We Can’t Control AI If We Can’t Control Where It Gets Its Information From
We Can’t Control AI If We Can’t Control Where It Gets Its Information From
Current artificial intelligence safety research heavily concentrates on behavior, yet every action begins with a judgment about what information deserves belief. As machine intelligence rapidly advances and gains autonomous capabilities, the obscure pedigree of web data and compressed authority create severe provenance problems.
Artificial intelligence safety research has historically concentrated on behavior, monitoring actions, and establishing human approval checkpoints. However, every autonomous action begins one step earlier with a judgment about what information deserves belief. OpenAI chief scientist Jakub Pachocki warned that current alignment and monitoring techniques may soon prove inadequate as machine intelligence accelerates. Autonomous systems now draw on search results, corporate records, and material produced by other AI models. Because the web values circulation over ancestry, information origins easily fade. Without a reliable system of evidentiary provenance, autonomous agents risk executing irreversible actions based on unverified premises.
The Obscure Pedigree of Web Data
Before an artificial intelligence sends a payment, rewrites software, or recommends a critical decision, it first assembles a picture of the world. That picture draws extensively from search results, corporate records, research papers, and material produced by other machine learning systems.
The modern internet was built for rapid circulation rather than strict ancestry. Information travels easily while its origins fade as it passes from publication to publication, making it difficult to reconstruct the route back to original evidence.
Compressed Authority
Peer-reviewed studies, corporate announcements, and anonymous forum posts arrive as identical text for an AI to combine.
Compounding Repetition
Generative models ingest material from other models, allowing weak claims to be republished until repetition creates artificial consensus.
Human Oversight and the Speed Gap
The effectiveness of human oversight depends entirely on the reviewer having superior information or a reliable means of checking what the AI utilized. In many settings, human reviewers access the same web that the AI searched, but only at a fraction of the speed.
While an artificial intelligence can scan thousands of pages instantaneously, the human responsible for approval must work through a summary and a handful of links. This capability gap risks turning supervision into a procedural signature rather than an independent examination.
Speed Imbalance
AI systems process thousands of pages while human supervisors struggle through brief summaries and limited links.
Procedural Review
High capability risks reducing human oversight to a routine signature instead of an independent examination.
Establishing an Evidentiary Discipline
Serious financial and institutional activity depends heavily on records that establish who acted, under whose authority, and with what audit trail. Autonomous artificial intelligence requires a comparable information discipline to ensure trust and accountability across operations.
Claims should carry robust provenance that survives copying and transformation, allowing agents to trace statements back to their origin. Furthermore, credible disagreement among reputable sources should remain intact to trigger further research or human intervention.
Auditable Knowledge
Institutions require clear audit trails to verify actions, matching the rigorous standards expected in traditional finance.
Preserving Uncertainty
When reputable sources diverge, systems should preserve competing views and confidence levels rather than forcing false consensus.
The Urgent Call for Shared Governance
OpenAI chief scientist Jakub Pachocki argues that current alignment and monitoring techniques may soon constrain the pace at which labs can responsibly scale AI tools. He expects voluntary slowdowns within the industry until shared safety thresholds and auditing mechanisms emerge.
Governance must operate at a shared level through independent scrutiny and broad participation from multiple institutions. Relying solely on the preferences of a single provider leaves supervisors judging decisions whose foundational evidence remains opaque.
Why It Matters
The rapid evolution of machine intelligence exposes a fundamental vulnerability in how systems gather and process information. By inheriting the internet's weakest habits around speed and replication, autonomous agents risk acting upon compressed claims and erased uncertainties. Establishing rigorous information provenance is no longer optional; it is the vital foundation required for meaningful human control. Until artificial intelligence systems can transparently show where their knowledge originated, supervisors will remain blind judges of automated decisions.
We can’t control AI if we can’t control where it gets its information from (techradar.com) · AI chip smugglers exploit trade control loophole to bring banned HBM from South Korea into China — expensive memory chips packaged with cheap, unsanctioned processors to become legal exports (tomshardware.com)
Reporting reviewed through November 3, 2026.
