# Methodology

## What this chart is

This map shows **our interpretation of selected public statements about AI**. Select a person or group to see the statements we used, when they were made, and why we placed that point there. The map is not a survey, an ideology score or a scientific risk estimate.

The research and writing were prepared with AI assistance. The atlas has not received an independent human fact-check. Read the linked sources and qualifications when drawing a conclusion about someone.

The map separates belief, policy, organizational identity and observed behavior in a 2 × 2 scatter plot.

## Horizontal axis: preferred pace of frontier capability growth

Here, **frontier capability growth** means making the most advanced AI systems more capable. The left side represents calls to slow or limit that development, including conditions, targeted pauses or a moratorium on superintelligence. The right side represents continued or faster development. The center represents continuing under conditions.

This axis is **not** the volume of product releases, capital spending, market adoption, open-weight availability or the rate of deploying existing systems. Faster deployment of existing models can coexist with slower growth in frontier capabilities. An actor can also favor faster safety research and slower capability progress at the same time.

Approximate editorial anchors:

| Anchor | Interpretation |
| --- | --- |
| -100 | Broad or enforced restraint on the relevant dangerous frontier |
| -50 | Explicit pacing, restrictions or slowdown advocacy |
| 0 | Conditional continuation; implementation or net preference may be ambiguous |
| +50 | Development with safeguards, without an established general pacing preference |
| +100 | Strong acceleration or opposition to catastrophe-driven restraint |

These anchors organize the canvas. They are not empirical units or a universal scoring rubric. Read the actor's **scope**: a conditional prohibition on ASI is not necessarily a ban on all AI.

## Vertical axis: public catastrophic-risk concern

The vertical axis reflects how strongly the reviewed public evidence emphasizes catastrophic or existential risk. It does **not** estimate a person's private beliefs or a numerical p(doom).

High concern can reflect precaution about severe consequences without a high assigned probability. Lower concern means relative skepticism or lower emphasis **in the reviewed position**, not absence of concern about discrimination, privacy, jobs, misuse, concentration of power or other harms.

A corporate risk framework establishes that catastrophic risks are explicitly considered. It does not identify a calibrated probability. Company placements therefore often carry wide vertical ranges.

## Why a nonempty lower-left quadrant is possible

Someone can oppose particular AI development or deployment because of existing harms, labor, rights, democratic accountability or concentrated power while remaining skeptical of AI-extinction scenarios. No names are plotted there in the initial curated sample. That is a coverage choice, not evidence that the viewpoint is empty.

## Point coordinates, uncertainty and evidence clarity

Coordinates use an internal range of `[-100, 100]`. Each point has `xRange` and `yRange` covering plausible alternative interpretations. These are **editorially assigned interpretive ranges**, not confidence intervals, probability distributions or survey error bars. Intersections with other quadrants are intentional when the evidence is ambiguous.

The default static graphic shows anchor points to remain readable. Selecting an actor in the interactive view reveals its range; the separate uncertainty SVG shows all ranges. Exact distances between nearby actors should not be interpreted as rankings.

`confidence` is displayed as **evidence clarity**:

- **High:** a direct and reasonably explicit statement supports the relevant public position. This does not certify the actor's conduct or the exact coordinates.
- **Medium:** a defensible synthesis, but interpretation, scope, institutional variation or reporting limits remain.
- **Low:** provisional placement, material ambiguity or a limited proxy such as a brief endorsement or historic organizational advocacy. Hollow/dashed markers distinguish these points.

There is no statistical model behind those labels. The position and its range should be reviewed together.

## Evidence rules

Every plotted actor must have evidence addressing **pace** and **concern**, plus a caveat and scope. Context-only evidence cannot satisfy an axis requirement.

Prefer first-person statements, company policies, original interviews and published proposals. Use attributed reporting when the original is inaccessible, and label it as reporting. Never convert an article headline or a search-result snippet into a directly verified first-person statement. A quote embedded in reputable reporting is still indirectly verified unless its original is read.

`checkedOn` records when the source was read, subject to the reading scope in `retrieval.scope` and `notes`, when present. An inherited record keeps its original review history; including it in a later build does not mean it was read again. `published` and `updated` retain their real precision: a month does not become an invented day. Reading a 2023 essay in September 2026 does not make it a new 2026 statement.

Selected source records now state what was read separately from links to original posts that could not be retrieved. A supplied export is a copy, and its author field does not independently authenticate authorship. Original-post links are not additional verified sources. The source panel shows this distinction before opening the linked original.

Archive notes distinguish a reviewed copy from an unsuccessful check. “No archived copy verified” can mean a lookup was blocked or rate-limited; it does not mean no copy exists. If there is no archive metadata, no check is recorded. An archive lookup does not update the source's reading date. The [agent guide](https://theaiatlas.org/agent-guide.md) describes these fields for automated readers.

`lastReviewed` is actor-level editorial review. It can be older than the current edition. Changing a page layout or running the build does not update this field.

No article text or source screenshot is copied into this repository. Summaries are original short annotations. External links remain subject to their publishers' terms.

## Separate entities and separate claims

A person is not their company. A research division is not necessarily the whole parent company. The following distinctions matter in this starting edition:

| Case | Treatment in the dataset |
| --- | --- |
| [Amodei](https://theaiatlas.org/evidence.html#entity-amodei) and [Anthropic](https://theaiatlas.org/evidence.html#entity-anthropic) | Separate records. The primary essay includes company commitments, but access for outside evaluators does not by itself establish a slowdown. Anthropic’s pace anchor is near the middle with a broad range; implementation is not independently verified. |
| [Altman](https://theaiatlas.org/evidence.html#entity-altman) and [OpenAI](https://theaiatlas.org/evidence.html#entity-openai) | A reported personal response is separate from OpenAI's own report of selective restrictions. No full-lab halt is inferred. |
| [Hassabis](https://theaiatlas.org/evidence.html#entity-hassabis) and [Google DeepMind](https://theaiatlas.org/evidence.html#entity-deepmind) | First-person proposals and a reported endorsement are separate from the institutional framework. |
| [Musk](https://theaiatlas.org/evidence.html#entity-musk) and [xAI / SpaceXAI](https://theaiatlas.org/evidence.html#entity-xai) | Musk's broader personal review combines direct interviews and an indirect endorsement, with a wide range. It does not establish a company-wide operational change; the organization remains in the review backlog. |
| [LeCun](https://theaiatlas.org/evidence.html#entity-lecun) and [Meta](https://theaiatlas.org/evidence.html#entity-meta) | LeCun's personal skepticism does not override Meta's explicit catastrophic-risk framework. |
| [Microsoft AI](https://theaiatlas.org/evidence.html#entity-microsoft-ai) and Microsoft | A named AI division's public code is not assigned to the entire parent company. |
| [Andreessen](https://theaiatlas.org/evidence.html#entity-andreessen) and [a16z](https://theaiatlas.org/evidence.html#entity-a16z) | The manifesto states that it expresses individual views, not the firm’s views. Andreessen’s personal advocacy is scoped separately; a16z is unplaced pending firm-level evidence for both axes. |
| [Trump](https://theaiatlas.org/evidence.html#entity-trump) and the US government | Reported personal statements and his signed order inform his record. This is not a separate placement of the administration or a finding that policy directions were implemented. |

Each entry has an `implementation` label such as public advocacy, published policy, or self-reported action. These are descriptions of the available evidence, **not an independent implementation audit**.

## What changed relative to the earlier informal graphic

The vertical axis is now public concern rather than an inferred private probability. Pacing and targeted restrictions are distinguished from shutdowns. People and companies have different shapes. Organizations with unreviewed axes are left off-chart. EA and AI safety are explained in the glossary rather than positioned as if they were single actors.

No earlier image is treated as a validated baseline. This repository's initial archive begins on 15 September 2026; movement from earlier images is not presented as measured history.

## Terminology is not a membership classifier

[EA](https://theaiatlas.org/evidence.html#idea-ea), [e/acc](https://theaiatlas.org/evidence.html#idea-eacc), [doomer](https://theaiatlas.org/evidence.html#idea-doomer), [decel](https://theaiatlas.org/evidence.html#idea-decel) and [AI safety](https://theaiatlas.org/evidence.html#idea-safety) are explained with their distinctions and original sources in the guide. They are not interchangeable categories. The chart does not automatically classify an actor into a tribe from its coordinate.

## From overview to research

### Start with the computer system

Our explanations start with [computer programs and their data](https://theaiatlas.org/evidence.html#idea-software), then distinguish the [model from the system that uses it](https://theaiatlas.org/evidence.html#idea-ai-system). This follows the software and system distinctions in the [NIST glossary](https://csrc.nist.gov/glossary/term/software) and [OECD explanatory memorandum](https://www.oecd.org/content/dam/oecd/en/publications/reports/2024/03/explanatory-memorandum-on-the-updated-oecd-definition-of-an-ai-system_3c815e51/623da898-en.pdf).

For language models, we separate programmed calculations, settings adjusted during training, and the surrounding application's tools and controls. The [system drawing](https://theaiatlas.org/evidence.html#llm-system) shows these parts with their sources. We describe the choices people make when building and operating an application. We assess capabilities, [claims about understanding](https://theaiatlas.org/evidence.html#idea-understanding), and effects on people using their own evidence; calling a system software is not our verdict on those questions.

### Choose how far to read

The site opens on the map, with links to the guide, search and source notes. Each glossary entry starts with a short summary. Open it for a fuller explanation, its relevance to the map, and important distinctions. References accompany each part. Reading routes suggest an order for learning. The timeline covers selected events, and connections between ideas explain our interpretation; neither gives a complete history or establishes someone's membership in a group.

Comparison presents the same evidence questions for two selected records. It does not compute a winner, combine institutional and personal records, or fill evidence gaps with inferred scores. Saved reading notes retain the selected claims' source links and caveats. Search and comparison remain independent of what is visible on the map.

The [LLM introduction](https://theaiatlas.org/evidence.html#llm-basics) distinguishes software execution, learned weights, interpretability, selection rules and reproducibility. Its [weather comparison](https://theaiatlas.org/evidence.html#llm-weather-analogy) is explicitly an analogy with stated limits. Each explanation is written at four reading depths that reword one sourced claim; a simpler depth leaves detail out and never substitutes a looser claim. The interactive decoding example uses invented probabilities and a fixed pseudorandom generator; it is a teaching illustration, not a measured model. Its in-depth temperature control reshapes those invented odds and does not make sampling deterministic. The accompanying [fixed-code example](https://theaiatlas.org/index.html#llm-lab) executes an ordinary multiplication in the browser. These demonstrations add no evidence about an actor's position.

Source citations open provenance, publication precision and retrieval notes before linking to the original. Some entries rely on a paper's abstract or a community retrospective; the registry identifies that scope. In particular, the [basilisk entry](https://theaiatlas.org/evidence.html#idea-basilisk) uses retrospective sources and does not claim to have retrieved the deleted original post. Philosophical arguments and speculative scenarios are presented as arguments, not established facts.

Factual explanations include references to the publications used. A reference lets you check an explanation; its presence alone does not establish that the explanation is accurate or balanced. Older saved editions preserve the evidence and format available when they were published.

## What automated checks establish

Checks reject malformed data, missing axis evidence, bad dates, dangling source references, out-of-range positions and label collisions. They also test rendering, escaping and the local server. Those checks establish structural and software properties only. They cannot prove that a source is accurate, that an interpretation is fair, or that a promised safeguard has been implemented.

## Comparing editions

Compare the underlying statements as well as the points. A changed position on the chart may reflect new evidence or a revised interpretation. It does not, by itself, show that a person or group changed their mind. Editions that use different axis meanings cannot be compared as if the scale stayed the same.