Geo founder Yaniv Tal has identified four weaknesses in online information that he says make AI answers unreliable: lost provenance, flattened authority, hidden disagreement, and repeated model-generated errors.
Summary
- Geo separates claims, sources, and supporting evidence inside community-governed knowledge Spaces.
- Tal says human judgment can help rank credible reasoning without removing competing views.
- A Nature study found that repeated training on synthetic material can cause model collapse.
- NIST recommends tracking training sources and incorporating expert human feedback into AI systems.
Geo founder Yaniv Tal told crypto.news that unreliable AI answers often begin with the material models receive, arguing that the internet was designed to distribute information rather than preserve its authority, origin, or accountability.
“AI doesn’t have a truth problem, the internet does,” Tal said.
According to Tal, information loses critical context as websites scrape and republish it. A claim may pass through several pages before entering a training set, leaving a model with the statement but no clear route back to its original source.
Authority also becomes difficult to measure when a research paper, company announcement, and anonymous forum post enter the same data pipeline as text. Tal said models may then treat material with different standards of evidence as if it carries similar weight.
“Provenance collapses. A claim gets scraped, restated, and re-scraped until the original source is unrecoverable,” he said.
Geo founder identifies four failures behind unreliable AI answers
Disagreement creates another problem because language models often combine competing positions into one response. Tal said such compression can hide genuine disputes among qualified experts, giving users a single confident answer without showing that credible alternatives exist.
The fourth weakness arises when AI-generated material returns to the data supply used by later models. Errors can survive repeated publication, while synthetic articles, posts, and summaries make it harder to locate the human-produced material from which a claim originated.
“Models increasingly train on output from other models, so errors don’t just persist, they amplify,” Tal said.
Independent research has documented a related risk. A 2024 Nature study examined what happens when generative models repeatedly learn from material produced by earlier models. Researchers described “model collapse” as a process in which systems gradually lose information about the original data distribution.
Less common material began disappearing during the early stages of the experiments, according to the paper, while later model generations produced distributions bearing little resemblance to the source data. Researchers said access to original, human-produced information remains important as AI-generated material spreads across the internet.
Tal does not classify the four weaknesses as failures confined to model design. In his account, models reproduce flaws already embedded in their training material, including missing sources and a lack of visible distinctions between evidence, opinion, and promotion.
Human verification would rank claims without removing disagreement
Geo’s proposed answer starts by separating a statement from its author and the evidence supporting it. Relationships between entries can show whether one argument supports another, contradicts it, or responds directly to a disputed point.
Once those links are recorded, Geo can order arguments by their assessed strength while retaining competing material beneath them, Tal said. The structure is intended to keep disagreement available without displaying every contribution as an undifferentiated comment feed.
“Pluralism doesn’t have to mean noise,” Tal said. “We think several well-structured competing perspectives is often the more honest answer.”
Geo organizes information through independent communities called Spaces. Its current website lists initial Spaces covering crypto, health, AI, education, world affairs, and US politics.
Members can contribute and participate in discussions, while people with relevant knowledge can apply to become editors. According to Geo, its curator program asks participants to evaluate sources, connect arguments, and add structured information to an open knowledge graph.
The model was developed from Geo Genesis, which entered early access in January 2025. As previously reported, its default governance structure divided participants into Editors and Members. Editors received voting authority, while Members could contribute information to individual Spaces.
Built on the Aragon OSx governance framework, Geo Genesis followed the release of GRC-20, a standard for representing connected knowledge. The Graph said in June 2025 that GRC-20 could let communities publish structured information onchain and retrieve it through subgraphs and Substreams.
Tal, who co-founded The Graph before starting Geo, said expertise would not be assigned by one central authority under the proposed system. Contributors would build a reputation through their work, and Geo intends for that record to follow a person between Spaces.
“Human judgment is the scarce input now, not the redundant one. Machines generate more content than people can process, and almost none of it has anyone accountable for its truth.”
Community governance still faces an identity problem
A human-run knowledge network must still determine whether contributors are genuine, qualified, and independent. Open governance systems can face Sybil attacks, in which one participant creates multiple identities to gain influence over voting, rankings, or rewards.
A June 2026 explainer examined how biometric checks, social trust networks, and zero-knowledge identity tools attempt to verify unique people. Each method involves trade-offs between privacy, attack resistance, accessibility, and control, while AI can help attackers create convincing personas at low cost.
Tal said Geo would rely on contribution records and domain-specific communities rather than allowing anonymous material to carry the same status as claims attached to people with public histories. Editors would apply through individual Spaces, and members would take part in governing the subjects they follow.
“Skin in the game without financial stakes changes the quality of what people are willing to put their name to,” Tal said.
The approach does not remove the need to decide who selects editors, how communities resolve coordinated manipulation, or whether expertise earned in one subject should carry weight in another. Tal described the gatekeeper issue as real but argued that independent Spaces would prevent one institution from controlling every field.
Human involvement has also gained attention in crypto security, where automated tools may detect malicious activity, but users still authorize transactions. In May 2026, a Ledger security report described manual verification and clear transaction displays as safeguards against AI-assisted phishing, fake interfaces, and automated scams.
Onchain records separate transactions from surrounding claims
Crypto is one of Geo’s first Spaces because blockchain activity offers a direct test of the difference between verifiable records and human interpretation, Tal said.
A blockchain can confirm that a transaction occurred at a recorded address and block. It cannot, without additional evidence, establish who controls the address, why funds moved, whether activity was organic, or what a project plans to do next.
“On-chain data is verifiable by construction. A transaction happened or it didn’t,” Tal said. “Everything wrapped around it is claims.”
Project announcements, partnership descriptions, market forecasts, and explanations of token movements therefore require separate treatment, according to Tal. Geo intends to attach such claims to named contributors and preserve their records, including earlier statements that proved wrong.
Promotional activity makes the distinction especially important in digital-asset markets, where teams may present selected metrics beside verifiable transaction data. Tal said a trustworthy knowledge layer should label the categories rather than presenting both in the same voice.
US guidance also calls for provenance and human review
For US developers and companies deploying generative AI, the National Institute of Standards and Technology has recommended several controls that overlap with parts of Tal’s argument.
NIST’s Generative AI Profile, published in July 2024, advises organizations to document training-data sources, monitor the origin of generated material, and evaluate feedback between provenance systems and human reviewers. The voluntary framework also calls for domain experts and affected communities to take part in certain assessments.
Content provenance standards offer another method for recording where digital material originated. The Coalition for Content Provenance and Authenticity uses cryptographically signed Content Credentials to preserve information about who created or changed an asset and how it was edited.
Under the C2PA specification, however, validated provenance does not determine whether an attached claim is good, bad, or truthful. The standard verifies that assertions are connected to the underlying asset, correctly formed, and free from tampering.
