AI Transparency Notice

Version: 0.1-draft Effective date: not yet in force (draft)

Draft pending review by a qualified Romanian lawyer before use.

In short

• When you chat in 1ads, you are talking to an AI system, not a person. For a human, write to support@1ads.net.
• AI proposes, people approve. The AI never changes your ad accounts and never publishes on its own. Changes run only after an authorised person in your team approves them.
• Generated images carry a digital "made with AI" marker (C2PA Content Credentials), and AI-generated text carries a machine-readable indicator when it leaves the Platform through exports, the API or the MCP server. When we publish to a network that offers an AI label, we switch it on.
• AI can be wrong. Check numbers, claims and facts before you approve.
• When you publish, labelling is your job too. Deepfakes, and AI text published to inform the public on matters of public interest, must be disclosed (section 6).
• We do not train AI models on your data.

1. Purpose and scope

1.1 This notice explains how the 1ads Platform, operated by ALTTAB PROFIT SRL, a limited liability company (the "Company"), uses artificial intelligence: which features use AI, which models and providers are involved, how content is marked, how humans stay in control, what the known limitations are and what the Customer must do when publishing AI-generated content.

1.2 It is written with regard to Article 50 of Regulation (EU) 2024/1689 (the "AI Act"), as amended by Regulation (EU) 2026/1744, whose transparency obligations apply from 2 August 2026, and with regard to the GDPR and the rules of Third-Party Platforms. The Platform was placed on the market after 2 August 2026, so the transitional period in Regulation (EU) 2026/1744 for systems already on the market does not apply to it.

1.3 This notice is the information referred to in Section 7.7(f) of the Terms (terms-of-service.md). It does not create obligations beyond those in the Terms and the AUP (acceptable-use-policy.md); if it conflicts with them, the Terms and the AUP prevail. Capitalised terms have the meaning given in the Terms. In this notice, "AI Features" means the Platform features that use AI models, and "Outputs" means content, analyses, proposals and other results generated by the AI Features. This notice is information for the Customer; it is not legal advice.

2. Who is who under the AI Act

2.1 The Company is the provider of the 1ads AI system. We build it on general-purpose AI models supplied by Anthropic and OpenAI (the providers of those models) and offer it under our own name.

2.2 The Customer is the deployer when it uses the AI Features under its own authority in its professional activity, in particular when it publishes Outputs or uses them in ads.

2.3 Agencies. An agency that uses the Platform for its clients is a deployer for the content it creates and publishes. Agencies and their clients should agree in writing who applies the disclosures described in section 6.

2.4 White-label. Agencies using white-label instances must keep all AI disclosures and labels provided by the Platform. An agency that offers the Platform to third parties under its own name or trademark may itself take on obligations of a provider under the AI Act and should seek its own legal advice. We will provide the technical disclosures and labels the white-label instance needs.

2.5 Not a high-risk system. The Platform is designed for marketing and advertising work. It is not intended for the high-risk uses listed in Annex III of the AI Act and must not be used for practices prohibited by Article 5 of the AI Act (see acceptable-use-policy.md, sections 7.4 and 7.5). In particular, the AI Features must not be used to place, target or optimise job advertisements, or to screen or filter job applications (Annex III, point 4(a)); these uses are excluded from the intended purpose. Performance Guardian and the AI chat do not make audience, targeting or bidding proposals for campaigns flagged as employment campaigns. [TBC: product to implement the employment flag and block before launch.]

3. Which features use AI

3.1 The table below lists the AI Features and the human control that applies to each.

3.2 Reporting and ROI. ROI, profit (including COGS, refunds and cash-on-delivery refusals) and attribution figures are calculated with deterministic formulas from the data you connect and configure. Incrementality estimates use statistical methods and come with uncertainty. The AI chat may explain or summarise these results, but it does not replace the calculations.

3.3 Changes to models. Model providers regularly release new versions. We may change the models used for a feature to maintain quality, cost or safety. The current list is shown in the Platform and in the sub-processor list (subprocessors.md); changes to sub-processors are notified as described in the Data Processing Agreement (data-processing-agreement.md).

3.4 Where AI processing happens. Anthropic processes data in the USA; OpenAI processes data in the USA, and in the EU where EU data residency applies [TBC]. Transfers rely on the EU-US Data Privacy Framework where the recipient is certified, otherwise on the Standard Contractual Clauses (Decision (EU) 2021/914) and a transfer impact assessment. We have requested zero data retention from Anthropic. Model providers may otherwise keep inputs and outputs for a limited period for abuse monitoring under their terms; details are in subprocessors.md and privacy-policy.md. EU-only AI processing ("EU-strict" mode) is planned for Enterprise customers but is not available yet.

3.5 No training on your data. The Company does not use Customer Data or Outputs to train or fine-tune AI models. Under the commercial API terms that currently apply to us, Anthropic and OpenAI do not use API inputs and outputs to train their models.

3.6 Tokens stay out of AI. OAuth tokens and other credentials for your connected accounts are encrypted and are never sent to AI models.

3.7 Features of Third-Party Platforms. AI features offered by Third-Party Platforms themselves (for example automated bidding or automatically generated assets inside an ad platform) are provided by those platforms, not by the Company, and are not covered by this notice.

4. You are interacting with an AI system (Article 50(1))

4.1 The chat assistant is clearly identified in the interface as an AI assistant. It is not a person. Human support is available at support@1ads.net.

4.2 Drafts and proposals created with AI Features are flagged as AI-generated in the Platform, including in the client approval portal, so that client approvers know what they are reviewing. [TBC: confirm the label is implemented in the client approval portal and in white-label instances.]

4.3 Notifications sent by automations (for example a Performance Guardian alert) are identified as automated.

5. How AI-generated content is marked (Article 50(2))

5.1 Images. Images generated by the Platform carry C2PA Content Credentials: cryptographically signed metadata stating that the image was generated or edited with AI, by which tool and when. OpenAI also embeds its own C2PA metadata in images its models generate.

5.2 AI labels on social networks. When you publish through the Platform (directly or through the publishing bridge Zernio) to a network whose API lets the publisher declare AI-generated content (for example Meta's "AI info" label, TikTok's AI-generated content label or YouTube's altered or synthetic content disclosure), the Platform sets that declaration for AI-generated media. Where a network's API does not support it, the Platform tells you to add a label manually.

5.3 Text. AI-generated text is labelled in the Platform. When it leaves the Platform through exports, the REST API or the MCP Server, it carries a machine-readable indicator (for example an ai_generated field, and C2PA or IPTC metadata in exported files). When it is published, the Platform sets the AI-content flag of the destination network where that network's API offers one. Once text is copied out manually, the marking may be lost, which is why your own disclosure (section 6) matters. [TBC: product must implement the text indicator before launch.]

5.4 Limits of marking. Metadata can be removed by editing tools, screenshots, compression or by some websites and networks. Marking supports, but does not replace, the deployer's disclosure. Intentionally removing markings to hide AI generation is prohibited (acceptable-use-policy.md, section 3.10).

5.5 Minor edits. Where the AI only performs an assistive function for standard editing (for example correcting spelling or resizing an image) and does not substantially change the content or its meaning, the content need not be marked as AI-generated.

5.6 Video and audio. The Platform does not currently generate video or audio. If we add these features, we will update this notice before they become available.

5.7 Code of Practice. We use the European Commission's Code of Practice on Transparency of AI-Generated Content (final version published on 10 June 2026) as a reference for our marking and labelling approach. [TBC: whether the Company signs the Code; plans for additional marking layers such as imperceptible watermarking and a detection interface.]

5.8 Prohibited imagery. The image generation features include safeguards that prevent the generation of non-consensual intimate imagery and child sexual abuse material (Article 5 of the AI Act, as amended). [TBC: describe the safeguards, for example input and output classifiers and blocked prompts.]

6. Your obligations when you publish AI content (Article 50(4))

6.1 Deepfakes must be disclosed. If you publish AI-generated or AI-manipulated images, audio or video that resemble existing people, objects, places, entities or events and would falsely appear to your audience to be authentic, you must clearly disclose that the content is artificially generated or manipulated. The test is how the audience would perceive the content, not what you intended.

Examples for marketers:

Even where content is not a deepfake, a Third-Party Platform may require an AI label. Follow the stricter rule.

6.2 Creative, satirical or fictional work. If the deepfake is part of an evidently artistic, creative, satirical or fictional work, the disclosure is still required but may be made in a way that does not spoil the work. As a matter of 1ads policy, we treat commercial use of a real person's likeness as outside this lighter treatment.

6.3 AI text on matters of public interest. If you publish AI-generated or AI-manipulated text to inform the public on matters of public interest (for example articles or posts on health, the economy, public policy or news), you must disclose that it is AI-generated, unless the text has undergone human review or editorial control and a person or company holds editorial responsibility for its publication. That review must be substantive (checking facts and content), not a quick glance. Ordinary product ads and promotional copy are usually not "text on matters of public interest", but they remain subject to advertising law and Third-Party Platform rules. The Platform's approval log records who reviewed and approved each item, which can help you document editorial control.

6.4 How to label. The disclosure must be clear and distinguishable, given at the latest when people first see the content, and accessible. Practical options are a visible caption such as "AI-generated" (Romanian: "Generat cu AI"), the common EU icon described in the Code of Practice, and the AI labelling tools of the Third-Party Platform.

6.5 A label does not make content lawful. Disclosure does not cure the lack of a person's consent, defamation, intellectual property infringement or misleading advertising. AI content in ads must not mislead consumers.

6.6 Political content. The Platform does not support political advertising in the EU (Regulation (EU) 2024/900). See acceptable-use-policy.md, section 3.5.

7. Human oversight and the approval model

7.1 The AI never acts alone. The AI never executes changes on ad accounts and never publishes content by itself.

7.2 How a change happens. Proposal by the AI or by a user, then review, then approval by an authorised person (with configurable multi-step workflows, for example agency first and then client), then execution through the Third-Party Platform's API, then an entry in the audit log.

7.3 What a proposal shows. Each proposal shows what would change, the current and proposed values, the reasons, and the data it is based on. [TBC: expected impact and uncertainty indicators.]

7.4 Auto-approval. In the future, Customers may configure an auto-approval policy for certain changes, always within the Platform's hard caps (for example a maximum budget change). Auto-approval policies are disabled during the Access Program (access-program-terms.md, Section 8.5), without exceptions. Where a Customer enables it, changes executed under the policy are the Customer's decisions. Trend Radar concepts and content on gated sensitive topics are never auto-approved or auto-published.

7.5 Hard caps. Platform-wide limits apply even when a change has been approved. [TBC: publish cap values or describe them in the documentation.]

7.6 Stop at any time. Authorised users can pause any automation at any time and can activate crisis mode (see brand-safety-and-sensitive-content-policy.md).

7.7 Audit log. Every proposal, approval, rejection and executed action is recorded in an audit log (with pseudonymised user identifiers), so you can see who approved what and when.

7.8 Approvals are decisions. Review proposals carefully before approving. An approval is the Customer's decision and the Customer remains responsible for it.

8. Limitations and known risks

AI is useful but imperfect. These are the main risks we know about and what we do about them.

8.1 Hallucinations. The AI may produce statements, figures or sources that sound plausible but are wrong. We ground answers in your connected data and show the underlying figures where possible, but errors are not eliminated. Check before you approve or publish.

8.2 Outdated or incomplete data. Third-Party Platforms report with delays; attribution windows, conversion lag, data sampling and API outages affect figures. The general knowledge of AI models has a cut-off date. Trends change quickly and may be over before you publish.

8.3 Estimates are estimates. ROI, profit, attribution and incrementality depend on the inputs you provide (for example COGS, refund and cash-on-delivery refusal rates) and on statistical assumptions. Treat them as decision support, not certainty.

8.4 Bias. AI models can reproduce stereotypes, for example in copy, images or audience suggestions. Review Outputs for unfair or discriminatory content and targeting (acceptable-use-policy.md, section 3.8).

8.5 Brand-safety errors. Classifiers can miss sensitive content (for example new slang, memes, coded language or nuances between languages) or block harmless content. When unsure, the Platform escalates rather than allows, but you remain the final check.

8.6 Prompt injection. Web pages, comments, product feeds or documents may contain hidden instructions aimed at the AI. The Platform treats such content as data and requires human approval for every action, but residual risk remains. Report suspicious behaviour to ops@1ads.net.

8.7 Similarity to existing works. Generated images or text may resemble existing works, logos, products or people. Check before use (acceptable-use-policy.md, section 3.11).

8.8 Language and quality. Quality can vary between languages and tasks. Outputs are not legal, tax, financial or medical advice.

8.9 Variability. The same request can produce different results, and results may change when providers update models.

9. How the trend brand-safety gate works

9.1 Every trend and every concept derived from it is classified as Blocked, Gated or Standard: (a) Blocked: for example the commercial exploitation of tragedies, disasters, wars and deaths; hate and harassment; sexual content; extremism; dangerous misinformation; elections and political actors. No concepts are generated. (b) Gated: for example social movements such as #MeToo, health topics, religion, commemorations and real people in the news. Concepts are allowed only with senior approval and, by default, in a "purpose-only" posture with no product, offer or call to buy. (c) Standard: normal approval workflow plus your brand rules.

9.2 Customers can make the rules stricter for their brands but cannot relax the platform floor. Concepts are re-checked before publication, because the context can change between approval and publication. Trend Radar concepts are never published automatically.

9.3 The full rules, including crisis mode, are in brand-safety-and-sensitive-content-policy.md.

10. No automated decisions with legal effects

10.1 The Platform does not make decisions about individuals that produce legal effects or similarly significantly affect them within the meaning of Article 22 GDPR. Its proposals concern campaigns and content, and decisions are taken by people.

10.2 Customers must not use the Platform for such decisions (acceptable-use-policy.md, section 7.5). Ad delivery decisions made by the algorithms of Third-Party Platforms are the responsibility of those platforms.

11. Reporting problems with AI Features

11.1 If an Output is wrong, harmful, biased, infringing or not labelled correctly, tell us:

• quality problems and wrong Outputs: support@1ads.net [TBC: in-app feedback button];
• harmful or illegal Outputs and labelling problems: support@1ads.net;
• prompt injection, data leaks or other security issues: ops@1ads.net;
• personal data questions: privacy@1ads.net.

11.2 Please include the date and time, the feature, the brand or workspace, and a screenshot or the identifier of the proposal or content. We review reports, may adjust filters, prompts or models, and tell you what we did where appropriate.

12. Changes to this notice

We update this notice when features, models, providers or the law change. The version number and effective date at the top show the current version.