As generative AI systems produce more answers, summaries and synthetic content, the reliability problem is increasingly shifting upstream toward the information those systems consume.
Geo founder Yaniv Tal argues that many failures associated with AI are rooted less in the models themselves than in the structure of the internet. In his view, online information often loses provenance, flattens distinctions between credible and weak sources, hides disagreement and increasingly recycles model-generated material back into future training data.
Tal’s central argument is that AI systems inherit these weaknesses from the information environment around them.
Provenance Can Disappear as Information Spreads
One of the main problems identified by Tal is the loss of provenance.
A claim can be copied from one website to another, rewritten, summarized and scraped repeatedly before eventually entering a training dataset. By that point, the original source may be difficult or impossible to recover.
That creates a major reliability problem.
A model may retain the statement itself while losing the chain of evidence behind it.
Without a clear source trail, users have less ability to distinguish original reporting, research, commentary and recycled information.
Different Sources Can Enter the Same Pipeline With Similar Weight
Tal also argues that authority becomes flattened when information moves through conventional web pipelines.
A peer-reviewed research paper, a corporate announcement and an anonymous forum post may all enter a dataset as text.
Yet those sources do not carry the same evidentiary value.
If a system does not retain information about who produced a claim, under what conditions and with what supporting evidence, it may struggle to differentiate between rigorous evidence and unsupported assertion.
AI Can Compress Legitimate Disagreement Into a Single Answer
Another weakness is the way language models can combine conflicting views into one apparently coherent response.
In areas where qualified experts genuinely disagree, that can create false certainty.
Rather than showing that multiple credible interpretations exist, a model may synthesize them into a single answer that sounds more definitive than the underlying evidence justifies.
Tal argues that preserving disagreement is therefore essential.
Reliability does not always mean identifying one universally correct answer. In some cases, it means showing that several well-supported positions remain unresolved.
Synthetic Content Creates a Feedback Problem
The fourth weakness concerns AI-generated material entering future data supplies.
As more synthetic articles, summaries and posts appear online, later systems may increasingly train on material produced by earlier models.
This can make it harder to identify the original human-generated source of a claim.
Errors may also be repeated.
Over time, repeated model-generated mistakes can be amplified rather than corrected.
Research Has Identified the Risk of Model Collapse
The concern is not limited to Geo’s own framework.
A 2024 Nature study examined what happens when generative models repeatedly train on output produced by earlier generations.
The researchers described a process called model collapse, in which models gradually lose information about the original data distribution.
According to the study summarized in the source, less common material began disappearing in earlier stages, while later model generations produced distributions increasingly different from the original data.
That suggests continued access to original human-created information remains important as synthetic content expands.
Geo Proposes Separating Claims, Sources and Evidence
Geo’s proposed solution is structural.
Instead of treating a piece of information as a single unit, the platform separates the claim itself from the person making it and the evidence supporting it.
Relationships between entries can then show whether one statement supports another, contradicts it or directly responds to a disputed argument.
That design aims to preserve the reasoning chain behind information rather than merely storing its final conclusion.
Human Judgment Would Rank Arguments Without Deleting Alternatives
The key role for humans is not to eliminate disagreement.
Tal’s proposal is to use human judgment to assess the relative strength of claims while retaining competing perspectives.
That would allow stronger reasoning to rank higher without erasing credible alternatives.
The goal is therefore not a single central authority deciding what is true.
It is a structured system in which evidence, reputation and reasoning remain visible.
Geo Uses Community-Governed Knowledge Spaces
Geo organizes information into independent communities known as Spaces.
The platform currently lists initial Spaces covering crypto, health, AI, education, world affairs and US politics.
Members can contribute to discussions.
People with relevant knowledge can also apply to become editors.
According to Geo, curators evaluate sources, connect arguments and add structured information to an open knowledge graph.
The System Builds on Geo Genesis and GRC-20
Geo was developed from Geo Genesis, which entered early access in January 2025.
Its default governance structure divided users into Editors and Members.
Editors had voting authority, while Members could contribute information to individual Spaces.
The system was built using the Aragon OSx governance framework.
It also followed the release of GRC-20, a standard designed to represent connected knowledge.
The Graph said in June 2025 that GRC-20 could allow communities to publish structured information onchain and retrieve it through subgraphs and Substreams.
Reputation Is Intended to Follow Contributors
Tal says expertise would not be assigned by one central organization.
Instead, contributors would build reputational records through their work.
Geo intends for that reputation to follow individuals across different Spaces.
The underlying idea is that accountability improves when claims are attached to identifiable people with visible histories.
That contrasts with information appearing anonymously or without a record of prior accuracy.
Human Governance Introduces New Risks
The approach does not eliminate governance problems.
A community-run system must still determine who is genuine, who is qualified and whether participants are acting independently.
Open governance networks can also face Sybil attacks, where one person creates multiple identities to influence votes, rankings or rewards.
AI itself can make such attacks easier by generating convincing online personas at low cost.
Identity Verification Comes With Trade-Offs
The source notes that biometric checks, social trust networks and zero-knowledge identity tools have all been proposed as ways to verify unique participants.
Each method creates trade-offs involving privacy, accessibility, attack resistance and centralization.
That means human verification is not a simple replacement for automated systems.
It introduces its own design problems.
Geo Favors Contribution Records Over Anonymous Claims
Tal says Geo would rely heavily on contribution histories and domain-specific communities.
Anonymous material would not carry the same status as claims linked to people with public records.
Editors would apply within individual Spaces.
Members would participate in governance around the subjects they follow.
The system therefore attempts to make expertise contextual rather than universal.
Crypto Provides a Useful Test Case
Crypto is one of Geo’s first Spaces because blockchain data offers a clear distinction between verifiable records and interpretation.
A blockchain can confirm that a transaction occurred at a particular address and block.
It cannot, by itself, establish who controls an address, why funds moved, whether the activity was organic or what a project intends to do next.
Those surrounding explanations remain claims requiring separate evidence.
Verifiable Data and Narrative Should Not Be Treated the Same
That distinction is important in digital-asset markets.
Project announcements, partnership claims, forecasts and interpretations of token movements can appear beside verifiable onchain records.
Tal argues that reliable systems should label those categories differently rather than presenting them with the same level of authority.
A transaction record and a marketing claim about that transaction are fundamentally different types of information.
US Guidance Also Emphasizes Provenance
Some of Geo’s ideas overlap with recommendations from the US National Institute of Standards and Technology.
NIST’s Generative AI Profile, published in July 2024, advises organizations to document training-data sources and monitor the origin of generated content.
It also recommends evaluating how provenance systems interact with human reviewers.
Domain experts and affected communities are included in certain assessment processes.
Provenance Alone Does Not Establish Truth
Content provenance standards such as C2PA address a related but narrower problem.
C2PA can cryptographically preserve information about who created or modified a digital asset and how it changed.
But verified provenance does not determine whether the claim itself is true.
It can confirm that an assertion is attached correctly and has not been tampered with.
That is different from evaluating the quality of the reasoning behind it.
Human Verification Is Better Viewed as an Additional Layer
The strongest implication is that human verification may be most useful when it complements, rather than replaces, automated systems.
Machines can process large volumes of information.
Humans can contribute judgment about expertise, context, disagreement and accountability.
The challenge is designing governance that preserves those benefits without introducing manipulation, centralization or arbitrary gatekeeping.
Conclusion
Geo’s approach highlights a broader problem for AI reliability: the quality of model answers depends heavily on the information structures surrounding them.
Lost provenance, flattened source authority, compressed disagreement and synthetic-data feedback loops can all weaken the reliability of generated answers.
Human verification offers one potential response by reconnecting claims to evidence, reputation and accountable contributors.
Final Takeaway
The most important insight is that improving AI reliability may require more than better models. It may also require better information infrastructure. Systems that preserve original sources, distinguish evidence from opinion, expose legitimate disagreement and attach accountability to claims could give future AI systems a more trustworthy foundation from which to generate answers.
