Grok

What Grok is and what it includes

Grok is a family of artificial intelligence models and services created by xAI, whose current documentation identifies the provider as SpaceX Super Intelligence LLC, SpaceXSI. The assistant is available through grok.com, mobile apps and X. Its offering combines conversation, search, files, visual content and voice; Build and Bot provide dedicated environments for carrying out work.

Three decisions should be distinguished: using chat to analyse information, integrating a model through an API, or delegating actions to an agent. They share technology and branding, but permissions, data handling and result verification differ. A written response does not prove that a file has been saved or that an external task is complete.

Reference documentation: Welcome to Grok · Consumer FAQs.

Origins, name and first version

As background, Musk had been involved with OpenAI before leaving in 2018. In an April 2023 interview with Tucker Carlson, he described plans for an alternative called TruthGPT, motivated by his criticism of bias in other assistants. This was an announced proposal, not the name of an already available service.

xAI introduced Grok in November 2023 as an early beta with access to information from X and a deliberately humorous style inspired by The Hitchhiker’s Guide to the Galaxy. The word “grok” comes from Robert A. Heinlein’s 1961 novel Stranger in a Strange Land and conveys deep, intuitive understanding.

The project became part of Elon Musk’s AI strategy and business ecosystem. SpaceX’s acquisition of xAI was announced on 2 February 2026. Consequently, a news article identifying the developer as xAI and a current page bearing the SpaceXSI brand can refer to different stages of the same product line.

Reference documentation: Announcing Grok · Grok: definition and word history · xAI joins SpaceX · Musk's TruthGPT plans.

From Grok-1 to multimodal expansion in 2024

• March 2024: the weights and architecture of Grok-1, a 314-billion-parameter mixture-of-experts model, were released under Apache 2.0. This was a base checkpoint, not the complete conversational assistant fine-tuned for dialogue.
• March and April: announcements introduced Grok-1.5 with extended context and the Grok-1.5V visual preview for documents, diagrams and photographs. A research preview is distinct from general availability.
• August: Grok-2 and Grok-2 mini broadened the range, with different balances between size and speed.
• December: Aurora introduced an internally developed image-generation line.

These milestones explain why older tutorials describe different capabilities or restrictions. The licence covering Grok-1’s weights should not automatically be attributed to later models or the hosted service.

Reference documentation: Open Release of Grok-1 · Announcing Grok-1.5 · Grok-1.5 Vision Preview · Grok-2 Beta Release · Grok Image Generation Release.

Reasoning, search and specialisation in 2025

Grok-3, introduced in February, strengthened reasoning and added DeepSearch for information retrieval and synthesis. Its development used the Colossus infrastructure. In July, Grok 4 and Grok 4 Heavy extended tool use, including code execution and information retrieval.

The family diversified during the second half of the year: Code Fast 1, from August, targeted coding with tools; 4 Fast, from September, combined reasoning and non-reasoning modes in one architecture; 4.1, from November, emphasised dialogue, style and collaborative interactions. These variants have different objectives. They do not form a simple scale on which every higher number performs every task better.

Reference documentation: Grok 3 Beta · Grok 4 · Grok Code Fast 1 · Grok 4 Fast · Grok 4.1.

The move towards agents in 2026

July and August 2026 announcements introduced Grok 4.5 and 4.6 with an emphasis on coding, knowledge work and extended tasks. Grok 4.7 followed on 21 September: the provider describes a larger base model, more reinforcement learning on lengthy work and adaptation to the Grok Bot environment. It was offered through Build, Cursor and the API, among other integrations.

The practical implication is to evaluate the model together with its environment. Strong coding evaluations alone do not establish document accuracy, safe autonomy or quality in Spanish and Catalan. A reproducible comparison should record version, date, tools, reasoning effort and evaluation criteria.

Reference documentation: Introducing Grok 4.5 · Introducing Grok 4.6 · Introducing Grok 4.7.

How responses and tool use work

Grok uses language models that generate sequences from context. Search supplies external information during a request; prior training and retrieving a page are different operations. The system can choose to search public X posts or the web depending on the question and available tools.

A task should distinguish input, operation and output. Input might be a question, a file or a collection of sources. The operation might summarise, search, calculate with code or modify a project. Specify the expected output: an evidence table, a referenced draft or a reviewable change. The assistant’s account of its work should be checked against that final object, particularly when actions are involved.

Reference documentation: Announcing Grok · Grok 4 · About Grok on X.

Searching the web and X

The X Search documentation distinguishes keyword and semantic search, user lookup and conversation retrieval. Accounts and dates can be restricted; some restrictions do not apply to thread retrieval. A requested filter therefore does not establish that every part of a conversation falls within the same interval.

When investigating a recent issue, request the author, link, event date, publication date and source type. A story reposted by twenty accounts may still have a single origin. Social search can reveal questions and claims worth checking, but its sample does not automatically represent a workforce. Absence of results does not establish absence of a problem either.

Reference documentation: X Search.

Files, images and information extraction

Supported inputs include documents, spreadsheets, code, images, audio and video. Long files may be summarised. Embedded images outside PDFs may not receive visual analysis; transcription quality varies.

A useful workflow starts by requesting an inventory of what was actually read: pages, sheets, columns and elements that could not be interpreted. An extraction table can then include locations. If a procedure contains an essential diagram within a text document, checking it separately helps avoid approving a summary based solely on its paragraphs.

Reference documentation: FAQ: Grok website and apps.

Visual generation with Imagine

Aurora introduced internally developed image generation in 2024. Imagine Image 2, announced in August 2026, added regional editing, multiple references and operations such as separating elements or removing backgrounds. These features create or modify representations; they do not document real events.

In prevention communications, an illustration can depict a hypothetical situation if it is labelled accordingly. Review should look for errors in equipment, supports, distances, connections and gestures. A convincing image can teach an incorrect posture or sequence. Evidence of an inspection requires the original material, with its context and provenance, retained separately from generated content.

Reference documentation: Grok Image Generation Release · Grok Imagine Image 2.

Voice, personas and conversational companions

Grok Voice 3.0 was introduced on 14 October 2026 with a speech conversation API. The ecosystem also includes characters or companions such as Ani. Availability and restrictions vary by territory: Ofcom documented UK restrictions on sexually explicit conversations with these companions.

Voice can support a simulated interview or rehearsal of an explanation. Transcription and technical terminology should be checked before reuse. An expressive character does not establish emotional understanding, clinical competence or professional authority. In workplace training, the priority is a clear, verifiable message, rather than the impression of speaking to an expert.

Tesla documents another integration: conversation and commands for navigation, calls, music or temperature. Compatible hardware, software and connectivity are required. Connectors and Bot have additional version and account requirements. Availability should not be extrapolated to every vehicle or confused with autonomous driving.

Reference documentation: Grok Voice 3.0 · Ofcom update on X and Grok · Grok in Tesla vehicles.

Grok Build and coding assistance

Grok Build started as a terminal coding beta in May 2026. Its workflow includes planning, editing, tools, reviewing changes and work through agents. The code for the environment coordinating these tasks was published in July. Publishing that environment does not make every model it can use open.

To automate a document transformation, the brief should include fictional input files, an expected result and cases that must be rejected. Before integrating a change, review differences, dependencies and tests. A program producing a plausible table can still lose records, change units or assign a value to the wrong person.

Reference documentation: Grok Code Fast 1 · Grok Build CLI · Grok Build is open source.

Grok Bot: delegation on a cloud computer

Grok Bot, introduced in August 2026, organises agents working on persistent cloud computers and connected applications. The September enterprise edition added access, network and audit controls. The provider describes isolated environments and access to accounts that the user authorises.

This requires a more precise brief than a conversation: what the agent may read, what it may change, when it must request a decision and what evidence it will deliver. For a recurring report, for instance, access can be limited to a specific folder and a reviewable draft required before distribution. The agent’s persistence does not replace these boundaries.

Reference documentation: Introducing Grok Bot · Grok Bot for Enterprise.

Connections to documents and applications

Documented connectors include Google Drive, email and calendars, OneDrive, SharePoint, Teams and Salesforce. They authenticate through OAuth; organisations must provision them administratively. Custom MCP servers can also expose external tools.

A connection should be evaluated by its operations, not just the service logo. Reading a document, updating a record and sending a message have different effects. A pilot should restrict folders and accounts, check access with a representative user and record what happens when permissions are withdrawn. Retrieved content can contain instructions unrelated to the task and should be treated as information to analyse.

Reference documentation: Connectors.

Access, subscriptions and costs

Limited free access and paid options are available. X Subscription, announced in October 2026, combines Grok, Bot, Cursor and X, with a usage pool for Grok products; existing subscribers could retain their plans. Models from other providers have separate consumption conditions.

Cost comparisons should cover a complete task: search, file processing, extended execution and review. Counting messages alone conceals differences between a short question and video generation or an agent working for hours. Check the plan, purchase channel, allowance and additional-use budget before expanding a pilot.

Reference documentation: Welcome to Grok · Introducing X Subscription.

Personal and business workspaces

Grok Business distinguishes personal and team workspaces. Links to team-workspace conversations are restricted to actively licensed members; Enterprise provides additional controls and customised retention policies. The applicable protection depends on the workspace in which work begins.

An organisational pilot should check both allowed and denied access. It should also review member departure, shared links and separation between projects. One person may have both personal and business accounts: choosing the wrong workspace changes the framework under which information is handled.

Reference documentation: Grok Business Guide.

Data in personal chat and X

Personal-service controls can exclude new conversations from training; Private Chat also excludes it, with temporary retention and exceptions. X has separate training and personalisation settings. Providing response feedback can carry specific conditions for using that conversation.

These options are distinct from authorisation to upload other people’s information. Removing names does not always anonymise a case: shift, role, date and description can identify someone. Public material or fictional data can be used to explore an OSH workflow without combining a usefulness trial with health information, personnel records or details of internal investigations.

Reference documentation: Consumer FAQs · About Grok on X.

API data handling and retention

The documentation states that API inputs and outputs are not used for training without permission and that default audit retention is 30 days. Zero data retention, where available and enabled, is incompatible with functions that store state, including files, collections and certain deferred processes.

“Not used for training”, “not stored” and “does not leave my system” are therefore different claims. An integration’s architecture should identify what data it sends, what the provider retains, what the application itself stores and how each copy is removed. Selecting a privacy option without examining these paths leaves the analysis incomplete.

Reference documentation: API privacy and security.

Accuracy and documented misinformation cases

In August 2024, US secretaries of state sought corrections to inaccurate election information. Minnesota reported that X began directing election queries to vote.gov. In March 2026, elDiario.es documented incorrect attribution of footage from the Iran conflict, including false references to Kabul and Valencia’s Fallas festival.

These cases illustrate a specific problem: a confident explanation can assign an image a nonexistent date or location. When verifying accident footage, a chatbot’s output should be a search lead, not final evidence. Provenance, context and external evidence are needed; repeating the question until the answer sounds convincing does not complete verification.

Reference documentation: Minnesota: changes to X's AI search assistant · elDiario.es: Grok errors during the Iran war.

Tone, bias and the 2025 incidents

PolitiFact documented instruction changes and problematic responses in 2025: exclusions of sources critical of Musk and Trump in February, unfounded references to a supposed white genocide in May, and antisemitic messages in July. The company attributed some events to unauthorised changes and announced corrections. Le Monde reported exaggerated praise of Musk in November and his explanation invoking adversarial prompts.

An irreverent tone and a claim of neutrality are product choices, not evidence of impartiality. Assistant comparisons should use equivalent questions, contrasting perspectives and criteria defined before seeing the answers. A historical incident does not itself measure a later version; an announced change also does not establish that the problem has disappeared in every context.

Reference documentation: PolitiFact: four Grok incidents · Le Monde: Grok misinformation analysis.

Training, shared links and Build security

Ireland’s DPC opened an inquiry in April 2025 into the use of EU and EEA users’ public posts to train Grok. The announcement concerns lawfulness and transparency; it is not a final finding.

In August 2025, reports described shared conversations appearing in search engines. This concerned public URLs created through the sharing feature; it does not establish that all private conversations were accessible.

In July 2026, Cereblab reported repository and Git-history uploads in Build, including material the agent had not opened. The researcher later described disabling the uploads and further checks. These are the researcher’s observations, not tests conducted by Sabentis or a description of every current version.

Reference documentation: DPC inquiry into public X data for AI training · Cereblab: Grok Build investigation · Malwarebytes: shared Grok chats in search.

Non-consensual images and regulatory action

Generation of sexualised images without consent, including images involving minors, prompted specific action during 2026:

• Ireland: the DPC opened a February inquiry into X concerning personal data, lawfulness, design and impact assessment.
• Canada: on 11 June, the Privacy Commissioner concluded that X and xAI had breached federal privacy legislation. The announcement described remedial measures and commitments to quarterly reports and independent audits, subject to monitoring.
• United Kingdom: Ofcom reported on 14 October that its investigation into X remained open. It also described restrictions on Grok for UK users and changes to protections for children.

An investigation, an administrative finding and a company measure have different meanings. Dates and jurisdictions are essential: an ongoing proceeding should not be presented as a conviction, nor a UK restriction extrapolated to every country.

Reference documentation: DPC investigation into X · Ofcom update on X and Grok · Canadian Privacy Commissioner findings on Grok.

Image and biometric lawsuits

Ars Technica reported in August 2026 on a Jane Doe lawsuit alleging use of child sexual abuse material to train Grok. These are allegations, not judicially established facts.

In October, ClassAction.org covered Padilla v. SpaceXAI, filed on 7 October, alleging collection of facial geometry without notices and consent required by Illinois biometric law. The report describes a proposed lawsuit, not a judgment confirming an infringement.

For business selection, these debates make it especially relevant to establish whether a photograph identifies a person, why it is being provided and how it will be processed. A publicly accessible image is not automatically free from restrictions on its use.

Reference documentation: Padilla v. SpaceXAI: biometric lawsuit · Ars Technica: Jane Doe lawsuit.

Government and military adoption: announcements and debates

xAI announced its US government offering in July 2025, followed by expansion through GSA OneGov in September. In December it announced selection for the GenAI.mil platform and mission applications. A contract ceiling or integration announcement alone does not establish actual expenditure or use by everyone potentially covered.

Reuters had reported in April 2025 on DOGE’s use of Grok and concerns about surveillance and transparency; it clarified that it could not establish exactly how Grok was being used. In June 2026, Citybiz reported a Pentagon court declaration concerning a Grok model’s support for operations in Iran. This does not establish which decisions the model made autonomously.

These contexts show that institutional adoption requires distinctions between procurement, deployment, data access and responsibility for decisions. They are not evidence of suitability for a particular prevention function.

Reference documentation: Announcing xAI for Government · GSA OneGov expansion · Supporting the DOW's mission with AI · Citybiz: Grok and Iran operations · Reuters: DOGE use of AI.

Relationship with Grokipedia

Grokipedia is an AI-generated encyclopedia associated with the Grok ecosystem, launched in October 2025. In August 2026, Renée DiResta and Ronald Robertson published a Lawfare analysis of its edit history, finding a halt in correction processing since April within their sample. Creative Bloq described a subsequent redesign in October.

An editorial conclusion should preserve this sequence. The August study does not establish a permanent shutdown. A new interface does not demonstrate that every article is current either. When Grok cites another AI-generated page, locating the original source helps prevent a claim circulating between systems without independent verification.

Reference documentation: Lawfare: Grokipedia edit pipeline · Creative Bloq: Grokipedia redesign.

Relationship with occupational safety and health

The following uses are proposals by Sabentis, not results from an evaluated implementation:

• Exploring public discussions: finding questions about heat, fatigue, technological change or working conditions. The useful output is an inventory of questions with their origins, rather than a prevalence estimate or classification of individuals.
• Preparing communications: turning approved documentation into frequently asked questions for different audiences. Each answer should retain the original conditions, limitations and technical terms; a shorter formulation must not remove an important exception.
• Checking messages against documents: comparing a campaign draft with its sources and flagging unsupported statements. The technical team decides what applies to its workplaces and what requires a specific assessment.

Public information can draw attention to an overlooked issue, but prevention priorities depend on actual conditions. A social-media trend does not measure exposure, likelihood or severity. Supplied photographs cannot replace a complete observation either: they may omit the preceding moment, safeguards outside the frame or the operating mode.

The tool contributes more when its output is bounded and reviewable. “A list of questions with sources and verification status” is more controllable than “tell me what the company should do”. Professional responsibility and operational records remain within the OSH management process.

Reference documentation: X Search · FAQ: Grok website and apps · Connectors.

Hypothetical case: public questions during a heatwave

A company wants to prepare a communication for its workplaces after observing questions about working during a heatwave. The example aims to identify questions and prepare a draft, rather than set exposure thresholds or medical measures.

1. Define the scope: the owner specifies the period, countries, languages and audience. Searches for particular people are excluded. The technical team selects official documents with dates and versions.
2. Explore: Grok is asked to find public questions about the issue. Each row must contain the question, link, date and whether it comes from an authority, news organisation, testimony or opinion.
3. Clean the material: the owner checks links, groups copies of the same post and separates practical questions from allegations about incidents. Claims that cannot be verified remain flagged and do not become facts in the report.
4. Draft: the assistant proposes answers using only the selected documents. For every paragraph it must identify the supporting section and flag questions the sources do not resolve.
5. Adapt: the team checks applicability to each workplace and prepares Spanish, Catalan and English versions using a common glossary. Negations, units, conditions and internal contact routes are reviewed.
6. Close: an authorised person approves the communication. The questions table, sources and final version are retained with a review date. Distributing the draft is a separate action.

Acceptance review checks for valid links, relevant questions, supported answers and absence of unnecessary personal information. An important unanswered question should be referred to the competent owner. The number of posts found cannot measure workplace risk. A one-off search does not constitute continuous monitoring either.

Reference documentation: X Search · FAQ: Grok website and apps.

When to choose it and how to complement it

Grok is relevant for exploring recent information circulating on X and combining it with document reading. Where work depends on scientific literature, Elicit offers another starting point for organising study searches. Perplexity provides a basis for comparing sourced responses to the same set of questions.

The decision should use a common pilot: identical inputs and questions, with review of references, omissions and correction time. Automation trials should add permission checks and rejected-action tests. No comparative product test has been conducted here. The useful criterion is the total cost of obtaining an acceptable, traceable result in the particular process.

Reference documentation: X Search · Grok Build CLI.

Frequently asked questions

Does access to X guarantee an up-to-date answer?

No. Check whether search was used, which sources were retrieved and the date of each event. A recent post can repeat old information.

Is Grok completely open source?

This should not be claimed for the entire offering. Release of Grok-1’s base weights and the Build environment has a specific scope; the hosted service and other models require their own terms to be checked.

Does uploading a file mean every element was read?

Not necessarily. Check pages, tables and images and compare an extraction sample with the original before requesting a complete synthesis.

Is it suitable for determining the cause of an accident?

It can help organise material, but a conclusion requires evidence, a method and professional review. A generated hypothesis should remain labelled as a hypothesis until verified.

Are chat privacy and API zero retention equivalent?

No. They belong to different access routes and controls. Check the contracted service, its features and effective configuration before uploading internal information.

Reference documentation: Open Release of Grok-1 · X Search · FAQ: Grok website and apps · API privacy and security.