When the Reader Is a Machine

An organisation used to publish to people. A policy, a product page, a set of results and a safety notice were written on the assumption that the person who needed them would read them, in the order they were written, on the page where they sat.

A growing share of that material is now read first by an assistant and relayed to a person as a summary. The person may never see the page. Whatever the summary says becomes, for that reader, what the organisation said.

The failure is sourcing, not invention#

The strongest measurement available is not about commerce. Twenty-two public service media organisations across 18 countries and 14 languages, coordinated by the BBC and the European Broadcasting Union, put 30 shared news questions to the free consumer versions of four general assistants and assessed the answers.

Forty-five per cent of responses carried at least one significant issue and 81 per cent carried an issue of some kind. The largest single cause was sourcing, at 31 per cent, ahead of accuracy at 20 per cent and insufficient context at 14 per cent. Gemini recorded significant issues in 76 per cent of responses against 37 per cent for the next product.

Three cautions travel with those numbers. These were free consumer versions tested in mid-2025 and all four have shipped new defaults since, so the figures describe a moment rather than current performance. The study was not adversarial and question difficulty was not controlled, so the rate is not a worst case. And the assessors work for the organisations whose journalism is the subject matter, which gives them a stake in the finding.

Read with both cautions, the shape of the finding is what matters for anyone who publishes. The dominant failure was not a machine inventing a claim. It was a machine attaching a real claim to the wrong source, or dropping the context that made the claim true. An organisation cannot protect itself from that by being accurate. It is already accurate. The claim is being detached from the thing that made it checkable.

The person on the other end will not check#

A summary with a sourcing problem would be a minor matter if the reader went back to the page. Two findings in this base say otherwise.

Parasuraman and Manzey, reviewing automation bias and complacency across aviation, medicine and military settings, found under-questioning of automated advice in novices and experts alike, resistant to training and worse under workload. The effect size for generative systems is far less predictable than for the automation they studied, so the finding transfers as a direction rather than a magnitude.

Lee and colleagues surveyed 319 knowledge workers about 936 real uses of AI at work and found higher confidence in the tool associated with less critical thinking, while higher confidence in oneself went the other way. It is self-report and cannot separate cause from selection: people who think differently may use the tools differently.

Taken together they describe a reader who receives a confident summary under time pressure and does not go to the source. That reader is a customer, a candidate, a regulator or a journalist.

What this changes about publishing#

The instinct is to write for the machine. That reads as an optimisation problem and leads to worse pages. The better move is to write so that a claim survives extraction, which is also how to write for a person in a hurry.

• Put the claim, its source and its date in the same place. A summariser takes sentences, not layouts. A claim whose evidence sits three paragraphs below it arrives alone.
• Name yourself inside the sentence that matters. Attribution carried only by a page header or a byline is the first thing to fall away.
• Date anything that will stop being true. A regulatory position or a price quoted without a date becomes a claim with no shelf life, and it will be relayed long after it expires.
• Say what you did not establish. A limitation written next to a finding travels with it more often than a caveat kept in a separate section.
• Check how you are being described. Ask the assistants the questions your customers ask about you, and read what comes back as a report on your own publishing rather than as a verdict on the tool.

That last one is the cheapest audit available to any organisation and almost nobody runs it on a schedule.

What would settle it#

The same methodology as the broadcasters' study, applied to commercial material: product claims, safety information, terms, published policy. Nobody has published that work. Until somebody does, the sourcing finding above is evidence about news answers and the extension to commercial publishing is an argument.

Where this sits in my own argument#

This research holds itself to a method of dated claims, named sources and stated limits, which exists for human readers who want to check. The same discipline turns out to be what survives machine summary intact. The alternative is for an organisation to be represented accurately by accident, which is drift in the ordinary sense.

Related SuperSkills research#

On the commercial half of the same shift, when your customers have agents. On the reader who does not check, automation bias and human in the loop is not a safeguard. On how this research handles its own sourcing, how this research works.

Key sources

• Fletcher, J. and Verckist, D., BBC and European Broadcasting Union (2025). News Integrity in AI Assistants.
• Parasuraman, R. and Manzey, D. H. (2010). Complacency and Bias in Human Use of Automation.
• Lee, H.-P. et al. (2025). The Impact of Generative AI on Critical Thinking.

About this research#

Written by Rahim Hirji, author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company.

How this research works · Reviewed quarterly · Found an error? Tell me and it is corrected on the page.