Who controls the answer?

A NOTIFICATION appeared on some iPhones carrying the BBC News logo. Luigi Mangione, the man accused of killing UnitedHealthcare chief executive Brian Thompson in New York, had shot himself, it said.

Except he had not. Mangione was alive and in custody. The BBC had published no such claim. Apple Intelligence grouped three BBC notifications into a single summary and turned one into a false statement. Following complaints from the BBC and other news organisations, Apple temporarily suspended notification summaries for news and entertainment apps.

The incident was widely treated as another AI error, but its implications went further. The BBC had done the reporting, yet AI had changed what potentially millions of users thought it had reported. Its name remained on the alert even though it no longer controlled the information attached to it.

Something similar is now happening on a much larger scale. People are turning to AI assistants with questions they once asked Google, journalists or experts, such as what happened in an election or why a war began.

In all fairness, search engines were never neutral to begin with. They decided whether a report from Dawn appeared beside an obscure Indian website presenting itself as a research platform, and most users rarely moved beyond the first few results. But users could still open the original reports, see where each claim came from and judge the sources. AI removes that step. It reads the material, summarises it and presents its own version of what happened.

This matters because malicious actors are already polluting the information space. Elections bring coordinated disinformation, synthetic audio and manipulated videos designed to promote candidates or discredit opponents. Governments and parties are no longer simply responding to news but producing their own version of events.

People are turning to AI assistants with questions they once asked Google, journalists or experts, such as what happened in an election or why a war began.

The same contest is visible in Gaza, where restrictions on journalists, competing claims and loaded language have shaped coverage. The killing of civilians, destruction of hospitals and homes, restrictions on aid and wider Israeli abuses are extensively documented, yet presented through radically different frames.

Months of reporting by journalists, UN agencies and human rights groups may be placed against a single denial from the Israeli or US government. An AI assistant may then present both as equally credible, without showing how much evidence supports each account. The history of occupation may disappear, leaving what appears to be a dispute between two sides.

The democratic danger is not limited to hallucination. It lies in allowing a handful of private companies to organise public knowledge. Citizens may receive one answer without knowing which newsroom established the facts, what evidence supports them or which accounts were left out. A confident summary can close a debate before the reader knows one exists.

This is already happening. A study coordinated by the European Broadcasting Union and led by the BBC examined more than 3,000 responses from four major AI assistants across 18 countries and 14 languages. It found that 45 per cent had at least one significant issue. Serious sourcing problems appeared in 31pc, while 20pc contained major accuracy issues.

There is an economic consequence too. Traditional search normally preserved the click. Readers reached the news website, encountered the original reporting and generated some advertising or subscription value. AI systems can extract the result of that work, compress it into a paragraph and retain the user.

Pew found that users were almost half as likely to click on a search result when an AI summary appeared. Publishers surveyed by the Reuters Institute expect search traffic to fall by 43pc over the next three years. Their reporting is still being used, but fewer readers are reaching the organisations that produced it.

Large international publishers are negotiating licensing agreements with AI companies but smaller organisations have little comparable leverage. This is particularly dangerous for Pakistan, where journalism is already weakened by declining revenues, political pressure, delayed salaries and dependence on government and corporate advertising. Most publishers have no subscription base, limited technical capacity and little understanding of how often AI systems collect their work.

The European Union has begun addressing parts of the problem. Its AI Act requires certain AI-generated or manipulated material, including deepfakes and unedited text on matters of public interest, to be labelled. Providers of general-purpose AI models must maintain copyright compliance policies and publish summaries of their training material. Systems designed to influence elections are treated as high risk, while the Digital Services Act requires the largest platforms and search engines to address risks to elections and civic discourse.

But these protections do not settle the central question. They do not require an assistant to present several credible accounts, explain why it preferred one source, direct users prominently to the original report or compensate the publisher whose journalism supplied the answer.

Pakistan needs to begin this discussion before the terms are set elsewhere. News organisations should measure AI crawling and referral traffic, decide how their archives may be used and bargain collectively with technology companies. Common agreements should require clear attribution, links to original reports, a process for correcting false summaries and fair payment when journalism is used commercially. Publishers must also rebuild direct relationships with audiences through memberships, newsletters and products that do not depend entirely on platforms.

Journalism has lost audiences to new technologies before. This time, AI does not only decide which report is seen. It can rewrite the report, mix it with weaker sources and deliver the result without the reader reaching the newsroom. If these systems depend on reporting, they cannot be allowed to make the reporter disappear.

The writer is the founder of Media Matters for Democracy.

Published in Dawn, August 26th, 2026