VIDEO | AUDIO | RECAP EN / ES / FR | ARCHIVE | PERMALINK
Speaker: Patrick Jones - Vice President, Global Stakeholder Engagement, ICANN
Moderator: Gabriella Schittek - Global Stakeholder Engagement Director, Nordic and Central Europe, ICANN
AI and Internet identifiers
Patrick Jones examined how popular artificial intelligence models interact with the Internet’s unique identifiers, including domain names, IP addresses, autonomous system numbers, and protocol parameters. Drawing on tests conducted approximately six weeks apart, he compared the responses of Anthropic’s Claude Sonnet 4.6 and Google Gemini.
He emphasized that AI is only the latest technological development to affect the environment in which the Domain Name System operates. Previous changes included the rise of the mobile Internet, deployment of DNS Security Extensions, growth of the Internet of Things, development of encrypted DNS protocols, and renewed interest in alternate namespaces and blockchain-based identifiers.
Jones referred participants to a March 2026 blog post by ICANN’s Office of the Chief Technology Officer. Its central point was that AI does not change ICANN’s mission or the fundamental architecture of the DNS, although it creates new operational and policy questions for the wider Internet ecosystem.
How the models work
Large language models generate text and other content based on information absorbed during training. They may connect to external files, applications, and services through tools such as the Model Context Protocol.
Jones stressed that an LLM is not ordinarily a networked application in the same sense as a browser or mail server. The model itself does not necessarily conduct DNS lookups or initiate IP connections. Networking instead takes place in the surrounding infrastructure or through tools that provide web search, web retrieval, or API access.
A model may understand the hierarchy of the DNS, DNSSEC, IP address allocation, regional Internet registries, routing registries, and Internet peering. However, that knowledge can become outdated. Claude Sonnet 4.6, for example, had a stated knowledge cutoff of August 2025. Without access to current external information, it could be unaware of subsequently introduced domains, delegations, or protocols.
Agentic AI introduces additional considerations because agents may autonomously contact domains or APIs. Jones said this raises questions about DNS abuse, the use of identifiers, and the accountability of automated systems.
Comparing Claude and Gemini
Jones first tested the free version of Claude Sonnet 4.6 and Google Gemini on 30 April 2026, then repeated the experiment approximately six weeks later. Gemini had meanwhile been updated with its 3.5 Flash model.
The later responses were not fundamentally different, but they had changed noticeably in content and thoroughness. This illustrated that the answer to the same question can vary according to the model, its version, available tools, and the date and context in which the prompt is submitted.
Claude’s use of identifiers
Claude has facts about Internet identifiers embedded in its model weights. It knows, for example, that .de is Germany’s country-code top-level domain, which IP address ranges are reserved for private networks, and that IANA maintains authoritative protocol parameter registries.
In its basic operating mode, however, Claude does not itself resolve domain names, query registries, maintain active IP connections, or possess a persistent network identity. It operates as a stateless mathematical model producing text.
When a person accesses Claude, the surrounding system depends on the Internet identifier infrastructure. A browser looks up claude.ai, the DNS returns the information needed to reach Anthropic’s service, and IP addresses and related allocation systems enable the connection. The .ai domain is Anguilla’s country-code top-level domain, selected by Anthropic for its association with artificial intelligence.
If Claude invokes web search or retrieval, external systems interact with the DNS on its behalf. Jones therefore characterized Claude as a consumer and indirect invoker of the identifier system, rather than a participant in its operation or maintenance.
Gemini and search grounding
Gemini differs because it can use Google Search grounding and operates within Google’s extensive back-end infrastructure. When current information is needed, it may draw upon Google Search and conduct real-time retrieval.
This gives Gemini a more direct path to current Internet information than a model relying exclusively on static training data. DNS resolution still occurs in the surrounding Google infrastructure rather than as an intrinsic function of the language model.
New domains and stale knowledge
Jones next asked what happens when new top-level domains are delegated and how frequently each model checks the DNS to verify its responses.
Claude does not automatically know that a new TLD has been added after its training cutoff. Unless web search or web retrieval is enabled, it does not regularly check the DNS and cannot independently verify whether a domain or delegation is current.
Gemini’s integration with Google Search allows it to make on-demand queries and retrieve more current information. Nevertheless, the reliability of its response still depends on whether search grounding is invoked and whether the information retrieved is accurate.
Jones concluded that identifier information held inside an AI model remains static unless the model has access to current external systems. It may contain errors, overlook recently introduced identifiers, or produce different responses at different times. Models without active web tools will not automatically recognize new domains, protocols, or other DNS changes.
Operational questions
Jones asked registry and registrar operators whether they had observed increased server queries associated with AI systems. He also invited Internet service providers to consider whether agentic AI was becoming a new category of network user and whether its activity could be distinguished in traffic patterns.
These questions matter because automated agents could generate new forms or volumes of DNS activity, access services autonomously, or participate in suspicious domain registrations. Jones presented these as emerging questions rather than areas in which ICANN already had definitive answers.
Scope of unique identifiers
In response to Riyadh Zehrah, Jones clarified that “unique identifiers” encompassed the full range of resources coordinated through the IANA functions, including top-level domains, IP addresses, autonomous system numbers, and protocol parameters.
He acknowledged that the subject was relatively new and that his aim was to translate a technical issue into an accessible presentation. He again recommended the ICANN OCTO blog as a useful framework for further discussion.
Accuracy and model evolution
Fanaka Chidakwa asked how the models might develop to reduce errors and what the greatest challenge would be. Jones observed that the systems were constantly changing and incorporating new information. The differences between his April and June results demonstrated how much their responses could evolve within only six weeks.
He did not identify a single solution to the problem of errors, but emphasized that users should expect continuing changes in both model capabilities and answer quality.
Identifying and blocking AI queries
Eberhard Lisse asked how operators could recognize AI-generated queries if they wanted to block them. Jones said this was an issue that network and service operators would need to incorporate into their operational planning, but he did not have a specific technical method to recommend.
Standards and the IETF
Elliot Hollerton-Hill asked whether the Internet Engineering Task Force was developing standards concerning AI use of the DNS. Jones was not aware of specific work, but expected AI and DNS interactions to become a topic at future IETF meetings.
He also anticipated continued discussion at ICANN meetings and other community events throughout 2026 and beyond.
AI-assisted DNS abuse
Alexander Mayrhofer suggested that AI agents could soon be used to create suspicious or malicious domain registrations and asked whether existing anti-abuse systems were sufficient.
Jones could not say whether the current regime fully covered this possibility. He noted that AI and DNS abuse had been discussed during ICANN86 in Seville and expected it to remain an active area of community work.
A DNS-focused ICANN model
Mikhail Anisimov asked whether ICANN might develop or commission an AI model trained specifically on DNS data, potentially for domain abuse mitigation.
Jones said the question would be better addressed by ICANN’s OCTO team. He did not know whether it planned to combine language models with Domain Metrica or its other technical tools. Schittek noted that Siôn Lloyd was scheduled to discuss Domain Metrica in a future webinar, when the question could be raised again.
Keeping models current
Schittek asked what the community could concretely do when an AI model failed to recognize a newly created domain. Jones did not offer an immediate remedy.
She then asked what had most surprised him during the comparison. He identified the amount by which the models’ responses changed between April and June. Although their overall explanations remained similar, the later responses were more thorough. Users should therefore not expect an AI model to provide a fixed or reproducible answer: its output will vary according to what is asked, when it is asked, and which version of the system responds.
Closing
Schittek announced that Jones’s presentation had already been published and that the webinar recording would be posted shortly afterward. The session concluded the current season of the ICANN Webinar Series for EMEA, with the series scheduled to return at the end of August 2026.
RESOURCES
Artificial Intelligence and the Work of ICANN — March 2026 OCTO blog post by Matt Larson, the framing document Patrick Jones returned to twice
ICANN EMEA webinar page — slides and recording for this session
ICANN Office of the Chief Technology Officer — the OCTO team Jones referred the custom-model question to
ICANN Domain Metrica — DNS measurement platform raised in the exchange with Mikhail Anisimov
Inviting ccTLD Operators to Join Domain Metrica — background from Siôn Lloyd, who presents on it in the autumn
ICANN86 Policy Forum, Sevilla — where AI and DNS abuse sessions were held earlier in June
ICANN DNS Abuse Mitigation — relevant to Alexander Mayrhofer’s question on the current anti-abuse regime
IANA Protocol Registries — the authoritative protocol parameters Jones cited as built into model weights
Model Context Protocol — the agent-to-tool bridge described in the terminology slide
Claude — one of the two models tested, on Anguilla’s .ai ccTLD