Forces

Pressures that push the open problems one way or another. Each is stated as its sources argue it; the catalogue records the argument and the evidence, not a verdict on where things will land.

• AI companies harvest what is published openly

  Training and answering with AI needs large amounts of text and data, much of it collected from the open Web. Hosts report heavier automated traffic and fewer human visitors, and some now block or charge crawlers.[src1][src2]

  on Who owns what the crowd built?: Pushes open projects toward paid access for heavy users, and some hosts toward closing their doors to crawlers.[src1][src2]
  on On what terms may anyone reuse it?: Revives the question of what an open licence allows, and calls for new machine-readable terms for AI use.[src1][src2]
  on Who pays for public data infrastructure?: Raises hosting costs while lowering the visits that bring donations.[src]
• Search answers instead of sending visitors

  Search engines and chat assistants increasingly answer questions themselves, so fewer people click through to the pages and datasets the answers come from.[src1][src2]

  on Why would anyone add structure?: Shrinks the reward of traffic that made publishers add markup in the first place.[src]
  on Who owns what the crowd built?: Cuts the human visits that bring volunteer editors and donations to commons projects.[src]
• General methods that scale keep winning

  Richard Sutton argues that across 70 years of AI research, general methods that make use of more computation have beaten approaches that build in human knowledge, and by a large margin.[src]

  on Hand-built meaning or learned meaning: Favours statistical learning over hand-written knowledge as computation grows cheaper.[src]
  on Expressiveness against tractability: Suggests that systems which predict at scale may displace those which derive with guarantees, whatever the cost to soundness.[src]
• Data protection law

  Laws such as the EU's GDPR and California's privacy act give people rights over data about them and oblige those who process it to respect their choices, which creates a need to state purposes and consent in a form machines can act on.[src1][src2]

  on Who controls data about me?: Makes consent and processing records a legal duty; how far the law will require machine-readable signals, such as browser consent settings, is still being negotiated in the EU.[src1][src2][src3]
  on Can the Web forget?: Turns forgetting into a legal right: delisting from search since 2014 and erasure under the GDPR.[src1][src2]
• Laws that require verified identity and labelled content

  Governments are writing trust into law. The EU's eIDAS 2 regulation of 2024 obliges member states to offer citizens digital identity wallets, and the AI Act requires AI-generated content to be marked in a machine-readable way from August 2026.[src1][src2]

  on Who vouches for it?: Puts states and regulated providers among those whose word a reader is asked to take, and creates demand for signed credentials and provenance marks.[src1][src2]
  on Who issues your name online?: Offers a state-issued identity as an alternative to platform accounts and self-made names.[src]
• Running infrastructure costs money and effort

  Servers, network-wide indexes and moderation all cost money and effort that most people and groups will not spend, so companies that run them for others tend to win. Moxie Marlinspike put it as 'people don't want to run their own servers, and never will'.[src1][src2]

  argued in Bluesky and the AT Protocol: Usable Decentralized Social Media
  on Central against decentralized: Pulls toward a few large hosts. AT Protocol's designers call the indexing that gives a whole-network view fairly expensive, and Marlinspike observes that users of decentralized systems end up trusting a handful of providers.[src1][src2]
  on Who pays for public data infrastructure?: Services that grow with their data, such as Wikidata's query service, need new engines and hardware long after launch money is spent.[src]
  on Querying a web, not a database: Answering arbitrary queries over billions of statements is expensive, which pushes public endpoints to split data or limit queries.[src1][src2]
  on Who issues your name online?: Few people will run a server or keep a domain to hold their own name, so platform accounts win by default.[src]
  on Can the Web forget?: Keeping old data costs money, so platforms may forget what users wanted kept; critics read Myspace's loss of 50 million songs as a choice not to pay for migration.[src]
• Laws that require portability and interoperability

  Regulators can oblige large platforms to let data and users move. The GDPR gave people a right to data portability in 2016, and the EU Digital Markets Act of 2022 requires the largest messaging services to interoperate. In April 2026 the Commission's first review of the Act declined to extend that duty to social networks.[src1][src2][src3]

  on Central against decentralized: Pushes toward portability where it applies. Its reach is set in periodic reviews, and social networks are outside it for now.[src1][src2]
  on Who controls data about me?: The GDPR's right to data portability lets people take their data from a platform, though not yet to keep it elsewhere and grant access from there.[src]
• Language models need knowledge to check against

  Language models produce fluent text that can be wrong, so their builders look for curated, structured knowledge to ground and check them against; the same models are also used to build that knowledge.[src1][src2]

  argued in Unifying Large Language Models and Knowledge Graphs: A Roadmap, W3C and the State of the Web in the AI Era
  on Hand-built meaning or learned meaning: Brings knowledge graphs back as a partner to statistical models rather than a rival.[src1][src2]
  on Agents acting for us: Makes structured data, vocabularies and policies more valuable as inputs and guardrails for agents.[src]
  on Can ordinary developers use it?: Puts a language model between developers and the data, so they may never write SPARQL themselves.[src1][src2]
  on Expressiveness against tractability: Creates demand for reasoners that can check what learned models produce.[src]
• Distrust of platform owners

  When a platform changes owner or policy, some of its users look for services that cannot be taken from them. Berners-Lee warned in 2010 that social sites were walling off their users' data. Bluesky's registrations grew from 2.89 million to 25.94 million in 2024, much of it from people unhappy with X.[src1][src2]

  argued in Long Live the Web
  on Central against decentralized: Brings users to decentralized services in waves; whether they stay is part of the open question. Bluesky's mobile app had 10.4 million monthly users in June 2026, down 27% on a year earlier.[src]
  on Names that outlive their hosts: Names tied to a company's domain or server last only as long as the company; that risk motivates persistent identifiers and DIDs.[src]
  on Who issues your name online?: Each lockdown, paywall or rename of a platform reminds people that their handle is the platform's to change; Bluesky's domain handles and the IndieWeb answer that.[src1][src2]
• Protocols change slowly, platforms quickly

  A shared protocol changes only when its implementers agree, while one company can change its platform at will. Marlinspike argues that a truly decentralized system 'becomes very difficult to change, and often remains stuck in time'.[src]

  on Central against decentralized: Gives central platforms the edge in features and speed. Standards groups such as the W3C Social Web Working Group and the IETF ATP Working Group are where decentralized protocols get changed.[src1][src2][src3]
  on Who sets the standard?: Fast-moving fields such as AI agents choose industry foundations, while slower, shared protocols move into standards bodies once many parties depend on them.[src1][src2]
• Public money for open internet software

  Much decentralized software has no business model and runs on grants. For a decade the EU's Next Generation Internet programme funded such projects through NLnet's NGI Zero; in 2026 NLnet paused new calls while NGI Zero gave way to a new Open Internet Stack effort.[src1][src2]

  on Central against decentralized: Pays for decentralized software that no company funds, including a hosting stack for the fediverse (NGI Fediversity).[src]
  on Who pays for public data infrastructure?: Grants are one of the few sources of money for upkeep rather than launch; when a programme pauses, so do its grants.[src]
• Search engines reward markup and set its terms

  Schema.org was founded by the major search engines, and publishers add much of its markup to earn their rich results, so the engines' choice of which types to support shapes what gets published.[src1][src2][src3]

  on Why would anyone add structure?: When Google retires a rich-result type, the reason to publish that markup goes with it.[src1][src2]
  on Who sets the standard?: Gives the search engines that founded schema.org the final say over the vocabulary most of the Web's markup uses.[src]