Independent Research Agenda — Not Client Work

Ten open questions on AI and public life

These are subjects I intend to research, set out in the open before the work is done: the exact question, how it will be answered, which records it will rest on, and where it could go wrong. Nothing on this page reports a finding. Each entry stays marked open until the sources have been read and the evidence chain is published.

Questions open
10
Findings published so far
None on these topics
Status of every entry
Open — research not yet run

Section 01

Standards these projects will be held to

Stated first, so the agenda can be judged against them later.
  • No finding is published without a document from the body that issued it. Summaries are a route to the record, never a substitute for it.
  • Every project states what could not be established, and treats a null or mixed result as a publishable result.
  • Each entry carries the date it was checked, because law, deployments and model behaviour all change.
  • Bias, human rights, safety, fairness and proportionality are treated as the areas of concern, in line with UNESCO's Recommendation on the Ethics of Artificial Intelligence.
  • Every one of these is independent research. None is client work, and none is done on behalf of any government, company or campaign.

Section 02

The ten questions

Each is written so that it can be answered from the record, or shown to be unanswerable from it.
RA-01 · Elections and democracyFirst findings published

AI and election deepfakes

How have governments and election authorities responded to verified cases of AI-generated political deepfakes, and which countries have introduced enforceable disclosure or removal requirements?

Why it matters: Elections, democratic process, misinformation and freedom of expression.

How it will be answered

  • Build a country table from the text of enacted law and election-commission rules, not from news summaries.
  • Separate enforceable duties (disclosure, takedown, penalties) from voluntary codes and draft bills.
  • Record, for each entry, the instrument, its date and the body that issued it.

Source base

  • National legislation portals and official gazettes
  • Electoral commission rules and guidance
  • EU AI Act and Code of Practice on Disinformation texts
  • Court judgments where obligations were tested

Known risk in this question: A proposal, a fine and a code of conduct are three different things. Anything not yet in force will be labelled as such.

RA-02 · Privacy and policingFirst findings published

Facial recognition and public surveillance

Which governments currently use facial-recognition technology in public spaces, what legal safeguards apply, and which deployments have been suspended or ruled unlawful?

Why it matters: Privacy, policing, security, race and civil liberties.

How it will be answered

  • Split the record into three columns: deployment confirmed by the deploying body, safeguard in law, and adverse ruling or suspension.
  • Only count a deployment when the police force, ministry or city itself documents it.
  • Log the date of each ruling and whether it was appealed.

Source base

  • Court judgments and data-protection authority decisions
  • Police and ministry procurement and policy documents
  • National data-protection legislation
  • Municipal council records for city-level bans

Known risk in this question: Deployment status changes quickly and pilot schemes are often reported as permanent. Each row will carry the date it was checked.

RA-03 · Employment and discriminationFirst findings published

AI discrimination in recruitment

What documented cases show automated recruitment systems disadvantaging applicants based on gender, race, age, disability or location, and what action did regulators take?

Why it matters: Employment, discrimination law and corporate accountability.

How it will be answered

  • Include only cases with a regulator's decision, a settlement, or a filed complaint on the public record.
  • Record the alleged mechanism separately from the finding actually made.
  • Note where a company denied liability or settled without admission.

Source base

  • Equality and employment regulators' enforcement records
  • Court and tribunal filings and judgments
  • Regulator guidance on automated hiring tools

Known risk in this question: Widely repeated anecdotes about hiring tools are often reported without a regulatory record behind them; those will be excluded or flagged.

RA-04 · Public services and human rightsFirst findings published

Automated decisions in welfare systems

Which governments have used automated systems to detect welfare fraud or determine benefit eligibility, and what documented errors affected legitimate recipients?

Why it matters: Poverty, public services, state power and human rights.

How it will be answered

  • Anchor each case in an official inquiry, audit or judgment rather than in campaign reporting.
  • Quantify harm only where an official body published figures.
  • Record the remedy: compensation, suspension of the system, or neither.

Source base

  • Parliamentary inquiry reports and national audit offices
  • Ombudsman findings and court judgments
  • Data-protection authority decisions
  • OECD and other inter-governmental analyses of automation in social protection

Known risk in this question: Numbers of affected people vary widely between sources; only officially published figures will be stated, with their issuing body named.

RA-05 · Policing and due processFirst findings published

Predictive policing

Where have predictive-policing systems been deployed, what data did they use, and what evidence exists that their recommendations disproportionately targeted particular communities?

Why it matters: Policing, racial bias, crime prevention and due process.

How it will be answered

  • Distinguish place-based forecasting from person-based risk scoring; they raise different questions.
  • Take the data inputs from procurement documents or force policy, not from vendor marketing.
  • Treat disproportionality as established only where an audit, inspectorate or court examined it.

Source base

  • Police inspectorate and oversight-board reports
  • Freedom-of-information disclosures published by the requesting body
  • City council decisions to adopt or discontinue systems
  • Peer-reviewed evaluations

Known risk in this question: Several well-known systems were discontinued years ago and are still described in the present tense elsewhere.

RA-06 · Safety and consentOpen — not yet researched

Non-consensual AI-generated intimate images

Which countries have criminalised creating or distributing non-consensual AI-generated intimate images, and how do their penalties and victim protections differ?

Why it matters: Sexual abuse, women's and children's safety, consent and platform duties.

How it will be answered

  • Compare the statutory text itself: what conduct is criminalised, the maximum penalty, and whether intent to distribute is required.
  • Record separately whether victims have a takedown right and against whom it runs.
  • Note the commencement date, which often lags enactment.

Source base

  • Criminal codes and amending acts on official legislation portals
  • Explanatory memoranda and parliamentary records
  • Regulator guidance to platforms

Known risk in this question: This is a fast-moving area and the comparison will be dated on its face. No case details or victim material will be reproduced.

RA-07 · Language access and digital inequalityFirst findings publishedPriority project

AI assistants in Ewe, Twi, Ga and Hausa

How accurately do leading AI assistants provide health and public-service information in Ewe compared with English, and how do Twi, Ga and Hausa compare?

Why it matters: Digital inequality, cultural representation and access to essential information.

How it will be answered

  • Build a fixed question set on health and public services where a correct answer exists in an official Ghanaian or WHO source.
  • Ask each question in English and in each language, with identical wording, and record the full response verbatim.
  • Score each response against the official source on three axes: factual accuracy, completeness and language quality.
  • Have the language quality of each response reviewed by a fluent speaker, and say plainly who reviewed it.
  • Publish the full prompt set, the raw responses and the scoring rubric so the test can be re-run.

Source base

  • Ghana Health Service and Ministry of Health guidance
  • WHO fact sheets for the same conditions
  • Ghanaian public-service portals for the administrative questions

Known risk in this question: This is original empirical work, not a literature review. It needs named models, fixed dates, a fluent-speaker reviewer and a published rubric before any result is stated — and the sample will be small enough that it must be described as an indicative test, not a benchmark of any model.

RA-08 · Work and the economyOpen — not yet researched

AI and entry-level employment

What measurable changes in entry-level employment occurred after companies adopted generative AI, and do official labour statistics support claims of widespread job replacement?

Why it matters: Employment, young people, inequality and the future of work.

How it will be answered

  • Start from national labour-force statistics rather than from company announcements.
  • Separate what the statistics show from what is attributed to AI, and state where attribution cannot be established.
  • Report a null or mixed result as a result.

Source base

  • National statistical offices' labour-force series
  • ILO and OECD employment databases
  • Central bank and treasury labour-market analyses

Known risk in this question: Current evidence is genuinely mixed. The likely honest finding is that headline claims outrun the statistics, and the write-up will say so rather than manufacture a trend.

RA-09 · Health equityOpen — not yet researched

AI diagnostics across rich and poor countries

Which AI diagnostic systems have received regulatory approval, and is there evidence that their performance differs across countries, ethnic groups or underrepresented populations?

Why it matters: Health inequality, patient safety, medical access and algorithmic bias.

How it will be answered

  • Take the approval record from the regulator's own device database, including the approval route.
  • Record what population the submitted evidence covered — this is frequently the finding itself.
  • Report performance differences only from peer-reviewed evaluation or a regulator's own review.

Source base

  • Medical-device regulators' public approval databases
  • WHO guidance on AI in health
  • Peer-reviewed external validation studies

Known risk in this question: Approval is not evidence of equal performance, and absence of evidence about a population is not evidence of equal performance either. Both will be stated as such.

RA-10 · Law and liabilityFirst findings published

Legal responsibility when AI causes harm

When an AI-assisted decision causes documented financial, medical, employment or reputational harm, which party is legally responsible — the developer, the deploying organisation, the professional user or the platform?

Why it matters: Law, consumer protection, corporate responsibility and emerging regulation.

How it will be answered

  • Work from decided cases and enacted liability rules, by jurisdiction.
  • Set out where the law is settled, where it is contested and where no case has yet been decided.
  • Avoid stating a general rule; liability turns on jurisdiction and on the role of the party.

Source base

  • Court judgments and regulator enforcement decisions
  • Product-liability and consumer-protection statutes
  • EU AI Act and related liability instruments
  • Professional-regulator guidance on AI use

Known risk in this question: Nothing here is legal advice, and much of the field has no decided authority yet. Open questions will be labelled open.

Section 03

What happens next

One question at a time, published in the same format as the existing benchmarks.

Each completed project will get its own page with the question, the method as actually run, every source with a working link, the evidence chain, the searches logged, and a plain statement of what could not be established. Until then these entries stay marked open — publishing a conclusion before the sources have been read is exactly the failure this portfolio is meant to argue against.