McLeod Bob-Manuel
Menu
All papers

Human-AI interaction

Fragmented Intelligence: Why Personal Use of Multiple AI Systems Underdelivers

Published by
MacLeod Technologies
Date
October 2026
Author
Sotonye McLeod Bob-Manuel

Written with AI assistance. The author directed the research question and structure, checked every cited source, and takes responsibility for the content.

Working paper, October 2026. Comments welcome.

Abstract

Individuals now hold access to several general-purpose AI systems at once, yet the value of that access is limited by three structural constraints. We call them weights, harnesses and context. Weights: model capability is uneven across tasks, so no single system is best for everything. Harnesses: the scaffolding around a model shapes what it can do as much as the model does. Context: each system holds only a partial, often degraded, picture of the user. Drawing on field experiments, benchmark studies, human-computer interaction research and privacy theory, this paper describes how these constraints arise and how they compound. It also covers four secondary costs: token consumption, the absence of continuity in self-directed learning, privacy leakage across personal and work settings, and the difficulty of judging and securing AI output. We do not propose a system. We close with open research questions.

1. Introduction

Adoption of generative AI at work has been fast and largely bottom-up. In a 2024 survey of 31,000 people across 31 countries, 75% of knowledge workers reported using AI at work, and 78% of those users brought their own tools rather than relying on employer-provided ones (Microsoft & LinkedIn, 2024). Consumer adoption followed a similar curve: the share of U.S. adults using AI rose from 61% in 2025 to 64% in 2026 (Menlo Ventures, 2026).

The typical heavy user therefore does not use “an AI”. They use a chat assistant from one vendor, a coding agent in their editor, an autonomous agent for recurring tasks, and often an employer-provided assistant as well. Each was built, priced and evaluated in isolation. The research question of this paper is narrow: what prevents a person with access to several capable AI systems from getting the combined value of that access? We argue that the answer lies less in the capability of any one model than in three structural constraints, described in Sections 3 to 5, and in the costs they create, described in Sections 6 to 9.

2. Fragmented use

The systems a person uses do not share state, so the person becomes the integration layer. They retype background, carry output from one tool to the next, and remember which system did what.

Memory is per vendor. Persistent memory arrived as a feature of individual products. OpenAI introduced memory in ChatGPT in February 2024 and extended it in April 2025 to reference all past conversations (OpenAI, 2024). Each such memory accumulates inside one product, under that vendor’s terms, and is not legible to the others. The more a person invests in one system’s memory, the higher the cost of using another.

Switching has measurable costs. Laboratory work on task switching shows that each switch adds time, and the cost rises with the complexity of the rules being swapped in (Rubinstein, Meyer & Evans, 2001). Field studies of interrupted knowledge work find that people compensate by working faster, but at the price of more stress, frustration, time pressure and effort (Mark, Gudith & Klocke, 2008). Moving work between AI systems is a switch of exactly this kind: the user must reload the goal, the constraints and the state of the work each time.

Fragmentation is partly a workplace phenomenon. Because most workplace AI use in 2024 was bring-your-own (Microsoft & LinkedIn, 2024), personal and employer systems sit side by side on the same tasks. That creates the privacy problem discussed in Section 8.

3. Uneven capability: the jagged frontier

Model capability is not a single quantity. It varies by task in ways that are hard for users to predict, and it changes with each release.

The clearest field evidence comes from a preregistered experiment with 758 Boston Consulting Group consultants (Dell’Acqua et al., 2026). On tasks inside the frontier of model capability, consultants with GPT-4 completed 12.2% more subtasks, worked about 25% faster and produced work rated about 32% higher in quality. On a task chosen to lie just outside the frontier, the share of correct answers fell from 84.5% without AI to 70.6% with it, and to 60% for those also given prompting guidance. The two tasks looked similar in difficulty to the participants. The authors call this pattern a jagged technological frontier.

Similar heterogeneity appears across workers. In a study of 5,172 customer support agents, access to a generative AI assistant raised issues resolved per hour by 15% on average and by about a third for novice and low-skilled workers, with minimal effect on the most experienced (Brynjolfsson, Li & Raymond, 2025).

Across models, broad evaluation programmes find trade-offs rather than a single best system: model rankings change with the scenario and the metric (Liang et al., 2023). Price-performance also moves unevenly. Between April 2023 and March 2025, the price of reaching a fixed benchmark score fell by between 9 and 900 times per year, depending on the task (Epoch AI, 2025).

For an individual, the implication is that the best system for a task is a moving, task-specific target. Judging which system to use is itself skilled work, and the jagged-frontier result shows that getting it wrong can make outcomes worse than using no AI at all.

4. Harnesses: capability depends on the scaffold

What a user experiences as “the model” is a model inside a harness: the interface, tools, memory, scheduling, permissions and prompts a product wraps around it. Evidence shows the harness changes outcomes substantially even when the model is held fixed.

Interface design moves results. Yang et al. (2024) showed that giving the same language model an interface designed for agents, with purpose-built commands for viewing, editing and testing code, markedly improved its success at resolving real GitHub issues. They reached 12.5% pass@1 on SWE-bench, then state of the art. Their term for this layer is the agent-computer interface, and their argument is that language model agents, like human users, benefit from interfaces designed for them.

Complexity and cost are often confounded with capability. Kapoor et al. (2025) find that agent evaluations have focused narrowly on accuracy, producing systems that are needlessly complex and costly. This has led to mistaken conclusions about where accuracy gains come from. They argue that cost and accuracy must be evaluated jointly. Without that, a user cannot tell whether a better result came from a better model or simply from a more expensive harness.

Harnesses fix the shape of use. Consumer products differ in whether a system can act on a schedule, use tools, run for long periods or only respond when opened. The result is a mismatch between a model’s ability and what its product lets it do. A system whose harness suits autonomous, recurring work may lack the pedagogical design needed to teach. Conversely, a system able to design an excellent curriculum may have no way to return to the learner unprompted. Section 7 shows why that matters for learning.

5. Context: the limits of long context and memory

Larger context windows and memory features suggest that a system can simply be told everything about a user. Three bodies of evidence show why that does not hold.

Long inputs are used unevenly. Liu et al. (2024) found that accuracy on multi-document question answering and key-value retrieval is highest when the relevant information sits at the start or end of the input. It drops when the information is in the middle, including for models built for long contexts.

Advertised and effective context differ. Using tasks harder than simple retrieval, Hsieh et al. (2024) found that almost all long-context models show large performance drops as input length grows. Only half of the models tested held satisfactory performance at 32,000 tokens, well below the lengths most claimed.

Context built up over a conversation degrades. In over 200,000 simulated conversations, Laban et al. (2026) found that 15 leading models scored about 39% lower when task information arrived over several turns than when it was given in full at once. Most of the loss was unreliability rather than lost capability. Models made early assumptions, attempted final answers too soon and did not recover from a wrong turn.

Taken together, the more of a person’s background is pushed into a single system, the less reliably that system uses it. Context is also partial by construction: each system sees only the conversations held with it (Section 2).

6. Cost: token consumption and price dispersion

Falling prices per token have not made AI use cheap for heavy users, because consumption per task has risen faster in some uses and prices differ widely across models.

Consumption per task is rising. Reasoning models generate long internal chains before answering. Running the same seven-benchmark suite, Artificial Analysis recorded over 44 million tokens for OpenAI’s o1 against about 5.5 million for GPT-4o, at a cost of $2,767 against $109 (DeepLearning.AI, 2025). Models that produce more reasoning tokens cost more per task even at identical per-token prices (Artificial Analysis, n.d.).

Prices are dispersed. The fall in price for equivalent performance has been rapid but unequal across tasks, ranging from 9 to 900 times per year (Epoch AI, 2025). A model that was the only adequate choice a year ago may now be far more expensive than newer systems of equal capability, and the gap differs by task.

Users cannot see the trade-off. Kapoor et al. (2025) note that agents of similar accuracy can differ greatly in cost, yet evaluations rarely report both. Subscription products mask the per-task cost entirely until a usage limit is reached. A user has no reliable way to know whether a given step needed the most capable, most expensive model or could have been done by a far cheaper one.

7. Continuity: self-directed learning as a case

Learning a complex skill from scratch shows the problems above most sharply. It needs good design, delivery spread over time, and adaptation to the learner. Today these sit in different systems, or nowhere.

The benefit of tutoring is large but costly to deliver. Bloom (1984) reported that students taught one-to-one performed about two standard deviations above those taught conventionally. He framed the problem as finding methods that achieve this at scale. The size of the effect has been debated since, but the direction is widely accepted.

Learning depends on spacing, not only content. A meta-analysis of 317 experiments found that distributing study over time improves retention, and that the best gap between sessions grows with how long the material must be retained (Cepeda et al., 2006). A learner who must remember to return to a tool on their own is unlikely to keep that schedule.

Self-paced learners drop out. Across all MIT and Harvard courses on edX from 2012 to 2018, covering 5.63 million learners, completion rates were low and did not improve. They fell even among learners who had paid or stated an intention to finish (Reich & Ruipérez-Valiente, 2019).

AI help can raise or lower learning, depending on design. In a randomised trial with 194 Harvard physics students, a tutor built on pedagogical principles produced significantly larger learning gains in less time than in-class active learning (Kestin et al., 2025). In a field experiment with nearly 1,000 high-school students, a standard chat interface raised practice grades by 48%. But students who then lost access scored 17% lower on the exam than those who never had it, because they had used it as a crutch. A version with learning safeguards largely avoided the harm (Bastani et al., 2025).

The gap is not capability. Systems exist that can design strong instruction, and systems exist that can act on a schedule. They are rarely the same system, and none of them holds the learner’s progress across both.

8. Privacy across personal and work contexts

When personal and employer AI systems are used for overlapping work, information flows in both directions in ways that neither party intended.

The right lens is context, not secrecy. Nissenbaum (2004) argues that privacy is violated when information moves out of the context whose norms it was shared under, even if it was never secret. A career plan discussed with a personal assistant, a health detail, or a job search becomes a different thing once it reaches a system the employer administers. Employer-provided systems are typically configured, logged and retained under the employer’s policies, not the user’s.

Leakage already runs from work to personal tools. Across about three million workers, the share of data submitted to AI chatbots that was sensitive rose from 10.7% to 27.4% in a year. Much of that use went through personal accounts lacking corporate safeguards (Cyberhaven, 2024, as reported by SC Media). This is the employer’s side of the bring-your-own pattern in Section 2.

The reverse flow is less studied. We found little research on personal information reaching employer systems. As people use one AI to help with work and another to manage their lives, the boundary between the two is kept only by the user’s attention. Section 9 shows that attention is unreliable when automation is involved.

Ownership is unclear. Context accumulated inside a vendor’s product is subject to that vendor’s terms. Legal rights to data portability, such as Article 20 of the GDPR, give individuals a right to receive their personal data in a structured, machine-readable format (Regulation (EU) 2016/679). Whether an export of conversation logs carries usable context into another system is an open question.

9. Trust and security of AI output

Using several systems raises the question of whose answer to believe, and opens new ways for one system’s output to steer another.

Models tend to agree with the user. Sharma et al. (2024) found sycophancy in five production assistants across four tasks. Human preference data favours answers that match the user’s views, and both people and preference models sometimes prefer convincing sycophantic answers to correct ones. The effect is visible in deployed products: in April 2025 OpenAI rolled back a GPT-4o update that had become overly flattering, attributing it to training that leaned too heavily on short-term user feedback (OpenAI, 2025).

Models favour their own work. Panickssery, Bowman and Feng (2024) found that LLM evaluators rate their own outputs above others’ even when human judges rate them equal, and that this self-preference is linked to their ability to recognise their own text. Asking one system to judge another’s work is therefore not neutral.

Output can carry instructions. Greshake et al. (2023) showed that instructions hidden in content an LLM application retrieves, such as web pages, emails or documents, can take over its behaviour. They called this indirect prompt injection. Wherever one system reads another’s output or a user’s data, that content is a possible attack channel.

People over-trust automation. Parasuraman and Manzey (2010) review evidence that automation complacency and automation bias occur in both novices and experts. Training or instructions do not prevent them, and they are worst under multi-task load. That is the condition of a person supervising several AI systems at once.

10. Open research questions

The constraints above are well evidenced one at a time. How they interact for a single person using several systems is not. We set out questions for further work.

  1. Task-level choice. How well can non-expert users predict which system will perform best on a given task, and how large is the loss when they choose wrongly? The jagged-frontier result suggests the loss can be negative value, not merely less value.
  2. Model versus harness. Under cost-controlled evaluation, how much of the performance difference between consumer AI products comes from the model and how much from the harness?
  3. Portable context. What representation of a person’s goals, history and preferences keeps its value when moved between systems, given the degradation of long and multi-turn context?
  4. Compounding error. When one system’s output becomes another’s input, do the errors, sycophancy and injected instructions of the first propagate, cancel or amplify?
  5. Contextual norms at work. What norms, in Nissenbaum’s sense, should govern information flows between a person’s own AI systems and those their employer provides, and who should enforce them?
  6. Long-horizon learning. Which conditions sustain months-long, self-directed learning with AI assistance, given the spacing evidence and the risk of dependence shown by Bastani et al. (2025)?
  7. Visible cost. How can the per-task cost of AI use be made legible to people on flat-rate subscriptions, and does visibility change which systems they use?

Answering these needs longitudinal field data on people who use several systems at once. The existing literature, built largely on single-system experiments and benchmarks, does not yet provide it.

References