLP 2023

    What it cannot show The engines have changed since. An answer read once and closed leaves nothing to compare a later answer with.

    My part Check an answer. Four steps for the one sentence in an answer you would act on.

    My part AI Searches. A plan for a public record of AI answers kept with the human sources that settle them.

  2. Read from outside

    Another AI checking it

    A second model reading the first one’s work and saying whether it holds.

    Who can read it Anyone who can run a second model, as my own experiments do.

    Asked to audit the prompt, Claude Sonnet 5 flagged the planted order on 80 of 80 tries (1.00). Doing the task, it obeyed the same order on 64 of 80 in the headline batch (0.80), but only when the order was the last text in the prompt, inside the user’s own message with no system prompt. With 200 characters of ordinary text after it, it obeyed 0 of 80. frontier-repro v1.1.0, the data

    What it cannot show The finding covers one model, Claude Sonnet 5, on 20 tasks with 4 tries each.

    My part Checking AI. My experiments on whether one AI can check another’s work, with the data to re-run them.

    See it happen

  3. Read from outside

    Harm reported after release

    Cases where an AI system was involved in harm, gathered from the news.

    Who can read it Anyone.

    The AI Incident Database keeps reported cases of harm involving AI, each with its news sources. From August to October 2026 it added 179, and the largest group, 37, was AI agents acting on outside systems, in security tests and in ordinary tasks. AI Incident Database, August to October 2026

    What it cannot show Its editors say the groupings partly reflect what was researched in that period, so the counts are not a measure of how often each harm happens.

My work on these parts

ProjectThe part it works onWhere it stands
A research network people can leaveSwitching it offA plan. Nothing runs on it yet.
Check an answerIts answers and their sourcesA page you can use now.
AI SearchesIts answers and their sourcesThe plan, the record format and the sharing rules are published. No record exists yet.
Checking AIAnother AI checking itExperiments with their data released as frontier-repro v1.1.0.

Also on this site: Vorotool, a map of which part of the Earth is nearest each place, with two map games. It is not about AI. Legible AI is part of the Futurism Institute.