Your Superintelligent Assistant or 

Someone Else’s Superintelligent Marketing Tool?

Meta’s Muse Wonder begins to turn Mark Zuckerberg’s vision of personal superintelligence into a working product. Whether it strengthens human capability or deepens dependence will depend on how the technology is built, how its business model develops, and how people learn to use it.

Will your personal AI become a superintelligent assistant that helps you understand more, accomplish more, and remain in control of your decisions? Or will it become a superintelligent marketing tool—one that understands your habits, insecurities, ambitions, and purchasing impulses well enough to influence you before you recognize that influence yourself?

There is a related question: Will using this assistant make you more capable, or will it gradually replace the mental effort through which capability develops?

Neither outcome is inevitable. An assistant that remembers your goals, handles tedious work, teaches unfamiliar subjects, and helps you evaluate alternatives could substantially expand your abilities. The same convenience could encourage you to stop practicing important skills, accept recommendations without examination, or become dependent on a service whose price and terms you do not control.

Meta’s announcement of Muse Wonder makes these questions increasingly practical. It brings together personalization, persistence, and creation—the beginnings of an assistant that knows something about you, continues working toward your goals, and helps turn intentions into finished work. Those capabilities fit Zuckerberg’s vision of personal superintelligence, but they do not yet establish that the vision has been achieved. [1–4]

The distinction matters. A powerful model is one component. A trustworthy personal assistant requires a much larger system: memory, tools, permissions, reliable execution, security, understandable controls, and incentives that support the person using it.

The documents behind the announcement

Three documents establish the progression from aspiration to implementation.

First, Zuckerberg’s July 30, 2025, essay, “Personal Superintelligence,” argued for putting advanced intelligence in individuals’ hands. Its central orientation was personal empowerment: helping people pursue their own ambitions and shape their lives. It also acknowledged that increasingly powerful systems would require consequential decisions about safety and openness. [1]

Second, his August 10, 2026, manifesto, “The Future Is for Everyone,” expanded that aspiration into a broader program. It addressed personal agents, creation, education, entrepreneurship, scientific progress, access, employment, infrastructure, security, privacy, openness, and alignment. The promise was no longer simply that more intelligent models would exist. It was that their benefits should become widely available. [2]

The earlier document is an essay rather than a scientific research paper. The later manifesto builds on its philosophy; it does not constitute an experimental demonstration that the philosophy will work.

Third, Meta’s October 12, 2026, Muse Wonder announcement and model card describe a model intended to support more personalized and sustained assistance. The model arrived after the September introduction of Muse, the agent product that supplies the surrounding execution environment. Wonder’s capabilities and Muse’s infrastructure therefore need to be understood together. [3–5]

These documents have different evidentiary roles. The essays express intentions. The product announcement describes features and availability. The model card reports evaluations and limitations. None, by itself, proves what long-term use will do to human independence, learning, or welfare.

What the new release means in ordinary language

The three characteristics can be expressed simply:

1. It gets to know you.
2. It keeps working toward your goals.
3. It helps you make things.

These descriptions are useful, provided their limits remain visible.

“It gets to know you” means that the system can retain and retrieve information that makes later assistance more relevant. It does not mean that it understands you completely or that its interpretation is necessarily correct.

“It keeps working toward your goals” means that an agent can maintain progress across multiple steps and continue certain tasks without requiring you to supervise every moment. It does not establish that the system can independently manage any project for days or weeks without mistakes.

“It helps you make things” includes turning instructions into documents, software, websites, and other outputs. A flyer or document layout is a reasonable example when the necessary creation tools are available. The quality of the finished result depends on the model, its tools, the clarity of the task, and the checks applied afterward. [3–7]

Together, these capabilities move AI beyond a series of isolated questions and answers. They begin to support an ongoing working relationship.

First: It gets to know you

A conventional chatbot may respond well within a single conversation yet lose useful context when that conversation ends. A personal assistant needs continuity: your preferred writing style, accessibility requirements, current projects, recurring obligations, and previous decisions.

Wonder’s Active Memory is intended to make relevant information from earlier interactions available when it is useful. This is different from placing every previous conversation into every new prompt. A practical memory system must select what matters, manage changes over time, and avoid letting outdated information override current instructions. [3,4]

From a systems perspective, useful personalization requires several distinct functions:

* Capturing information that may matter later.
* Representing and storing that information.
* Retrieving relevant details for the current task.
* Resolving contradictions and changes.
* Letting the user inspect, correct, and remove stored information.

Those are general requirements for dependable personal memory, rather than a claim that Meta has publicly disclosed every component of its implementation.

Consider a user who previously preferred lengthy technical explanations but now asks for a short message suitable for a phone screen. A good assistant should preserve the general preference while following the immediate request. Memory should inform assistance, not imprison the user in an old profile.

The potential benefit is substantial. People should not have to repeatedly explain their circumstances to a system that is supposed to assist them personally.

The danger is equally significant. The more useful the profile becomes, the more sensitive it becomes. A collection of ordinary conversations can reveal health concerns, financial pressure, family conflict, political interests, and emotional vulnerabilities—even when the user never explicitly labels any of those facts.

Personalization therefore creates two forms of power simultaneously: the user gains more relevant assistance, and the provider may gain access to a more revealing account of the user’s life.

Second: It keeps working toward your goals

Most substantial work does not fit into one prompt and one answer.

Planning an event, investigating a research question, developing software, or preparing a publication requires intermediate decisions. Sources must be checked. Drafts must be revised. Errors must be corrected. Sometimes the system must wait for information or permission before continuing.

Wonder’s Focus capability is presented as an improvement in maintaining task direction and working through obstacles. Muse supplies the surrounding agent environment in which actions can occur. [3–5]

Technically, this distinction separates the model from the operating system around it.

The model can propose a plan, interpret results, and choose a next action. The surrounding software must preserve task state, invoke tools, enforce permissions, record actions, and recover from failures. Scheduling and background execution also depend on that surrounding software.

A project lasting weeks therefore involves much more than a model producing a long answer. It requires durable records of what has been completed, what remains unresolved, which instructions still apply, and what the user has authorized.

Persistence also needs stopping rules. An assistant that continues pursuing the wrong interpretation of a goal can do more damage than one that stops early. More autonomy makes it increasingly important to distinguish reversible preparation from consequential external action.

The current release is a step toward sustained assistance. It should not be interpreted as a general guarantee of reliable, unattended management of every long-running project.

Third: It helps you make things

The third characteristic concerns execution: moving from discussing an outcome to producing it.

A user may want a website, a report, a document, a piece of software, or material for an event. The assistant’s value increases when it can help construct the requested result, inspect it, and revise it in response to feedback. Wonder’s development tools and Muse’s agent environment support this broader movement toward practical creation. [5,6]

However, creation is not a single capability.

Writing the words for a flyer differs from laying out the flyer. Generating website code differs from checking whether the website works on a phone. Producing a report differs from verifying its sources.

A capable system needs an iterative process: understand the request, create a draft, examine the output, detect defects, and revise. Where accuracy matters, it also needs evidence beyond its own confidence.

This is why tool access matters so much. A model may describe a beautiful design without producing a usable file. Conversely, an agent connected to suitable software may be able to generate, render, inspect, and improve the actual artifact.

The meaningful advance is the combination of reasoning and execution. Its practical measure is whether the person receives a useful, verifiable result.

How far does this implement the manifesto?

The relationship is strongest at the level of personal agency.

An assistant that remembers context can become more personally useful. An assistant that persists can undertake more substantial work. An assistant that creates can reduce the distance between an idea and a finished result.

That combination begins to implement the manifesto’s promise that advanced intelligence should help individuals act on their own goals. It does not yet fulfill the entire promise.

The manifesto reaches well beyond product features. Its principles of empowerment, invention, and a broader distribution of power are accompanied by ambitions involving agents, creative work, entrepreneurship, tutors, science, and affordable access. It also addresses economic transition, infrastructure, misuse, privacy, national leadership, openness, and governance of increasingly powerful systems. [2]

Each part raises a different test.

Personal agents and empowerment. Does the assistant reliably pursue goals the user has chosen? Can the user understand, redirect, and stop it? Completing tasks is valuable, but empowerment also requires meaningful control over which tasks are pursued.

Creation and entrepreneurship. Does the technology enable people to produce work or start ventures that would otherwise be beyond their reach? Generating an initial artifact is only the beginning. Commercial usefulness may require testing, maintenance, customer understanding, and accountability.

Education. Does the assistant improve what the learner can subsequently understand and do? A tutor that delivers correct answers may improve homework while weakening the learning that homework was intended to develop.

Scientific progress. Does the system contribute to discoveries that withstand independent scrutiny? Fluent explanations and plausible hypotheses are useful inputs, but scientific progress depends on validation.

Broad access. Can ordinary people obtain meaningful capability at an affordable price? Free access to a limited service does not necessarily provide equal access to sustained execution, better models, or more computing resources.

Employment and infrastructure. Who benefits from productivity gains, and who bears the costs of transition? These are questions about institutions, investment, and distribution that a model release alone cannot answer.

Safety, privacy, openness, and governance. Can the system become more powerful without making users or society more vulnerable? Technical controls, public commitments, and oversight must be evaluated by how they operate in practice.

The manifesto’s discussion of additional computing capacity through market mechanisms also raises a distributive question. Willingness to pay can allocate scarce resources, but it is not equivalent to educational need, social benefit, or equal opportunity. That is an implication of the proposed economic structure, not evidence that its eventual results are predetermined. [2]

More capable AI does not automatically mean more capable people

This is the central educational and cognitive issue.

An AI-assisted person can produce better work while learning less. Another person can use the same broad class of technology to understand more deeply, attempt harder problems, and develop new skills.

These outcomes are not contradictory. They reflect different relationships between assistance and mental participation.

Research on a purpose-built AI tutor provides an example of the positive possibility. Kestin and colleagues studied 194 students in a college physics setting and found greater immediate learning gains with the AI tutoring condition than with the comparison classroom instruction. The intervention incorporated a deliberate educational design; it was not simply unrestricted access to a general chatbot. The findings support the potential of well-designed tutoring, while leaving longer-term retention and broader generalization to further investigation. [9]

Research by Bastani and colleagues demonstrates the opposing risk. In a high-school mathematics experiment involving nearly 1,000 students, access to a general AI interface improved performance during practice but was associated with worse subsequent unaided performance. A tutor designed with learning safeguards substantially improved assisted practice and avoided that disadvantage, although it did not establish a corresponding improvement on the unaided examination. [10]

The lesson is not that AI necessarily teaches well or necessarily damages learning. The design of the interaction matters.

An assistant can give an answer, ask the learner to explain a step, supply a hint, challenge an assumption, or require an attempt before intervening. Those choices affect which cognitive work remains with the person.

Output, learning, retention, and atrophy are different things

The research proposal “Is AI Making Us Smarter or Dumber?” provides a useful framework for separating questions that are too often collapsed into a single judgment about whether AI is helpful. Its companion annotated bibliography supplies the broader research foundation. These are research-planning documents, not completed validation studies. [19,20]

The framework distinguishes assisted performance from independent capability. It also separates acquiring a skill, retaining it, transferring it to a different situation, and recovering it after disuse.

A person may produce an excellent AI-assisted report without acquiring the ability to evaluate the report independently. Another person may already possess that ability but gradually exercise it less. Those are different developmental problems.

The first can involve failure to acquire a skill. The second can involve loss of a previously demonstrated skill, or atrophy in the functional sense. Neither description, by itself, establishes physical brain damage.

Research on procedural skill retention and decay provides reasons to take disuse seriously. Earlier quantitative reviews, more recent meta-analysis, studies of automated cockpits, and simulated process-control research examine how performance changes when practice and task demands change. These studies do not prove that conversational AI produces the same effects, but they identify relevant mechanisms and measurement approaches. [13–16]

This distinction prevents two misleading conclusions.

One is that better output proves cognitive improvement. The other is that any reduced effort proves cognitive harm.

Reducing effort can be beneficial when it removes unnecessary work. It becomes more concerning when it removes the practice needed to build or preserve a capability the person still needs.

What current cognitive research does and does not show

A study by Shen and Tamkin examined developers learning to work with an unfamiliar programming library. AI assistance was associated with weaker immediate understanding and related skill performance, without a statistically significant average completion-time advantage. The findings raise concerns about outsourcing work during skill acquisition, but they do not establish that all AI-supported programming weakens expertise. [11]

Research by Lee and colleagues surveyed knowledge workers about their use of generative AI. Greater confidence in the AI was associated with less reported critical-thinking effort. Because the evidence is based on self-reports and associations, it cannot establish long-term cognitive decline. It does, however, identify a plausible pathway: confidence in the system may reduce the scrutiny applied to its output. [12]

Research on cognitive forcing functions addresses a possible response. Buçinca and colleagues found that interaction designs requiring more deliberate consideration could reduce overreliance in AI-assisted decisions. Some effective designs were also less favorably received by users. [17]

That tension is commercially important. The interaction that feels easiest may not be the interaction that best preserves independent judgment.

The ICAP framework offers another useful distinction: receiving information, performing an activity, generating explanations, and engaging interactively do not represent equivalent forms of cognitive participation. Applying that perspective to AI suggests that the quantity of use matters less than what the person actually does while using it. [18]

These findings justify serious attention to dependence and skill development. They do not justify pretending that researchers have already measured every long-term consequence of deeply personalized AI agents.

A better objective: Developmental assistance

If an assistant is meant to empower its user, its objective should extend beyond producing the next satisfactory answer.

Sometimes it should complete a task. Sometimes it should teach. Sometimes it should help the user practice. Sometimes it should challenge a conclusion or make uncertainty easier to inspect.

The research proposal distinguishes three complementary interventions: AI-supported subject education, education in effective AI use, and the developmental alignment of everyday assistance. [19]

That last category is particularly relevant to personal agents. A person may learn or lose practice during ordinary work without opening anything labeled a tutoring application.

For example, an assistant helping with a budget could simply supply totals. It could also explain the assumptions, let the user change them, and ask whether the result matches the user’s priorities. A research assistant could produce a polished conclusion, or it could help the user compare competing explanations and inspect evidence.

The appropriate approach depends on the person’s goal. Someone seeking emergency help with a deadline may reasonably want more direct completion. Someone developing a new professional skill may benefit from a more demanding interaction.

The proposal’s three dimensions—exposure and interaction, cognitive engagement, and capability amplification—are intended to help study these differences. Its categories and assessment schedules remain provisional. They should not be treated as established diagnostic scales. [19]

A useful evaluation would combine measures of assisted performance with retention and transfer assessments, and with independent checks where appropriate. Accessibility support should remain available: removing a screen reader, for example, would confound access with the capability being measured.

The AI should not be the sole judge of whether its own assistance makes the user more capable.

Privacy: What Meta has addressed

It would be inaccurate to describe Muse as having no privacy or security architecture.

Meta describes an isolated execution environment, controls around credentials, limits on permissions, and mechanisms for reviewing or governing sensitive actions. Its published security approach includes separation of privileges and monitoring intended to restrict inappropriate information flows and external actions. [5,7]

These controls address real risks. An agent that interacts with applications can expose information or take unwanted actions in ways that a text-only system cannot.

Meta also states that Muse’s virtual-machine and conversation data are not shared with its advertising systems. Its documentation distinguishes this from activity that can affect advertising indirectly—for example, a user-attributable visit or transaction on an external service. It also describes user controls over memory and files, and an opt-out for the use of sanitized task trajectories in training. [7]

Those commitments should be represented accurately. It would be misleading to write as though the company had announced that private agent conversations would be fed directly into advertising profiles.

But acknowledging safeguards does not end the analysis.

At launch, operational restrictions on access were not the same as a cryptographic guarantee that the provider could not access the environment. Meta described stronger confidential-computing protections as forthcoming. A planned protection should be evaluated as a commitment until it is deployed and independently assessed. [7]

Privacy: What still needs scrutiny

Security against an outside attacker is one problem. Protection from inappropriate use by the provider is another. Protection from excessive collection is a third.

A system could be secure from intrusion while accumulating more personal information than necessary. It could comply with its published policies while making those policies difficult to understand. It could offer deletion while leaving users uncertain about derived profiles, previously trained models, or information already transmitted to other services.

A mature personal assistant therefore needs understandable answers to practical questions:

* What information is retained, and for how long?
* Which inferences about the user are stored?
* Can the user correct a mistaken inference?
* What does deletion remove?
* Who can access information, under what conditions?
* Can the user move useful memory and project history elsewhere?
* How are changes to commercial use communicated and controlled?

These are evaluation questions, not assertions that every listed protection is absent.

Persistent memory raises the stakes because conversations become cumulative. A single interaction may reveal little. Years of interactions can reveal patterns that neither party would infer from one exchange.

The resulting asymmetry is important. The assistant may know much about the user, while the user knows comparatively little about how the assistant ranks recommendations or how its provider evaluates commercial success.
 Will the business model reward empowerment—or dependence?

The possibility of getting hooked should be taken seriously.

Free access can lower barriers and produce real benefits. It can also begin a progression in which a service becomes increasingly difficult to replace.

A user starts with occasional questions. The assistant learns preferences. Projects accumulate. Connected services become convenient. The user begins relying on remembered context and automated routines. Eventually, a limit on usage or capability creates a reason to upgrade.

That progression can reflect genuine value. It can also produce dependence and switching costs. Both may occur together.

Dependence has at least three relevant forms.

Practical dependence develops when important work relies on the service.

Cognitive dependence develops when a person increasingly relies on the service to frame problems, evaluate evidence, or make judgments that the person no longer practices independently.

Commercial dependence develops when accumulated history, integrations, and habits make it difficult to refuse higher prices or less favorable terms.

A company does not need to intend cognitive harm for these outcomes to arise. Revenue targets, engagement measures, product defaults, and convenience can shape behavior without an explicit decision to weaken users.

Nor should the analysis assume that users are passive. They can use the technology deliberately, continue learning, inspect important claims, and preserve alternatives. But responsibility is not distributed equally: providers control the interface, pricing, defaults, and much of the information about how the system works.

What is free, what costs money, and what remains uncertain?

Muse already has a free entry level with usage limits. The initial commercial offering also included paid consumer tiers reported at $20 and $100 per month. Those are announcement-period figures, rather than a guarantee of future pricing. [5,8]

Wonder’s developer pricing is a separate matter. Its launch documentation listed charges of $4 per million input tokens, $20 per million output tokens, and $0.20 per million cached input tokens. Developer usage charges should not be confused with the consumer subscription tiers or with Muse Code subscriptions. [6]

Wonder’s availability also differs from Muse’s availability. The model rollout initially prioritized specified developer and subscriber channels, while Meta described plans for broader access to specialized models. A free Muse account should not be assumed to provide the same capability or usage allowance as a paid Wonder offering. [3,6]

The answer to “When will Meta offer a free entry level?” is therefore that it already does. The unanswered question is how much capability that level will contain over time.

There is no basis here for giving a firm date when a free tier will equal the strongest paid offering.

The manifesto’s broad-access philosophy will ultimately be tested by practical limits: memory quality, tool access, execution time, model capability, and the amount of work users can complete before paying more.

Is anyone else moving in this direction?

Yes. Meta is part of a broader movement toward assistants that remember context, connect to personal information, and offer more capability through paid tiers.

OpenAI’s description of ChatGPT memory and “Dreaming” concerns making remembered information more useful over time, including background synthesis and updating. ChatGPT’s pricing structure provides both free and paid access with differing capabilities and limits. These offerings are relevant comparisons in personalization and access; they do not establish that ChatGPT and Muse have identical architectures or autonomy. [21,22]

Google’s expansion of Personal Intelligence similarly emphasizes more personalized assistance through permissioned connections to information in services such as Gmail and Google Photos, including broader availability for eligible personal accounts. [23]

The competitive direction is clear: assistants are becoming more persistent and personally informed.

The consequences are not settled. Competition may improve capability, affordability, and privacy controls. It may also encourage providers to make their ecosystems harder to leave.

The important comparison is therefore not simply which assistant has the strongest model. It is which system gives users the most useful combination of capability, control, portability, privacy, and economic choice.

What remains to be built?

Wonder begins implementing the three characteristics, but substantial work remains.

For “it gets to know you,” the challenge is dependable, correctable, and appropriately limited memory. Remembering more is not always better. Remembering accurately and forgetting appropriately matter just as much.

For “it keeps working toward your goals,” the challenge is reliable execution over time. The assistant must preserve context, recognize changed conditions, recover from failures, and seek human direction at the right moments.

For “it helps you make things,” the challenge is verifiable quality. A finished-looking result must also be usable, accurate, accessible where needed, and maintainable.

Across all three, the larger challenge is alignment with the user’s interests. That includes situations in which the user’s interests conflict with greater engagement, more spending, or deeper dependence on the provider.

The model card’s evaluations are useful evidence about particular capabilities and risks. They are not a substitute for measuring real-world reliability or the developmental effects of prolonged use. [4]

If the promise is human empowerment, evaluation eventually needs to ask whether people become better able to understand, decide, create, and act—not simply whether the assistant completes more tasks.

Remaining the master is an outcome to maintain

A free assistant can be an opportunity. It can also be a hook.
 The hook may be convenient access that encourages initial use. It may be accumulated memory that makes the service hard to replace. It may be a progression toward increasingly expensive subscriptions. It may also be the emotional relief of having a system supply answers whenever independent thinking feels difficult.

These possibilities do not erase the benefits. A user can become deeply committed to a tool because it helps the user learn, create, and participate more fully in life.

The distinction is whether that commitment expands or constrains the person’s agency.

Remaining in control does not mean refusing assistance or doing every task unaided. It means retaining enough understanding to judge important results, enough practice to preserve needed capabilities, and enough freedom to change course.

It also means that the provider must accept meaningful forms of user independence: rejecting a recommendation, spending less, choosing another service, exporting useful information, or stopping altogether.

Whether personal AI enhances cognition or encourages atrophy will depend partly on company intent and product design, and partly on how users engage with the opportunity. It will also depend on whether educational institutions, researchers, and public policy develop ways to distinguish genuine empowerment from attractive dependence.

Muse Wonder makes the manifesto more concrete. It does not settle the manifesto’s deepest questions.

The question is not simply whether the AI becomes more intelligent. It is whether the person becomes more capable of directing that intelligence—or becomes increasingly dependent on a provider that benefits from directing the person.

References

1. Zuckerberg, Mark. (2025, July 30). Personal superintelligence. Meta.
    https://www.meta.com/superintelligence/
2. Zuckerberg, Mark. (2026, August 10). The future is for everyone. Meta.
    https://about.fb.com/news/2026/08/the-future-is-for-everyone/
3. Meta. (2026, October 12). Muse Wonder. Meta AI Research.
    https://research.meta.ai/blog/muse-wonder
4. Meta. (2026, October 12). Muse Wonder model card. Meta AI Research.
    https://research.meta.ai/static/muse-wonder-model-card
5. Meta. (2026, September 8; updated September 30). Introducing Muse, a personal AI agent. Meta Newsroom.
    https://about.fb.com/news/2026/09/introducing-muse-personal-ai-agent/
6. Walters, Josh. (2026, October 12). Build with Muse Wonder. Meta for Developers.
    https://dev.meta.ai/resources/blog/build-with-muse-wonder
7. Sheasha, Tarek. (2026, September 8). Security and safety for AI agents: Our approach with Muse. Meta AI Research.
    https://research.meta.ai/blog/security-and-safety-for-ai-agents-our-approach-with-muse
8. Reuters. (2026, September 8). Meta launches AI agent that can access other apps, send emails, make payments. Reuters.
    https://www.reuters.com/business/meta-launches-ai-agent-that-can-access-other-apps-send-emails-make-payments-2026-09-08/
9. Kestin, Greg; Miller, Kelly; Klales, Anna; Milbourne, Timothy; and Ponti, Gregorio. (2025). AI tutoring outperforms in-class active learning: An RCT introducing a novel research-based design in an authentic educational setting. Scientific Reports, 15, Article 17458.
    https://doi.org/10.1038/s41598-025-97652-6
10. Bastani, Hamsa; Bastani, Osbert; Sungu, Alp; Ge, Haosen; Kabakcı, Özge; and Mariman, Rei. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26), e2422633122.
    https://doi.org/10.1073/pnas.2422633122
11. Shen, Judy Hanwen, and Tamkin, Alex. (2026). How AI impacts skill formation. arXiv preprint, arXiv:2601.20245. Submitted January 28; revised February 1, 2026.
    https://arxiv.org/abs/2601.20245
12. Lee, Hao-Ping (Hank); Sarkar, Advait; Tankelevitch, Lev; Drosos, Ian; Rintel, Sean; Banks, Richard; and Wilson, Nicholas. (2025). The impact of generative AI on critical thinking: Self-reported reductions in cognitive effort and confidence effects from a survey of knowledge workers. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, Article 1121, pp. 1–22. Association for Computing Machinery.
    https://doi.org/10.1145/3706598.3713778
    Author-hosted publication page: https://www.microsoft.com/en-us/research/publication/the-impact-of-generative-ai-on-critical-thinking-self-reported-reductions-in-cognitive-effort-and-confidence-effects-from-a-survey-of-knowledge-workers/
13. Tatel, Corey E., and Ackerman, Phillip L. (2025). Procedural skill retention and decay: A meta-analytic review. Psychological Bulletin, 151(6), 696–736.
    https://doi.org/10.1037/bul0000481
14. Arthur, Winfred, Jr.; Bennett, Winston, Jr.; Stanush, Pamela L.; and McNelly, Theresa L. (1998). Factors that influence skill decay and retention: A quantitative review and analysis. Human Performance, 11(1), 57–101.
    https://doi.org/10.1207/s15327043hup1101_3
15. Casner, Stephen M.; Geven, Richard W.; Recker, Matthias P.; and Schooler, Jonathan W. (2014). The retention of manual flying skills in the automated cockpit. Human Factors, 56(8), 1506–1516.
    https://doi.org/10.1177/0018720814535628
16. Kluge, Annette; Frank, Barbara; and Miebach, Julia. (2014). Measuring skill decay in process control—Results from four experiments with a simulated process control task. In D. de Waard, K. Brookhuis, R. Wiczorek, F. di Nocera, R. Brouwer, P. Barham, C. Weikert, A. Kluge, W. Gerbino, and A. Toffetti (Eds.), Proceedings of the Human Factors and Ergonomics Society Europe Chapter 2013 Annual Conference, pp. 79–94. ISSN 2333-4959. 
17. Buçinca, Zana; Malaya, Maja Barbara; and Gajos, Krzysztof Z. (2021). To trust or to think: Cognitive forcing functions can reduce overreliance on AI in AI-assisted decision-making. Proceedings of the ACM on Human-Computer Interaction, 5(CSCW1), Article 188, pp. 1–21.
    https://doi.org/10.1145/3449287
    Open-access manuscript: https://arxiv.org/abs/2102.09692
18. Chi, Michelene T. H., and Wylie, Ruth. (2014). The ICAP framework: Linking cognitive engagement to active learning outcomes. Educational Psychologist, 49(4), 219–243.
    https://doi.org/10.1080/00461520.2014.965823
19. Hayes, Dennis C. (2026, October 18). Is AI making us smarter or dumber? A research proposal for measuring human development and improving AI use through cybernetic feedback. Includes the complete references. SpartansFirst Technology Blog.
    Link: http://www.spartansfirst.org/Is%20AI%20Making%20Us%20Dumber%20or%20Smarter

20. Same as reference 19.
21. OpenAI. (2026, June 4). Dreaming: Better memory for a more helpful ChatGPT.
    https://openai.com/index/chatgpt-memory-dreaming/
22. OpenAI. (n.d.). ChatGPT plans and pricing. Consulted October 18, 2026. Pricing and included capabilities may change.
    https://chatgpt.com/pricing/
23. Google. (2026, March 17). Bringing the power of Personal Intelligence to more people. The Keyword.
    https://blog.google/products-and-platforms/products/search/personal-intelligence-expansion/