Alevli framework
The Alevli filter
An AI marketing framework that decides whether a given use of AI or persuasion is acceptable by asking four questions in order, can the customer see it and refuse it, would they expect this use of what you know about them, does a named person stand behind it with sources, and can a human stop it and measure it, and grading the use by its weakest answer, so that the verdict follows the lines drawn by EU law and by the customer’s own expectations rather than by taste.
Most AI marketing frameworks are lists of things to switch on. This one is an AI filter: four questions that decide whether a given use of AI, or of any persuasion technique, is acceptable, and grade it by its weakest answer. The Alevli filter exists because marketing now answers to legal lines it did not have until recently: the AI Act's prohibitions since February 2025 and its transparency rules since August 2026, on top of the Unfair Commercial Practices Directive and, for online platforms, the Digital Services Act, with the Digital Fairness Act the Commission plans to table in the fourth quarter of 2026. What it does not have is a practitioner test that tells a marketer on a Thursday afternoon which side of those lines a feature sits on.
The filter borrows its spine from the European Commission's own guidance, which draws the line between lawful persuasion and manipulation at awareness and autonomy: manipulation is influence people are not aware of, or cannot resist even when they are. The filter extends the same logic to the three other places AI enters marketing: what a business infers about a customer, what it publishes under its name, and what it lets a model decide without a human. Each of the four questions protects one value the customer would name if asked: autonomy, expectation, honesty, accountability. The short version to repeat is the first question, because the other three are variations of it: could the customer see this, and would they still say yes?
The Alevli filter
Persuasion
Personalisation
Provenance
Control
The verdict
Use it when
- Deciding whether to switch on an AI feature in an ad platform, a CRM, a chatbot or a content pipeline
- Reviewing a landing page, checkout, pricing page or campaign before it goes live, whether a person or a model built it
- A vendor pitches a personalisation, neuromarketing or "behavioural AI" tool and cannot say which side of the law it works on
Do not use it when
- A specific enforcement question is on the table; the filter sorts uses, it does not replace a lawyer
- Audiences of minors or other groups the law names as vulnerable, where every gate is stricter than the questions alone suggest
The steps
- Name the use precisely, and name who decides. "Personalisation" is not a use; "a model that raises a financing offer to visitors it scores as financially stressed" is. Say what it does, at what moment, to whom, and whether a person or a model makes the call. Half of all disagreements about acceptability are disagreements about which use is being discussed, and the same technique can sit in different bands depending on who executes it.
- Gate one, persuasion: can the customer see it and refuse it? Three sub-questions, in order. Can they see it working: a price anchored next to a higher one passes, an imperceptible cue fails. Can they refuse it without cost, delay or shame: a highlighted default with a one-click alternative passes, a countdown that resets, a six-screen cancellation and a decline button that reads "No, I don't want to save money" fail. Do they end up no worse off by their own standard: a guarantee passes, a price tuned to inferred distress fails. Failing the first sub-question with AI involved puts the use close to the AI Act's ban, which applies once the distortion is material and the harm significant.
- Gate two, personalisation: would the customer expect this use of what you know about them? Place the use on the personalisation ladder below. Contextual and declared data pass. Behavioural data passes when the basis is disclosed and refusable. Inferred traits pass only when the customer would recognise the basis. Sensitive inference, health, finances, mood, vulnerability, fails: health data usually needs explicit consent, and a model that exploits financial hardship to push a harmful decision can fall under the AI Act's outright ban.
- Gate three, provenance: does a named person stand behind it, with sources? The test for AI-generated content is not whether a machine wrote it but whether a named human would sign it, at this volume, with these numbers sourced and these people and events real. Unattributed claims fail. Unsourced numbers fail. Synthetic people or events presented as real fail as fake endorsements under the UCPD, and deepfakes of real people must be labelled under the AI Act. Content nobody would put their name to is the definition of slop: LinkedIn began demoting it in 2026, and Google's August 2026 spam update was read by SEOs as hitting sites built on mass-produced pages.
- Gate four, control: can a human stop it, and is its effect measured? A budget a model can raise on its own, a campaign with no kill switch, an automation with no baseline and no holdout fail. The question is not whether AI executes but whether a person can halt it in an hour and prove in a month what it did. Then grade the use by its weakest gate, write the verdict down, and re-run whenever the decision moves from a person to a model.
The personalisation ladder
Gate two needs a ruler, because "how much personalisation" is the question marketers actually face. The ladder runs from what the customer can plainly see to what only a model would know.
| Rung | What the personalisation is based on | Verdict | Why |
|---|---|---|---|
| 1. Contextual | The page, the query, the time of day | Pass | The customer chose the context |
| 2. Declared | A form, a stated preference, a chosen service | Pass | The customer supplied it for this purpose |
| 3. Behavioural | Pages visited, cart abandoned, the whitening page viewed twice | Caution | Expected only if the basis is disclosed and an opt-out is one click away |
| 4. Inferred | Life stage, income band, intent scored by a model | Caution to fail | Passes only where the customer would recognise the basis if shown it |
| 5. Sensitive inference | Health, finances, mood, vulnerability | Fail | Health data usually needs explicit consent under GDPR Article 9; a model exploiting financial hardship can fall under AI Act Article 5(1)(b); unfair personalisation is a named target of the planned Digital Fairness Act |
A use that sits on rung three or four is not wrong. It is a use that must be able to answer "how did you know that?" without embarrassment. Rung five cannot answer the question at all, which is why the law answers it instead.
Worked example
A local dental practice's marketing stack, reviewed through all four gates, illustrative. The verdict is the weakest gate.
| Use, as deployed | Persuasion | Personalisation | Provenance | Control | Verdict |
|---|---|---|---|---|---|
| Whitening price shown beside the implant price | Pass | Pass | Pass | Pass | Acceptable |
| "Next appointment within 48 hours", true, checked weekly | Pass | Pass | Pass | Pass | Acceptable |
| Smart bidding with a budget cap and a monthly holdout by district | Pass | Pass | Pass | Pass | Acceptable |
| Retargeting people who viewed the whitening page with the whitening offer, opt-out in the ad | Pass | Caution (rung 3) | Pass | Pass | Acceptable with disclosure |
| AI-drafted blog posts edited, sourced and signed by the practice owner | Pass | Pass | Pass | Pass | Acceptable |
| Chatbot that answers as "Emma" without saying it is software | Pass | Pass | Fail | Pass | Fix before launch: AI Act Art. 50(1), UCPD |
| Countdown on a whitening offer that resets on reload | Fail | Pass | Fail | Pass | Unfair: UCPD Annex I point 7, false urgency |
| Twelve AI-generated "patient stories" with stock faces, unlabelled | Pass | Pass | Fail | Pass | Unfair: UCPD Annex I point 23c, fake endorsements |
| Goal-only ads where the platform sets creative, placement and spend with no cap | Pass | Caution | Caution | Fail | Not yet: no human stop, no measurement |
| Model raising financing offers to visitors it scores as financially stressed | Fail | Fail (rung 5) | Fail | Fail | Likely prohibited: AI Act Art. 5(1)(b) |
Five of the ten fail at least one gate, and all five would have passed a "does it convert" review. Three would have broken the law on the day they went live, and one risks the AI Act's outright ban. The filter's value is not that it catches them; a lawyer would, later. Its value is that the marketer catches them on Thursday and can say why in one sentence per gate.
Why it works
The four questions of this AI marketing framework track the law's own tests, in the order a marketer meets them. Gate one is the Commission's 2025 guidelines on prohibited AI practices in plain language: personalised advertising is not inherently manipulative when it uses no subliminal, manipulative or deceptive technique, and manipulation is influence people are not aware of, or cannot resist even when they are. The AI Act's prohibition is cumulative, requiring a subliminal or manipulative technique, a material distortion of behaviour and significant harm, which the guidelines say includes psychological and financial harm. Because the Commission reads "material distortion" in the light of the Unfair Commercial Practices Directive, the same sub-questions sort human-executed techniques onto the UCPD's lines and, for online platforms, interface design onto Article 25 of the Digital Services Act.
Gate two is the GDPR's purpose-limitation principle and Nissenbaum's contextual-integrity test rendered as one question: information flows are appropriate when they match the norms of the context in which the information was shared. Three GDPR provisions decide the rungs in practice. Behavioural personalisation needs a lawful basis, and where it runs on cookies or similar tracking it needs consent under the ePrivacy rules, in the Netherlands article 11.7a of the Telecommunicatiewet. Rung five touches the special categories of Article 9, health among them, which are closed to marketing without explicit consent. And a price or offer decided solely by a model with a significant effect on the person falls under Article 22's rules on automated decision-making. The Digital Fairness Act's announced scope, unfair personalisation that takes advantage of consumers' vulnerabilities, is the ladder's fifth rung written into a legislative agenda.
Gate three is Article 50 of the AI Act (a chatbot must say it is software, a deepfake must be labelled) plus the UCPD's prohibitions on misleading practices and fake reviews, with a 2026 twist: LinkedIn reported a 46% rise in detected inauthentic activity in the first half of the year and gave users a button to flag AI slop, and Google's August spam update was read by SEOs as targeting mass-produced content, so the provenance gate is now enforced by distribution as well as by law.
Gate four borrows the AI Act's human-oversight principle, which the Act mandates only for high-risk systems but which any marketer can adopt voluntarily, and pairs it with the holdout test: a use that cannot be stopped in an hour or measured in a month is a promise no one is in a position to keep.
Underneath the legal lines sits an older principle. Sunstein's test for a legitimate nudge is publicity: you should be willing and able to defend it openly to the people it is used on. Every gate is that test applied to a different part of the machine.
The four values and the UNESCO Recommendation
The four gates were not chosen by taste. UNESCO's 2021 Recommendation on the Ethics of Artificial Intelligence, adopted by all 193 member states, lists ten principles; the four gates map onto the five a marketer can act on directly, and leave the rest to the people who build and run the models.
| Gate | UNESCO principle | What the marketer controls |
|---|---|---|
| Persuasion | Human oversight and determination | Whether the customer decides, or is steered past deciding |
| Personalisation | Right to privacy; fairness and non-discrimination | What is inferred, and whether vulnerability becomes a target |
| Provenance | Transparency and explainability | Whether a person and a source stand behind the claim |
| Control | Responsibility and accountability | Whether someone can stop it and prove what it did |
The Recommendation is not binding, and the filter does not pretend it is. The law draws the lines the bands sit on; UNESCO supplies the reasons the lines are where they are, and gives the filter an anchor outside the EU for readers to whom the AI Act does not apply.
The pair: the filter and the gap
The Alevli gap says where to set a promise: at the highest level the business can still beat on a bad week. The Alevli filter says which ways of getting to yes would survive the customer seeing them used. The two share one law, satisfaction is delivery minus expectation, and divide one job: a marketing framework for the promise and an AI filter for the method. Every gate of the filter is a gap failure by another route:
| Gate | The promise it protects | The gap failure it prevents |
|---|---|---|
| Persuasion | The customer chose freely | Expectation debt manufactured on purpose |
| Personalisation | The business uses what it knows as the customer expected | A broken implicit promise about data |
| Provenance | Every claim has a person and a source behind it | Insight promised in the hook, filler delivered in the body |
| Control | Someone can keep the promise the automation made | A promise nobody stands behind |
The bridge sentence belongs on both pages: a promise the customer would not have accepted with full sight is debt, whoever or whatever made it. Overselling fails both instruments at once. It sets the promise above the bad week, and it usually needs an unfair technique to get it signed. A business that passes the gap and the filter is one that could show every customer both the promise and the method and lose nothing by it, which is the plain-language definition of the beatable band.
The legal lines are summarised from the sources below and from public legal commentary; the filter is a sorting method, not legal advice. The four gates, the weakest-gate verdict, the personalisation ladder and the mapping of each gate to a gap failure are original to this model.
Sources
- European Commission (2025). Guidelines on prohibited artificial intelligence practices established by Regulation (EU) 2024/1689, C(2025) 5052 final, approved 4 February 2025 and adopted 29 July 2025.
- Regulation (EU) 2024/1689 (AI Act), Articles 5(1)(a) and (b) (applicable from 2 February 2025), Article 14 (human oversight, high-risk systems) and Article 50 (transparency, applicable from 2 August 2026; the Art. 50(2) marking duty from 2 December 2026 for systems already on the market).
- Regulation (EU) 2016/679 (GDPR), Article 5(1)(b) purpose limitation, Article 9 special categories of data, Article 22 automated individual decision-making; Directive 2002/58/EC (ePrivacy) and, in the Netherlands, Telecommunicatiewet article 11.7a on consent for tracking.
- Directive 2005/29/EC (Unfair Commercial Practices Directive), Articles 5 to 9 and Annex I, points 7 and 23c.
- Regulation (EU) 2022/2065 (Digital Services Act), Article 25.
- European Commission (2024). Fitness Check of EU consumer law on digital fairness, SWD(2024) 230 final; Digital Fairness Act planned for Q4 2026.
- European Commission (2022). Behavioural study on unfair commercial practices in the digital environment: dark patterns and manipulative personalisation.
- Nissenbaum, H. (2004). Privacy as contextual integrity. Washington Law Review, 79(1), 119–157.
- Sunstein, C. R. (2016). The Ethics of Influence: Government in the Age of Behavioral Science. Cambridge University Press.
- UNESCO (2021). Recommendation on the Ethics of Artificial Intelligence, adopted by the General Conference at its 41st session.
- LinkedIn (2026). Transparency disclosures under the Digital Services Act, first half of 2026, on detected inauthentic activity.
- Srinivasan, H. (30 July 2026). LinkedIn post announcing the "Seems like AI slop" flag.
- Google Search Status Dashboard (August 2026). August 2026 spam update.
Who this is for
The same method, read three ways.
01
Running it
Name the use and who decides, then ask the four questions in order and write down the weakest answer. That answer is the verdict, not the average of the four.
02
Teaching it
The gates are the law's own tests in the order a marketer meets them: the Commission's perceive-and-resist line, contextual integrity, the AI Act's transparency rules and human oversight. What is new is the weakest-gate verdict and the personalisation ladder.
03
Buying a tool
Before you sign, ask the vendor which rung of the ladder the tool works on and how you stop it in an hour. A vendor who cannot answer both is selling you the risk.