Berlin has given AI search product teams a useful, but incomplete, map of the legal terrain. In a June 16 ruling reported by The Decoder, a Berlin court rejected the idea that Google’s AI Overviews should be treated as original content. Instead, it described them as a new search format that aggregates third-party material, with users able to recognize that the system is drawing from other sites and Google exercising no “decisive influence” over the underlying content.
That matters because it shifts the center of gravity in the liability analysis. If an AI Overview is legally understood as a display layer rather than an authored statement, the product is less like a publisher making a fresh factual claim and more like a retrieval interface that reorganizes existing sources. That does not eliminate risk. It changes where the risk sits: in sourcing, attribution, ranking, and the mechanics of how the summary is rendered and linked.
The complication is that Germany does not yet have a single, settled answer. A separate Munich court took a different view in an AI-summary case, treating generated summaries as independent content and holding Google responsible for errors because it controlled the algorithms. That split does not just create a doctrinal curiosity. It creates a product problem. Teams rolling out AI-assisted search in Germany now have to design for two competing legal theories at once: one that treats the feature as a format, and another that treats it as a statement.
What changed in Berlin
The Berlin decision arose from a trademark dispute involving a perfume company. According to the reporting, the AI Overview mentioned protected brand names alongside cheaper knockoffs and linked to their websites. The court nonetheless found no trademark infringement, reasoning that users could tell the search engine was pulling together information from outside sources and that Google did not have decisive control over the content itself.
That framing is the important part for product teams. A “display format” theory implies that the AI layer is a presentation mechanism layered on top of indexed material, not a wholly new authored work. In technical terms, that means the legal analysis will likely focus less on the generative model as an abstract black box and more on the upstream and downstream systems around it:
- which sources were retrieved,
- how source snippets were selected,
- whether attribution is visible and understandable,
- how aggressively the system compresses or paraphrases,
- and whether the interface signals uncertainty or provenance.
If the court’s logic holds, the feature’s defensibility depends partly on whether a reasonable user can see it as synthesized search output rather than an independent editorial claim. That is a product-design standard as much as a legal one.
The subtle but crucial implication is that the model’s internal generation step may matter less than the system architecture around it. A ranking model that selects sources, a summarizer that rewrites text, a UI that shows citations, and a fallback layer that suppresses low-confidence responses all become part of the liability story. In other words, the “content” is not just the tokens the model emits; it is the end-to-end assembly of retrieval, synthesis, attribution, and presentation.
Liability now tracks control, not just output
The Berlin ruling’s emphasis on lack of “decisive influence” is the kind of language product and policy teams should read carefully. It suggests that liability could turn on how much control the platform has over the specific claim being shown. If the system is clearly surfacing third-party material and labeling it as such, the feature looks more like a search result page. If it is generating a statement that appears to stand on its own, the risk profile starts to resemble publication.
That distinction is not academic. For an AI-assisted search product, the following choices can materially affect legal exposure:
- Attribution density: Are source links visible above the fold, inline, or only in a separate click-through?
- Claim granularity: Does the system quote, paraphrase, or synthesize across sources into a new assertion?
- Source diversity: Does the response rely on one source, or does it show clearly mixed provenance?
- Confidence gating: Does the system suppress outputs when sources conflict or confidence is low?
- Fallback behavior: Does the UI degrade to classic links when synthesis is uncertain or when the query touches sensitive claims?
Berlin’s format-based framing encourages a design posture that keeps the output anchored to source material. Munich’s independent-content stance points in the opposite direction: if the system appears to make its own factual claims, the operator may be treated as responsible for the statement even when the model is only probabilistic. The tension between those views is why this ruling should be read as a liability map, not a liability exemption.
Why product teams should care about the Berlin-Munich split
For search teams, the central question is not whether a model is “intelligent” or “original.” It is whether the legal system will see the feature as a presentation of indexed information or as a fresh assertion attributable to the platform.
That has several concrete product implications.
First, source presentation becomes a core control surface, not a decorative UX choice. If the court’s reasoning relies on users being able to tell that the system is drawing from other websites, then provenance needs to be legible, persistent, and hard to miss. Hidden citations or vague “learn more” affordances are not enough if the feature wants to preserve the search-format argument.
Second, licensing strategy becomes more important. If the output is materially derived from third-party content, rights holders may push for commercial terms even when the platform argues it is only displaying a format. That means legal and product teams need to align on when the system can summarize, when it must quote minimally, and when it should only deep-link.
Third, quality assurance has to expand beyond factual accuracy. A classic search product can often tolerate some ranking noise. An AI summary cannot. If the summary compresses multiple sources into a single assertion, the error surface changes: a small source mismatch can become a user-visible falsehood. Under a Munich-style view, that falsehood could be treated as the platform’s own statement.
The engineering consequences are practical, not abstract
Teams building AI-assisted search should treat Berlin’s ruling as a prompt to harden the provenance pipeline.
At a minimum, that means building systems that can answer four questions for every generated response:
- What sources were retrieved?
- Which passages were actually used to construct the summary?
- What confidence or relevance threshold allowed the answer to ship?
- What fallback path was available if the evidence was too thin or conflicting?
The most defensible architectures will likely have explicit provenance metadata from retrieval through rendering. That includes source IDs, timestamps, ranking scores, and traceable mappings from retrieved passages to displayed claims. If a legal dispute arises, teams need to show not just that the model was “grounded,” but how grounding was enforced.
Suppression logic also matters. A system that always generates an answer is harder to defend than one that can refuse to summarize, especially on branded, medical, financial, or highly contested factual queries. In practical terms, a robust AI search stack should be able to fall back to classic search results when:
- source coverage is thin,
- retrieved pages disagree materially,
- the query concerns protected marks or regulated claims,
- or the model cannot reliably attribute a specific assertion.
That fallback is not merely a safety feature. It is evidence of editorial restraint, which may become relevant if a court examines the platform’s degree of control over output.
Licensing and policy need to move together
The Berlin ruling does not settle whether platforms owe compensation for use of third-party material. But it does sharpen the business question of what kind of content the platform believes it is shipping.
If AI Overviews are a display format built from third-party sources, then licensing conversations shift from abstract model training debates to output-level rights. Product managers should expect pressure on at least three fronts:
- direct licensing for high-value content classes,
- whitelisting or exclusion workflows for rights-sensitive publishers,
- and clearer policies on what can be excerpted, paraphrased, or synthesized.
This is especially relevant for teams that want AI search to work across news, commerce, travel, health, and local services. The more the feature behaves like a substitute answer, the more likely it is to collide with content owners who do not want their pages turned into machine-generated summaries.
Policy teams should also be careful not to overread Berlin as a green light. The ruling narrows one theory of originality, but it does not erase consumer-protection, trademark, copyright, or unfair-competition concerns that may arise when summaries are wrong or misleading. And Munich’s independent-content approach means that even technically well-sourced summaries may still be treated as statements for which the platform bears responsibility.
What teams should do now
For engineering and product leaders shipping AI search features, the right response is not to wait for doctrinal clarity. It is to make the system more explainable, more suppressive when uncertain, and more traceable end to end.
A practical implementation plan looks like this:
| Workstream | Owner | Timeline | Success criteria | |---|---|---:|---| | Provenance logging for all AI summaries | Search engineering | 2-4 weeks | Every rendered summary stores source IDs, passage spans, and retrieval scores | | Visible citation redesign | Product + UX | 2-6 weeks | Citations are inline and readable without extra navigation on desktop and mobile | | Confidence gating and fallback rules | ML platform + search ranking | 3-6 weeks | Low-confidence or conflicting-source queries automatically fall back to classic results | | Rights-sensitive query taxonomy | Policy + legal + PM | 2-4 weeks | Branded, regulated, and high-risk queries are tagged and routed to stricter handling | | Summary QA and red-team review | ML quality + trust & safety | Ongoing, weekly | Material error rate on sampled summaries drops by at least 50% over baseline | | Licensing review by content class | Legal + business development | 4-8 weeks | Priority content categories have documented licensing or exclusion decisions | | Jurisdiction-specific rollout gates | PM + legal + platform ops | 2-6 weeks | Germany-specific policies are distinct from other markets and version-controlled |
A useful target for teams is not perfection but measurable risk reduction. For example: reduce uncited summary instances to near zero; ensure every high-risk query class has a documented fallback; and track the percentage of AI answers that can be fully traced back to source passages. If the team cannot explain how a summary was assembled, it is not ready for broad rollout.
The strategic read
Berlin did not declare generative search legally safe. It drew a line around one specific feature framing: AI Overviews, at least on the facts before the court, can be understood as a search display format that aggregates third-party content rather than as original publication.
That may sound like a win for platforms, but it is really a reminder that the legal status of AI search depends on product architecture. If the system looks like search, cites like search, and behaves like search, Berlin gives it a better chance of being treated like search. If it starts to look like the platform is making independent factual claims, Munich shows how quickly the analysis can flip.
For AI teams, the response should be straightforward: build provenance first, attribution always, fallback by default, and licensing where the content value justifies it. The legal ambiguity is not going away soon. The product should be designed as if both sides of that ambiguity will matter.
