Standard Intelligence Limited
Effective from 2 August 2026 | Version 1.0 | Last reviewed 2 August 2026
Why this page exists
Article 50 of Regulation (EU) 2024/1689 (the AI Act) requires that people are told when they are interacting with an AI system, and that AI-generated content is marked and, in defined cases, labelled. Those obligations apply from 2 August 2026.
The disclosures that satisfy Article 50 appear where the interaction happens: in the product, on the image, next to the article. This page is the register behind them. It sets out every surface we operate, what AI does on it, who reviews the output, and which limb of Article 50 we have applied.
We are a UK company. We publish this because our systems are made available to users in the European Union, and because a governance vendor that will not document its own AI use is not worth buying from.
1. Scope and our roles
Article 50 assigns different duties to providers and to deployers. We hold both roles, on different surfaces.
As a provider (Article 50(1) and 50(2)) we design and place on the market AI systems that interact with users and generate content. The design-level duties to signal AI interaction and to apply machine-readable marking sit with us.
As a deployer (Article 50(3) and 50(4)) we use third-party AI systems in our own operations, including marketing and content production. Where that use produces content within scope, the duty to label sits with us.
We are not a provider of general-purpose AI models. We build on models supplied by third parties and are a downstream provider of the AI systems listed below.
Where processing happens. Inference for our products runs in the European Union. The model providers we use are named in our subprocessor agreement, and we do not add or change one without notifying customers under the terms of that agreement.
Our own Article 4 position. Article 4 of the AI Act requires providers and deployers to ensure a sufficient level of AI literacy among their staff. Every member of Standard Intelligence staff has completed the Literacy programme we sell, against the same standard our customers are held to. We would not ask a customer to buy an obligation we had not met ourselves.
Related documents. This page covers transparency about AI. Personal data is covered by our Privacy Notice. Customer data processing and the model providers we use are covered by our Data Processing Agreement and Subprocessor Agreement.
Surfaces under development. We are building further product surfaces with more advanced AI capabilities. They are not in the inventory below because they are not available to anyone yet. They will be added here, with their classification, before general availability rather than after.
2. Surface inventory
| Surface | AI involvement | Human review | Article 50 limb | Disclosure mechanism | Reviewed |
|---|---|---|---|---|---|
| Marketing website | Drafting assistance for some copy; some illustrative images generated with AI; no autonomous publication | Every page and image reviewed and approved by a named person before publication | 50(4) considered for text, out of scope where editorial control applies; 50(2) marking applied to generated images | Footer statement linking here; visible label on generated images | 2 August 2026 |
| AI Regulatory Navigator | AI system interacting directly with users; generates text grounded in our regulatory corpus | Corpus authored and reviewed by us; individual responses are generated | 50(1), 50(2) | In-product notice at first use; response-level provenance indicator | 2 August 2026 |
| Literacy (Article 4 courseware) | Synthetic narration audio; AI-assisted drafting of learner-facing text | All content reviewed and approved before release | 50(2) | Statement on each module landing screen; machine-readable marking on audio | 2 August 2026 |
3. Human review and editorial control
Article 50(4) requires deployers to label AI-generated or manipulated text published to inform the public on matters of public interest, unless that text underwent human review or editorial control and a person holds editorial responsibility.
For everything we publish on regulation, compliance or the AI Act, editorial responsibility sits with the Chief Executive Officer. Review means a person reads the full text against source instruments, corrects it, and approves publication. Nothing goes out on an automated schedule without that step.
Where we have used AI in drafting, we say so. Where we have not reviewed something, we do not publish it.
4. Marking and detection of AI-generated content
Article 50(2) requires providers of generative AI systems to mark synthetic output in a machine-readable format so it can be detected as artificially generated.
What we apply
- Synthetic audio in Literacy carries C2PA Content Credentials embedded at generation.
- Generated text in Navigator and Platform carries a provenance indicator in the interface.
- Images generated for marketing use carry C2PA Content Credentials where the generating tool supports them.
Limits we will state plainly
Machine-readable marking survives normal distribution. It does not reliably survive screenshotting, re-encoding, transcoding or deliberate stripping. Detection tools return probabilistic results, not proof. We do not claim our marking makes content tamper-proof, and no current method does.
Timing
Article 50(2) as amended by the Digital Omnibus allows until 2 December 2026 for marking on generative AI systems placed on the market before 2 August 2026. Systems we place on the market on or after 2 August 2026 comply from the date they ship. Any surface operating under the December window is noted in the inventory above.
5. Accuracy and limitations
Our products can be wrong, and you should treat their output accordingly.
Navigator generates answers from our regulatory corpus. Platform extracts obligations, proposes classifications and assembles evidence records. Both produce output that is generated rather than retrieved verbatim, and generated output can misstate a requirement, miss an applicable instrument, apply a provision that has been amended, or read a source correctly and summarise it wrongly.
Verify against the source instrument. Every substantive output cites the provision it derives from. That citation is there to be followed, not to reassure you. Before you rely on any classification, mapping or obligation statement in a regulatory filing, a submission to an enforcement body, or a decision with legal consequences, check it against the instrument itself.
What we are not. We are not a law firm and we are not your lawyers. Standard Intelligence is not authorised or regulated by the Solicitors Regulation Authority, or by any other legal services regulator in any jurisdiction, and we do not carry out reserved legal activities within the meaning of the Legal Services Act 2007 or their equivalent elsewhere. We do not provide legal advice, and nothing our products produce is legal advice however closely it reads like it.
Two consequences worth stating. No relationship between you and us is a solicitor-client relationship, so nothing you put into our products and nothing they produce attracts legal professional privilege. And where a matter needs a regulated opinion, you need a regulated adviser; we can tell you what an instrument says, and we cannot tell you what you should do about it in a way you can rely on as advice.
A classification recommendation is an input to your judgement, not a substitute for it. Your organisation remains the decision maker for its own compliance, and the certification step in Platform exists to make that a deliberate act rather than an assumption.
Corpus currency. Regulation moves. Our corpus is versioned, every record shows the corpus version it was produced under, and we publish the date each instrument was last reviewed. Output produced under an older corpus version is not automatically wrong, but it is not evidence that nothing has changed.
If you find an error, tell us at ai-disclosure@standardintelligence.com. Corrections to the corpus are tracked and released under the process at section 6.
6. Human oversight
The principle we build to is that a person decides and the system assists. Two places make that real rather than rhetorical.
Changes to the corpus and to product logic. No material change to the regulatory corpus reaches a customer without a named person reviewing and approving it. That covers new or amended instruments, changes to obligation extractions, changes to archetype definitions, changes to the mappings between them, and changes to the logic that produces classification recommendations.
Review means a person with the relevant regulatory competence reads the change against the source instrument, accepts, amends or rejects it, and records the decision. AI proposes; it does not merge. Approval is recorded against the named reviewer, the corpus version and the date, and that record is retained so any output can be traced back to the change that produced it and the person who approved it.
Decisions about a customer's own estate. Standard Intelligence is a regulatory information product. It produces recommendations from what a customer tells it about their own systems, and that input varies in accuracy and completeness. Incomplete or inaccurate input changes the recommendation, sometimes materially.
For that reason we do not produce a classification decision record, and nothing in Platform should be read as one. Producing that record, and the rest of the risk analysis lifecycle around it, is the customer's responsibility, taken with their own legal counsel.
A recommendation does not become a record, and does not carry a customer forward into documentation or evidence production, until a named and competent individual at that customer has reviewed it and accepted it. Acceptance is a positive act, never a default, and it is recorded against that person. Our recommendations are an input to a customer's compliance position. They are not a determination of it.
Human oversight measures for any high-risk system we place on the market are documented in the technical file for that system under Article 14 of the AI Act, and are available to customers under diligence.
7. Assessment in Literacy
Literacy separates the assessment that produces your record from the assessment that helps you learn. The first never involves AI. The second does, and what it is allowed to affect depends on the mode your organisation has selected.
7.1 The certification assessment
Knowledge checks and a deterministic rubric produce your completion record, and where your employer has designated you for a human oversight role, your competence evidence for it. Marking is done by rule against a fixed answer key. No AI system marks it, nothing about the outcome is inferred or predicted, and a human assessor will review any result you dispute.
This is the only assessment that gates anything, in every mode.
7.2 Assessment modes
Organisations buying Literacy choose one of three modes. The choice determines who marks scenario work, what it affects, and what the organisation takes on as a deployer under the AI Act.
| Mode | Who marks scenario work | Effect on your record | AI Act position for the buying organisation |
|---|---|---|---|
| Optional scenarios (default) | Nobody. Scenarios are practice, submitted for your own reflection, with model answers you compare against | None | No high-risk AI system is deployed. No deployer obligations arise from assessment |
| Human-graded | A qualified human assessor, with AI assisting the assessor rather than deciding. The assessor is accountable for the mark | Contributes to your record | The decision is not solely automated. The organisation deploys no high-risk assessment system |
| Agentic scoring | An AI system, with a human escalation and appeal route as the oversight safeguard | Contributes to your record | The organisation deploys a high-risk AI system under Annex III of the AI Act and takes on deployer obligations, including human oversight, logging and informing affected workers |
The default is deliberate. Most organisations do not want an automated assessment affecting a member of staff's competence record, and the default mode means they never have to think about it.
7.3 What is disclosed, in every mode
You are told which mode applies before you begin. Where AI is involved in marking or in assisting an assessor, that is stated on the assessment screen and every AI-generated mark or comment carries a label. Where a human assessor decides, you can ask who, and you can appeal.
7.4 What your employer receives
Completion records: who passed, which module, on what date, and any acknowledgement of a named human oversight role. This is not optional, because the Article 4 literacy obligation sits with the employer and they need evidence of who was trained and to what standard. Learners are told before they begin.
In the default mode, practice scenario work is yours alone and forms no part of that record.
7.5 Classification
Agentic scoring evaluates learning outcomes and contributes to a competence record, so it falls within Annex III of the AI Act and we treat it as a high-risk AI system with Standard Intelligence as provider. Standalone Annex III obligations apply from 2 December 2027 under Regulation (EU) 2026/1744.
The default and human-graded modes do not put a high-risk assessment system into a customer's hands. Our determination for each mode is documented and available to customers under diligence.
8. What we do not do
We do not operate emotion recognition systems. We do not operate biometric categorisation systems. Article 50(3) does not apply to any surface we run.
We do not generate deepfakes. We do not produce synthetic image, audio or video content depicting real people, whether public figures, customers or staff, in any context.
We do not use AI to make decisions about individuals in credit or in access to services.
There are two places where AI touches an outcome about a person, and we would rather name them than let the sentence above imply otherwise. Literacy assessment, covered at section 7. And recruitment: we use LinkedIn Recruiter as our applicant tracking system, and it applies AI to search, matching and ranking. Nothing is rejected by it. Every application is read by a person, every candidate progresses on a human decision, and hiring follows a full interview process run by people.
We do not train models on customer content submitted to Platform or Literacy. That holds for our own models and for the third-party models we build on, and it is written into our processing terms rather than being a policy we could change quietly.
One thing we do do, stated here so it is not a surprise. Customers who opt in contribute to industry benchmarks. What contributes is aggregate data about AI system inventories and the risk classifications a customer has assigned to them. Not the systems themselves, not their documentation, not any content a user has entered.
The output looks like this: your inventory is currently 34% high risk, against a sector average of 21%. The exchange is reciprocal and built into how the feature works: a customer who contributes can see where they sit against their sector, and a customer who does not contribute cannot. There is no view of the benchmark that does not involve being in it.
Nothing reaches a published benchmark until the sample behind it meets a k-anonymity threshold of 10, so that no figure can be traced back to a contributing customer. Where a sector or sub-sector is too thin to clear that threshold, the benchmark is not published for it, and we do not merge sectors to manufacture a sample.
Almost none of this is personal data. It describes AI systems and risk classifications, which are facts about a customer's estate rather than about people. Where any element of it is personal data, the contribution is governed by the processing terms in your contract.
Opting in is a positive election, never a default. Withdrawal is available at any time. It removes you from every benchmark calculated afterwards, and it also switches off your own comparison view, since the two are the same mechanism. Nothing else about the service changes. Figures already published stand, because a distribution that has gone out cannot be recalculated, and at a threshold of ten contributors your figures were never separable from it in the first place.
Literacy contributes nothing to any benchmark. Learner results are not pooled, not aggregated into sector figures, and not compared against other learners. There is no ranking, no leaderboard, no percentile and no badge.
9. Code of Practice
The Commission published a Code of Practice on Transparency of AI-Generated Content in June 2026 as a voluntary route to demonstrating compliance with the marking and labelling obligations. We are adopting its technical measures and intend to sign by 30 September 2026. This page will be updated on signature.
Where we label AI-generated content in the interface, we use the Commission's optional icon set.
10. Accessibility
Article 50(5) requires this information to be provided clearly and distinguishably at the latest at the point of first interaction or exposure, and to meet applicable accessibility requirements. Our in-interface disclosures are readable by assistive technology, carry text alternatives where an icon is used, and meet WCAG 2.2 AA. If a disclosure is not reachable for you, tell us and we will fix it.
11. Questions and complaints
Write to ai-disclosure@standardintelligence.com. Queries reach the Chief Technology Officer and are answered within 10 working days.
If you believe we have failed an Article 50 duty, you can also raise it with the market surveillance authority in your Member State.
12. FAQ
Am I talking to a person or an AI?
There is no chatbot on this website. Anyone who replies to you from Standard Intelligence is a person. Inside the AI Regulatory Navigator you are querying an AI system, and it says so before you type anything.
Was this page written by AI?
Drafted with AI assistance, then read line by line against the AI Act and approved by a named person before publication. Same process as everything else we publish on regulation.
How do I tell whether an image or audio clip on your site is AI-generated?
Visible label next to the asset, using the Commission's icon set. Where the file format supports it, there is also machine-readable provenance data embedded in the file, which you can check with any C2PA-compatible tool.
Does Platform or Navigator make decisions about me?
No. They produce analysis and records about AI systems and regulatory obligations, which a person then acts on. Neither makes an automated decision about an individual.
Is what I put into your products used to train models?
No. Not ours, not anyone else's. See section 8, and the data processing terms in your contract, which govern.
So what are the industry benchmarks built from?
Aggregate counts of AI systems and the risk classifications customers have assigned to them, from customers who opted in. That is what produces a line like "your inventory is 34% high risk against a sector average of 21%". No system documentation, no user-entered content, and nothing published until the sample is large enough that no contributor can be identified from it.
We have not opted in. Are we in the benchmarks?
No, and you cannot see them either. Contributing and viewing are the same switch. Withdraw later and you drop out of every benchmark run after that and lose the comparison view; figures already published cannot be recalculated.
I am a Literacy learner. Who sees my results?
Your employer receives your completion records: which modules you passed and when. They need that because the Article 4 literacy obligation sits with them, and we cannot switch it off at your request or theirs. In the default mode, practice scenario work is yours alone and is not reported. Other learners see nothing about you, and nothing you do goes into any benchmark or comparison.
Is my certification assessment marked by AI?
No, in every mode. It is marked by rule against a fixed answer key. AI marking applies only to scenario work, and only where your organisation has selected a mode that uses it. Section 7 sets out the three modes and how to find out which one applies to you.
Is my Literacy coursework marked by a machine?
Scenario and practicum submissions are graded by an AI system, and you are told before you start. If you disagree with a grade, ask for human review using the link shown next to it. A person then re-marks the submission and their decision stands.
Can I rely on what Navigator or Platform tells me?
Not without checking. Both generate output, and generated output can be wrong. Every substantive answer cites the provision behind it, and that citation is there for you to follow before you rely on anything in a filing or a decision with legal consequences. See section 5.
Are you a law firm?
No. We are not regulated by the SRA or any other legal services regulator, we do not carry out reserved legal activities, and nothing we produce is legal advice. Nothing you put into our products is privileged. Section 5 sets out what that means in practice.
Who approves changes to your regulatory corpus?
A named person with the relevant competence, on every material change, against the source instrument. The AI proposes; it does not merge. Section 6 sets out how that works and what is recorded.
I think I have found AI content you have not labelled. What now?
Tell us at ai-disclosure@standardintelligence.com with a link. We will check it, correct it if you are right, and record the correction in the version history below.
Why is there so much detail here?
Because we sell AI governance software, and a disclosure that says "we sometimes use AI" would not survive our own product.
13. Version history
| Version | Date | Change |
|---|---|---|
| 1.0 | 2 August 2026 | First publication, on the date Article 50 became applicable. |
We review this page quarterly and on any material change to a surface in the inventory. Determination records supporting each classification are held internally and available to customers under diligence on request.
We tell customers when this changes. Any material change to this page is notified directly to current customers, at the contact address held for them, at the same time the change is published. Material means a new surface, a change to what AI does on an existing surface, a change to a classification, or a change to what an employer or learner receives. Cosmetic edits are logged here without notification.
Surfaces under development. New product surfaces with more advanced AI capabilities are in build. They will be disclosed here, with their classification and their Article 50 mechanism, before general availability.