A position paper posted on arXiv on 18 August 2026 argues that AI lock-in, the progressive loss of human skills and system resilience as people and institutions depend more heavily on AI, is already emerging and needs to be treated as a safety priority before dependencies become irreversible S¹. The authors contend the threat spans individual skill atrophy, societal fragility, and national-scale infrastructure failure. The question they raise is one nobody has a clean answer to: what happens when the AI systems we have built our workflows around go down, or get cut off?

My read: This is the first paper I've seen that frames AI dependency as a safety problem in the same breath as alignment and regulation, rather than a productivity or labour-market concern. I'm skeptical of the claim that lock-in is "already emerging" at national level, because the paper offers scenarios rather than documented case studies. But the core argument, that we are building dependencies faster than we are building fallbacks, is the kind of thing that is obvious in hindsight and invisible until something breaks.

What the paper actually says

The authors start from an observation about the field: AI safety research has concentrated on two things, technical alignment (making sure AI outputs match human intent) and regulating the societal impacts of generative AI S¹. Both matter. Neither, they argue, covers the slow erosion that happens when people and organisations hand off tasks to AI and then forget how to do them without it.

They call this gap "AI Lock-In" and define it as excessive reliance on AI systems that leads to human deskilling, reduced capacity for independent functioning, and systemic vulnerabilities when those systems become unavailable or compromised S¹. The framing matters because it shifts the risk from "the AI does something bad" to "we can no longer do the thing without the AI."

The paper traces how lock-in escalates across three levels S¹:

  • Individual: skills atrophy as people delegate cognitive tasks to AI tools
  • Societal: institutions and workflows restructure around AI availability
  • National: critical infrastructure embeds AI dependencies that create single points of failure

The authors argue this is not a future risk but a process already underway at all three levels S¹. They also warn that service disruptions or geopolitical conflicts could amplify the damage dramatically, for instance if a country depends on AI systems hosted by a rival nation and access is severed S¹.

Scenarios, not case studies

The paper draws on "detailed scenarios" to investigate how lock-in emerges and escalates S¹. The abstract does not specify whether these are fictional thought experiments or analyses of real events. The evidence confirms that no documented, real-world national infrastructure failures caused by AI lock-in are cited, and no specific AI systems or vendors are named as confirmed sources of lock-in.

This is a position paper, not an empirical study. The claims are normative arguments about what the field should prioritise, not measurements of what has happened. The distinction matters: the authors are saying "prepare now," not "here is the data showing it has already cost X lives or Y dollars."

Why the framing matters

The paper's central contention is that proactively addressing AI lock-in before dependencies become entrenched or irreversible is essential for preserving individual autonomy and national security S¹. That is a strong claim. It positions AI dependency as a strategic risk on par with cybersecurity or supply-chain resilience, rather than a workplace productivity concern.

The authors also provide guidance on mitigating and preparing for lock-in risks at each level, individual, societal, and national S¹. The abstract does not detail what those mitigations are, but the framing suggests they go beyond "use AI less" and into structural safeguards, redundancy, and maintained human capability.

What to do about it

Consider a radiology clinic that has spent two years integrating an AI model into its workflow for preliminary scan readings. The radiologists still sign off on every report, but over time they see fewer raw images and more AI-annotated summaries. The model is fast and convenient. Six months in, the junior staff who trained alongside the AI are faster at reading AI output than at reading unassisted scans. Two years in, the senior radiologists trust the annotations enough to skim rather than study. If the model goes offline, or the vendor raises prices tenfold, or a regulator flags a systematic bias that forces a shutdown, the clinic does not revert to its pre-AI workflow. It reverts to a degraded version of it, staffed by people who have partially forgotten how to work without the tool.

That is lock-in at the individual and organisational level. The paper's argument is that the same dynamic is playing out at larger scales, and faster than anyone is tracking.

One practical thing a reader could do this week: pick one task you currently delegate to an AI tool and do it manually, start to finish, without the assistant. Time yourself. If it takes dramatically longer than you expect, or if you find yourself reaching for the tool partway through, you have a data point on your own dependency curve.

What we don't know yet

Several things remain unclear. The paper's scenarios are not characterised in the abstract as real or hypothetical, so the strength of the evidence base is uncertain until the full text is read. No specific AI systems, vendors, or documented infrastructure failures are named, which makes it hard to assess whether the national-level claims are grounded in observed events or projected from first principles. The paper is unreviewed and single-sourced, posted to arXiv as a position statement rather than submitted as empirical research S¹.

The deeper question is whether any government or regulator has formally recognised AI lock-in as a distinct policy priority. The evidence indicates none have, at least not in a way that surfaces in this paper's scope. If the argument gains traction, the first sign will be a policy document or regulatory framework that names dependency and deskilling as risks separate from bias, privacy, or alignment.

The next signal: community response and citation patterns on arXiv in the coming weeks, which will show whether other researchers treat this as a serious research agenda or a thought piece. We'll check whether the scenarios hold up under scrutiny and whether the mitigation guidance is concrete enough to act on.

If this kind of reporting is useful, subscribe to keep reading.

Sources: S1 — Position: AI Lock-In Is in Progress, and We Must Be Prepared · P2 — blucz/beprepared · P3 — The Lock-In Phase Hypothesis: Identity Consolidation as a Precursor to · P4 — HanseulJo/position-coupling · P5 — PKU-Alignment/ProgressGym

More from Not A Tech Guy

Generated from an audited evidence pack with primary-source research. Social-media items are discussion signals, not verified facts. Nothing here is financial, legal or medical advice.