Four steps to avoid giving more than you gain from AI. #
Posted July 20, 2026 [ Reviewed by Michelle Quirk
](/us/docs/editorial-process)
Key points
- AI is increasingly part of how we formulate questions, interpret evidence, and decide what deserves attention.
- Our skills erode because AI is always available and independent practice feels inefficient.
- When the system fails, control returns to the person whose readiness has been weakened by its success.
We are taking part in a global social experiment, largely without signing up. Every day, drivers request directions for routes they once knew. Writers open a chat window before facing a blank page. Hospitals, courts, and employers lean further on model-generated scores. Each choice appears minor. Together, they alter where knowledge resides, who exercises judgment, and what happens when the system fails. Tools may empower us to do more, faster at lower cost. But what are the costs?
The consequences of generative artificial intelligence (AI) go far beyond the convenience once offered by elevators or calculators. AI has entered the thought process itself. It is increasingly woven into the chain through which we formulate questions, interpret evidence, and decide what deserves attention. But AI-powered empowerment is an oxymoron. When individuals and institutions continuously transfer creation, judgment, execution, and information to a system, they gradually lose the ability to create, grasp, and master the processes that the tool was only meant to make better. Ergo, the tool is empowered at the expense of those who deploy it.
Agency decay—the luring danger of compounding comfort #
Agency decay is not an irreversible verdict, yet. It resembles a limb kept in a cast: the capacity remains, yet strength and coordination diminish through disuse. Skills rarely disappear in one dramatic moment. They erode because the tool is always available, the shortcut is rewarded, and independent practice begins to feel inefficient.
The pattern predates generative AI. In “Ironies of Automation,” psychologist Lisanne Bainbridge showed already in 1983 that as automated systems become more reliable, human operators get less practice handling the situations for which they are still held responsible. When the system fails, control returns to the person whose readiness has been weakened by its success.
Memory research found a related effect. A study published in 2011 showed that when people expect information to remain available online, they remember where to retrieve it more readily than the information itself. GPS research likewise found that heavier lifetime reliance on navigation systems was associated with weaker spatial-memory performance. We did not suddenly become less intelligent; we adapted to the environment and stopped retaining what the environment promised to supply, to only remember the shortcuts to access that supply.
Early findings on generative AI deserve caution, yet they point in the same direction. A 2025 MIT Media Lab preprint reported weaker neural connectivity among participants writing essays with an AI assistant than among those working without one. Microsoft researchers surveying 319 knowledge workers found that greater confidence in generative AI was associated with less critical-thinking effort, while greater confidence in one’s own ability was associated with more. The constructive goal is hybrid intelligence: artificial assets extending natural intelligence, rather than gradually replacing its exercise.
The reverse power paradox #
AI appears to give us more information. In practice, the flow can reverse. The system receives our prompts, documents, corrections, preferences, operating rules, and process histories. It accumulates context while the user may retain less of the reasoning that produced the result. The machine becomes more informed about the work as the human becomes less able to reconstruct, challenge, or perform it independently.
At the organizational level, this becomes collective agency decay. Companies are not only buying software; many are embedding external AI systems into research, customer service, coding, recruitment, logistics, and strategic analysis. They transfer proprietary data and operational knowledge into tools that increasingly mediate how work is done. Over time, part of the organization’s practical intelligence migrates beyond its boundary.
The dependency is technical, commercial, and cognitive. Data may be difficult to move. Processes are redesigned around a provider’s models and interfaces. Staff stop practicing the underlying work. Switching supplier then means more than replacing a license: It can require extracting data, rebuilding workflows, retraining people, and recovering capabilities that have atrophied. Organisation for Economic Co-operation and Development (OECD) work on data portability links poor interoperability to higher switching costs and lock-in, while the National Institute of Standards and Technology (NIST) generative-AI risk profile explicitly advises organizations to identify overreliance on third-party data and systems, maintain fallbacks, and scrutinize vendor arrangements.
When oversight becomes theatre #
“Human oversight” is often presented as the safeguard. Yet there is a big “but”—oversight is only meaningful when the overseer understands the task, the system’s limits, and the evidence needed to contest its output. A person who cannot reproduce the reasoning, recognize a plausible error, or act without the tool is not supervising it. They are authorizing it.
This takes the aforementioned irony of sophisticated automation to an institutional scale. As systems grow more capable, organizations may retain a human approval step while removing the knowledge, time, and authority that make approval substantive. The European Union AI Act reflects the underlying requirement by linking oversight to competent, trained, and authorized human beings. A signature at the end of an opaque process cannot substitute for informed judgment. Ultimately, the human decision-maker must remain accountable.
AI-mediated peer review offers a vivid example. Automating parts of evaluation may relieve an overloaded academic system. Yet handing over the judgment that determines which ideas a field rewards also shifts intellectual steering toward patterns favored by the model... The same distinction applies in business: Delegating a first draft is different from delegating the decision about what should be built at all, whom it should affect, and which risks are acceptable.
The following is a simple four-step process to be and remain proactive players in a hybrid landscape:
The A-Frame takeaway #
**Awareness: **Notice which abilities, data, and decisions are moving from yourself, your team, and your institution into AI systems.
**Appreciation: **Value AI’s speed and reach, alongside human capacity—starting with your own—to critically question, contextualize, and imagine from scratch.
**Acceptance: **Friction is part of competence. Most relevant tasks must be learned and practiced before they can be delegated safely.
**Accountability: **Before using AI, ask not only, “Can it do this?” Ask: “Could we understand, defend, and continue this work without it—and why do we need AI in this context at all?”
We can still do something to preserve our place in the hybrid driver’s seat. But for how much longer?