OpenAI chief scientist Jakub Pachocki seemed to issue a warning on September 6.
In “An Alien Mind,” he argued that current alignment and monitoring techniques may soon prove inadequate for the pace at which machine intelligence is advancing. He also described a future in which AI takes a growing role in improving AI, compressing the time available for human supervision and institutional response.
Much of AI safety research concentrates on behavior. For example, what an AI system is permitted to do, how its actions are monitored and where humans retain approval. Yet every action begins one step earlier, with a judgment about what information deserves belief.
Latest Videos FromTechRadar
That earlier step is incredibly important.
AI acts on information whose pedigree is often obscure
Before an AI sends a payment, rewrites software or recommends a decision, it first assembles a picture of the world. That picture can draw on search results, corporate records, research papers and material produced by other AI systems.
The web was built for circulation. Information travels easily, while its ancestry often fades as it passes from one publication to another. A claim can begin in a research paper, appear in a news story, be condensed into a blog post and surface months later in a forum thread. By the time an AI encounters the statement, the route back to the original evidence may be difficult to reconstruct.
Authority also gets compressed. A peer-reviewed study, a corporate announcement and an anonymous post can all arrive as text. An AI system can rank and combine them into a polished answer, while the reader sees little of the difference between documented evidence, advocacy and repetition. Disagreement adds a further complication.
Credible sources regularly reach different conclusions, especially in finance, medicine, politics and emerging technology. AI is usually rewarded for producing a useful answer, which creates pressure to reconcile conflict into a single sentence. Uncertainty becomes easy to erase precisely when it deserves to remain visible.
The consequences grew substantially once AI acquired the ability to act. An agent can place a trade, approve a purchase, alter software or send a message on someone’s behalf. An uncertain claim can therefore become the premise for an irreversible action.
Human oversight depends on better evidence
The effectiveness of human oversight depends on the reviewer having superior information or a reliable means of checking what the AI used. In many settings, the human reviewer has access to the same web that the AI searched, only at a fraction of the speed. The AI can scan thousands of pages while the person responsible for approval works through a summary and a handful of links. As capability rises, review risks becoming a procedural signature rather than an independent examination.
Pachocki writes that OpenAI’s ability to rely on chain-of-thought monitoring is diminishing as reasoning becomes more complex and as AI becomes more capable of manipulating its own reasoning process. He calls for safety thresholds that can be enforced by third-party auditors, government agencies or international bodies.
Those safeguards govern capability and behavior. Reliable oversight also requires an evidentiary record showing what information produced the conclusion, where each claim originated, how trustworthy the source is and where credible disagreement remains. Without that record, the reviewer is judging the answer more than the evidence behind it.
AI inherited the web’s weakest habit
AI absorbed an information culture built around speed, reach and replication. Search engines, social feeds, aggregators and reposts made access extraordinarily efficient. They also encouraged a world in which the same statement could appear hundreds of times while tracing back to a single uncertain source.
Generative AI intensifies that effect because AI systems now produce material that later AI systems can ingest. A weak claim can be generated, scraped, summarized and republished until repetition gives it the appearance of consensus.
This creates a provenance problem with compounding consequences. Every additional generation can separate a statement further from the evidence that gave rise to it. Once an autonomous agent treats that statement as reliable, the information chain becomes part of the decision chain. While for consumers, this can mean a poor recommendation, for a financial institution, insurer or government agency, the same weakness can affect money, access, security or public services.
Autonomous AI needs a system of provenance
Take finance. Serious financial activity depends on records that establish who acted, under whose authority, on which terms and with what audit trail. Institutions spend heavily on those records because trust requires more than a plausible outcome. Autonomous AI needs a comparable discipline for information.
Claims should carry provenance that survives copying and transformation. A person or an agent should be able to trace a statement back to its origin, inspect the evidence behind it and see how the statement changed as it travelled. Credible disagreement should remain intact.
Where reputable sources diverge, the system should preserve the competing views and the degree of confidence behind each one. For an autonomous agent, uncertainty can serve as a trigger for further research, human approval or a narrower course of action.
Accountability should travel with information as well. A claim from an identifiable institution with supporting evidence deserves different treatment from an unattributed assertion repeated across dozens of pages. The aim is to give AI a richer sense of evidentiary quality before it commits resources or acts on someone’s behalf.
Governance belongs at a shared level. Provenance standards, verification methods and accountability rules need independent scrutiny and participation from many institutions, rather than dependence on the preferences of a single provider.
Final thoughts
Pachocki argues that current alignment and monitoring techniques may soon constrain the pace at which labs can responsibly scale AI tools, and he expects voluntary slowdowns until shared safety thresholds emerge. He also warns that continued progress could carry AI into recursive self-improvement, with AI driving a larger share of its own development.
That leaves every company building autonomous systems with a fundamental task. Establish what the agent knows before it acts and make that knowledge auditable.
Behavioral safeguards remain essential. So do permissions, audits and human approval. Their value rises sharply when the information beneath every decision carries provenance, context and accountability. The next phase of AI safety, therefore, begins with the evidence an agent sees, the authority it assigns to that evidence and the uncertainty it preserves before taking action.
Autonomy built on verifiable information gives humans a stronger basis for control. Opaque information leaves supervisors judging decisions whose foundations remain difficult to inspect.
Pachocki is right to warn that the window for preparation may be narrow. The industry should use that window to build systems that can show where their knowledge came from before those systems are trusted to act on the world.
We’ve featured the best AI chatbot for business.
This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.
The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit
