The dilemmas of delegation: An analysis of policy challenges posed by Advanced AI Assistants and natural-language AI agents
A 65-page policy analysis of Advanced AI Assistants, defined as foundation-model products that hold fluid natural-language conversation, personalise heavily to the user, and adopt human-like roles. It proposes a three-part taxonomy (Executors that act on the world, Advisers that set out options or steps, Interlocutors that aim to change the user's mental state, with Wysa as the example) and maps eight classes of policy challenge that would arise even if assistants worked exactly as advertised, including privacy, market distortion, political influence, deskilling, mental health, and professional standards.
Publisher
Ada Lovelace Institute
Published
11 Nov 2025
Added
today
DOI
—
Key Findings
- Assistants are split into Executors (act directly on the world, example OpenAI Operator), Advisers (set out advisory options or steps, example DoNotPay) and Interlocutors (aim to bring about a change in the user's mental state, example Wysa); most commercial agents with a natural-language interface fall inside the Assistant definition.
- Because Assistants combine personality with personalisation they are easy to anthropomorphise, form emotional attachments with and trust, which makes their outputs persuasive enough to influence life decisions; the report describes the evidence for this as ample but anecdotal.
- Eight harms are listed as possible even under lawful use and full functionality, including hard-to-detect influence on political views, widespread cognitive and practical deskilling, undermining mental health and flourishing, and degrading the quality of public and professional services.
- The deskilling and wellbeing section notes reports of AI companions confusing, upsetting, undermining or manipulating users and cultivating dependency, and early evidence that many companions are unsupportive in crisis and therefore poorly suited to vulnerable users, while cautioning that assistant-specific evidence is thin and mostly associational.
- Interlocutors are projected to become a key means by which many people regulate their mental health and a key mediator of interpersonal relationships within five years.
- The report frames Assistants as prospective universal digital intermediaries with gatekeeping power over what information, viewpoints and options users see, and calls for governments to develop an Assistant-specific strategy rather than rely on existing AI governance work.
Methodology Notes
Desk-based policy analysis with no empirical component. Dated November 2025 on the PDF cover; the publication page's structured metadata gives 2025-11-11 and lists three authors (Farmer, Smakman, Birtwistle), while the PDF cover names Harry Farmer. The PDF (65 pages) was fetched and read on 2026-09-15; the report page is reachable with curl but 403s browser-emulating fetchers. Companion piece to the December 2025 legal analysis 'The regulation of delegation'.
Topics
Authors
Harry Farmer, Julia Smakman, Michael Birtwistle
Tags
Cite This
APA
Harry Farmer, Julia Smakman, Michael Birtwistle. (2025). The dilemmas of delegation: An analysis of policy challenges posed by Advanced AI Assistants and natural-language AI agents. Ada Lovelace Institute. https://www.adalovelaceinstitute.org/report/dilemmas-of-delegation/
Related Insights
The regulation of delegation: Are AI advisers, agents and companions regulated in the UK?
Ada Lovelace Institute (legal analysis by AWO) · 1 Dec 2025
Friends in our phones? Legal responses to AI companions
Information & Communications Technology Law (Taylor & Francis); University of Bristol Law School; University of Otago · 4 Sept 2026
AI Assistance Reduces Persistence and Hurts Independent Performance
arXiv (Carnegie Mellon University; University of Oxford; Massachusetts Institute of Technology; University of California, Los Angeles); published at COLM 2026 · 6 Apr 2026
LLMs as Oracles: Reliance on LLMs for Subjective Personal Questions
arXiv (Stanford University and collaborators; affiliations not printed on the abstract page) · 13 Sept 2026