An EU-resident publisher using LLMs to generate retail finance content sits at the intersection of four regulatory regimes: BaFin / WpHG for general finance content supervision1, MiCA for crypto-asset content2, the EU AI Act for the LLM-tooling itself3, and ESMA's MiFID II suitability guidelines for anything that might be construed as advice4. The honest posture is documented restraint: no personalised advice, AI involvement disclosed, methodology declared, holdings disclosed. The compliance burden is modest if the operation is disciplined and material if it is not.

TL;DR

  • Four overlapping EU regimes apply: BaFin/WpHG, MiCA (for crypto content), EU AI Act, ESMA MiFID II.
  • The defensible posture: explicit no-advice framing, AI-involvement disclosure, methodology published, holdings disclosed.
  • BaFin's 2024 finfluencer factsheet is the most-current operational guidance for retail content1.
  • MiCA applies if the content covers crypto-assets within the EU regime2.
  • The AI Act's deployer obligations apply to LLM-using publishers, disclosure and human-oversight documentation.

The four regimes

BaFin / WpHG (general finance content)

The Wertpapierhandelsgesetz (WpHG) supervises content that discusses regulated investment products in Germany. BaFin's enforcement is registration-based for advisory services and substantiation-based for content. The 2024 BaFin Fachartikel on finfluencers documents the active enforcement perimeter for social-media finance content1, updated by BaFin's 2026 finfluencer factsheet5. The relevant operating posture:

  • No personalised advice. Specific buy/sell recommendations to identified persons cross the line.
  • Disclosure of paid endorsements, holdings, and affiliate relationships.
  • Substantiation of performance claims with documented evidence.
  • 5-7 year record retention on published claims and the data behind them.

MiCA (crypto-asset content)

The Markets in Crypto-Assets Regulation (MiCA) supervises content about crypto-assets within the EU regime2. For a finance publisher whose content covers crypto-assets:

  • Content describing specific crypto-assets falls under MAR / MiCA disclosure if recommending an investment strategy6.
  • Promotional content for unregistered crypto offerings is restricted.
  • Stablecoin and exchange-token coverage has tighter rules than other crypto content.

For a publisher covering crypto narrowly (e.g., one article on Bitcoin per quarter), MiCA is largely a documentation overlay. For a publisher whose primary focus is crypto, MiCA is the binding regime.

EU AI Act (LLM-tooling)

The EU AI Act applies to the LLM tooling itself, both at the provider level (Anthropic, OpenAI, Google) and at the deployer level (the publisher using the LLM)3. Deployer obligations under Article 50 include:

  • Transparency disclosure when content is AI-generated or AI-assisted.
  • Documentation of the AI use case and its risk class.
  • Human-oversight documentation showing how the publisher reviews AI-generated content before publication.
  • Annual risk-assessment update.

The cost of compliance for a small publisher is modest, €500-2,000/year in documentation time. The cost of non-compliance is potentially material, particularly if the content is later used to substantiate a complaint or enforcement action.

ESMA MiFID II suitability

ESMA's suitability guidelines apply to any content that could be construed as an investment suitability assessment4. The defensible boundary:

  • Content that describes investment strategies for a general retail audience without targeting specific risk profiles: lighter scrutiny.
  • Content that personalises recommendations to identified profiles ("if you are a moderate-risk investor, consider X"): suitability assessment, requires registration.

For a content publisher operating in the general-information lane, the suitability rules are about phrasing — avoid personalisation, frame content as education, do not assess individual circumstances.

The defensible posture

Five operating rules that satisfy all four regimes:

1. Explicit no-advice framing

Every article carries a disclosure: "This is educational content. It is not personalised investment advice. Consult a registered advisor before making investment decisions." The phrasing is conventional; the legal effect is to position the content outside the advisory regime.

2. AI involvement disclosure

If the article was generated or substantively assisted by an LLM, disclose it. The AI Act expects this; readers expect it; the disclosure does not materially affect reader trust if the methodology is documented.

The pattern: "Drafted with Claude Opus 4.7 assistance and reviewed by the editor before publication. The methodology is documented at /methodology/."

3. Methodology published

For any quantitative claim (Sharpe ratios, backtests, performance comparisons), the methodology section documents:

  • Data sources.
  • Time window.
  • Computation method.
  • Trial count (if multiple candidates were tested).
  • Assumptions (cost model, slippage, fees).

The Research Diary Schema walks the structured format. Publishing the methodology shifts the compliance posture from "claim X" to "claim X, supported by methodology Y," which is materially easier to defend.

4. Holdings and conflicts disclosed

Any position the author holds in covered securities, any paid relationship with covered vendors, any affiliate compensation: disclosed in the article. The disclosure has to be specific enough to be useful — "the author holds positions in technology stocks" is too vague; "the author holds long positions in MSFT and NVDA at the time of publication" is the right specificity.

5. No personalised recommendations

The hard rule. The moment content moves from general-information to personalised-advice, the registration regime applies and the compliance burden multiplies. Stay on the general-information side of the line.

What the LLM specifically changes

Pre-LLM finance content had the same regulatory framework, just at lower volume. The LLM changes:

  • Volume. A solo publisher can produce 5-10x more content per week. The compliance overhead per article must be low enough to absorb this volume.
  • Substantiation cost. Each article needs the same level of substantiation; the LLM does not reduce the substantiation work, it just generates more content that needs substantiating.
  • AI Act disclosure overhead. New regime layer specifically for LLM-using publishers.
  • Hallucination risk. LLMs occasionally invent facts. For finance content, hallucinated facts are a substantiation failure that creates enforcement risk. Editorial review must catch these — see LLM Finance Error Taxonomy: The Bond-Yield Trap for the failure modes.

A defensible workflow

1. PRE-REGISTER topic, scope, target audience.
2. DRAFT with LLM assistance; document the prompts used.
3. REVIEW manually for:
   - Hallucinated facts.
   - Personalised recommendations creeping in.
   - Required disclosures present.
   - Substantiation chain intact for every quantitative claim.
4. PUBLISH with methodology and disclosures inline.
5. RETAIN records: prompts, drafts, sources, methodology, edits — 5-7 years.

The workflow scales: each article carries ~30-60 minutes of editorial review on top of the LLM-assisted draft. For a publisher producing 2-3 articles per week, that is 1-3 hours of weekly editorial time — manageable as a solo operation.

Cost summary

For a solo EU publisher producing 100 finance articles per year:

Cost line Annual (€)
LLM API spend (Claude Sonnet 4.6, ~$10/article) 900
Editorial review time (own labour, opportunity cost) 5,000-15,000
Insurance 2,000
Legal review (4 hours × €550) 2,200
AI Act documentation overhead 1,200
BaFin / MAR documentation overhead 1,500
Total annual cost €12,800-22,800

The number is small if the operation is disciplined. It compounds rapidly if the operation cuts corners — a single substantiation failure that triggers BaFin inquiry can cost €10,000+ in defensive legal time on top.

What this excludes

  • Personal advisory engagements. Different regime, different cost stack.
  • Active money management for clients. Requires BaFin registration; outside this article's scope.
  • Trade recommendations to identified profiles. Crosses into MiFID II suitability framework.
  • Crypto exchange or token issuance. MiCA framework, separate analysis.

For a publisher staying in the general-information lane on traditional finance content, the cost above is the relevant envelope.

Failure modes

  • Treating no-advice framing as a magic phrase. The framing is necessary but not sufficient. If the content personalises recommendations regardless of disclaimer, the rule applies.
  • Skipping the substantiation chain. Every claim needs sources; "Claude said so" is not a source.
  • Mixing the publisher's own positions into recommendations without disclosure. This is the most-enforced violation in published finance content.
  • Ignoring the AI Act overlay. The Act is in force; disclosure is required.

FAQ

Do I need to register with BaFin to publish finance content?

For general-information content without personalised advice, no. For content that crosses into advisory or active money management, yes. The line is operational, not aesthetic — the BaFin + EU Guide walks it in detail.

Does the AI Act apply if I only use LLMs for editing, not for generation?

The deployer obligations focus on the publisher's use of AI in the editorial process. Editing-only use is lighter than generation use, but the disclosure threshold is low — if AI substantively affected the output, disclose. When in doubt, disclose.

How does this compare to the US FTC framework?

The US FTC framework is reactive (enforcement after a complaint) and lighter on registration. The EU framework is more active and carries more documentation overhead but does not require registration for general-information content. The cost gap is real but bounded — see FTC vs BaFin: Publishing Rule Cost Compared.

Connects to

References

Footnotes

  1. BaFin (2024). "Generation Y und Z setzen auf Finfluencer." Fachartikel. bafin.de ↩ ↩2 ↩3

  2. EU (2023). "Markets in Crypto-Assets Regulation (MiCA), Regulation (EU) 2023/1114." eur-lex.europa.eu ↩ ↩2 ↩3

  3. EU (2024). "Artificial Intelligence Act, Regulation (EU) 2024/1689." eur-lex.europa.eu ↩ ↩2

  4. ESMA (2023). "Guidelines on certain aspects of the MiFID II suitability requirements." esma.europa.eu ↩ ↩2

  5. BaFin (2026). "Mehr Transparenz für Finfluencer — Factsheet." bafin.de ↩

  6. EU (2014). "Market Abuse Regulation (MAR), Regulation (EU) No 596/2014." eur-lex.europa.eu ↩